An International Journal of Computing and Informatics

rumber 1 March 1993 ISSN 0350-55%

InformaticaAn International Journal of Computingand Informatics

3Cc

Profiles: Terry WinogradJapan: Knowledge ArchivesStanford: Language and InformationALPHA Project

I

o

5o

.§

The Slovene Society Informatika, Ljubljana, Slovenia

Informatica

An Intemational Journal of Computing and Informatics

Subscription Information

Informatica (ISSN 0350-5596) is published four times a year in Spring, Summer,Autumn, and Winter (4 issues per year) by the Slovene Society Informatika, Vožarskipot 12, 61000 Ljubljana, Slovenia.

The subscription rate for 1993 (Volume 17) is- DEM 50 for institutions,- DEM 25 for individuals, and- DEM 10 for studentsplus the mail charge.

Claims for missing issues will be honored free of charge within six months after thepublication date of the issue.

Technical Support: Borut Žnidar, DALCOM d.o.o.Stegne 27, 61000 Ljubljana, Slovenia

Printed by Biro M, d.o.o., Žibertova 1, 61000 Ljubljana, Slovenia.

Orders for subscription may be placed by telephone or fax using any major creditcard. Please call Mr. R. Murn, Department for Computer Science, Jožef StefanInstitute: Tel (+38) 61 214 499, Fax (+38) 61 158 058, or use the bank accountnumber LB 50101-678-51841.

According to the opinion of the Ministry for Informing (number 23/216-92 of March27, 1992), the scientific journal Informatica is a product of informative matter (point13 of the tariff number 3), for which the tax of trafRc amounts to 5%.

The issuing of the Informatica journal is financially supported by the Mimstry forScience and Technology, Slovenska 50, 61000 Ljubljana, Slovenia.

Informatica 17 (1993)

TOWARDS AN INFORMATIONAL ORIENTATION

The international issue of Informatica is only a beginning of the new informa-tional orientation, which is already on hand in research and technology commu-nities concerning the realm of the informational. This realm extends and willmore and more extend into the cultural, knowledge, and technology, irrespec-tive of the national, regional, continental, or global. The new perspective fromthe computer scientific to the informational began with the criticism at the verybeginning of artificial Intelligence, in 1965, when Hubert L. Dreyfus launchedhis report to RAND, in which he argued that "Early success in programmingdigital computers to ezhibit simple forms of intelligent behavior, coupled withthe beliefthat intelligent activities differ only in their degree of complexity, haveled to the conviction that the processing underlying any cognitive performancecan be formulated in a program and thus simulated on a digital computer."(Alchemy and Artificial Intelligence, The RAND Corporation Paper P-3244,December 1965.)

The next important milestone of understanding computers and cognitionwas the book published by Terry Winograd and Fernando Flores, in 1886,by Ablex Publishing Corporation (see Profiles on the next page). By thiswork, philosophical thought concerning hermeneutics, phenomenology, Beingand autopoiesis (philosophers Husserl, Austin, Gadamer, Heidegger, Habermasand biologist Maturana) was introduced, that is, meaningly legalized withinthe global artificial intelligence community.

The last push came from Japan when, in 1993, the Knowledge ArchivesProject started its work (look at Mission and Research Reports). The goal ofthis project is to become a project for projects and to connect and manageinternationally, interdisciplinary, and interindustrially the fields of artificial in-telligence, knowledge engineering, humanities (cultures over the globe), socialsciences, computer sciences, multimedia technologies, etc. with the goal toimplement the global knowledge system by means of both humans and newtechnological tools which will emerge during the run of the project.

Among others, Informatica will follow the problems of research and tech-nology, which fall into the domain (niches) of the informational, through aninternational cooperation, joining disciplines and researchers of different pro-fessional orientations. It will enable the publishing of submitted matters, whichfall into the field of informational conceptualism, theory, design, and technol-ogy. In this respect, cybernetics, humanities, social sciences, philosophy, etc.concerning informationally oriented problems will be treated, searching newpaths of informational research and technology.

At the beginning of the international issuing of Informatica there are stillunsaid editorial and organizational problems, which have to be solved in thecoming years. For instance, we are establishing the final Mjp t style for severalkinds of submitted matters (articles, reports, news, special colurans, e tc) . TheE-mail net of editors of Informatica over the globe is beginning to functioneffectively. Informatica is becoming "our journal" for all participating parties.We are still in acquiring of new editors and authors at the instantaneous journalcirculation of a thousand copies. And this beginning is promising.

—Anton P. Železnikar, Editor-in-chief

Informatica 17 (1993)

PROFILES

On this page we present editors of Informatica ina spontaneous and free way, in the form of thereply for which the Editor-in-chief has asked ed-itors to deliver their biographies. These biogra-phies may be of particular interest for readersof Informatica to learn about the circumstancesin which editors live and work. In the first at-tempt of Informatica's editors profiles, we havethe opportunity to give a short story of profes-sor Terry Winograd who together with FernandoFlores surprised the artificial intelligence commu-nity by the work Understanding Computers andCognition. This work gave a substantial impulsenot only to a new orientation but also to the crit-ical look to the phenomenon of intelligence andpossibilities of machine intelligence in particular.

Terry Winograd

Terry Winograd is Professor of Computer Scienceat Stanford University. He received his B.S. inmathematics from The Colorado College in 1966and a Ph.D. in Applied Mathematics from MITin 1970. He taught at MIT from 1970 to 1973 andhas been on the faculty of the Computer ScienceDepartment of Stanford University since 1973. Healso has appointments in the Department of Lin-guistics and the Program in Values, Technology,Science, and Society and is on the advisory boardof the Stanford Humanities Center.

Winograd's research on natural language un-derstanding by computers is often cited as a majormilestone in artificial intelligence. It was the ba-sis for his book Understanding Natural Language(Academic Press, 1972), and his textbook Lan-guage as a Cognitive Process (Addison-Wesley,1983) as well as numerous articles in both schol-arly journals and popular magazines. His mostrecent book, co-authored with Fernando Flores,takes a critical look at work in artificial intelli-gence and presents an alternative theory of lan-guage and thought, which suggests new directionsfor the design of intelligent human/computer sys-tems. The book, entitled Understanding Comput-ers and Cognition: A New Foundation for Design(Addison-Wesley, 1987), was named as the bestinformation science book of 1987 by the AmericanSociety for Information Science. He recently co-

edited a book with Paul Adler entitled Usability:Turning Technologies into Tools (Oxford Univer-sity Press, 1992).

Winograd's current research on the foundationsof the design of computer systems builds on thetheories developed in his book with Flores. Heis developing a "language-action perspective" inwhich current and potential software and hard-ware devices are analyzed and designed in the con-text of their embedding in work and communica-tive structures. The language/action perspectivegrew out of his earlier work in artificial intelli-gence, but it shifts the focus of attention awayfrom the mental and the individual, to the socialactivity by which we generate the space of cooper-ative actions in which we work and to the technol-ogy that is the medium for those actions. He alsohas developed several new courses at Stanford, in-cluding one on Computers, Ethics and Social Re-sponsibility, and a series on the Design of Human-Computer Interaction (sponsored by the NationalScience Foundation). He directs aproject at Stan-ford called the Project on People, Computers andDesign. During the 1992-93 academic year he ison leave from Stanford, working with Interval Re-search, a new research laboratory in Palo Alto.

IVinograd was the keynote speaker for the 1988Conference on Office Information Systems, the1990 Conference on Computer-Human Interac-tion (CHI'90), and the first National Conferenceon Computing and Values (1991). He edited aspecial issue of the ACM Transactions on Of-fice Information Systems (Spring 1988) on the"Language/action perspective." He is on the ed-itorial board of a number of journals, includingAI Ezpert, AI & Society, Journal of Computingand Society, Human-computer Interaction, andComputer-Supported Cooperative Work.

Winograd is a board member and consultantto Action Technologies, a developer of workgroupproductivity softvvare. He was a founding memberof Computer Professionals for Social Responsibil-ity. He has been on the national board since theorganization was founded, and served as nationalPresident from 1987-1990.

—A.P. Železnikar

Informatica 17 (1993) 3-12 3

METHODOLOGIES FROM MACHINE LEARNING IN DATAANALYSIS AND SOFTWARE

D. MichieThe Turing Institute, George House, 36 North Hanover Street,Glasgow Gl 2AD

Keywords: artificial intelligence, machine learning, software

Edited by: Matjaž Gams

Received: February 25, 1993 Revised: March 8, 1993 Accepted: March 15, 1993

In the last few decades the rise of computing and telecommunications has flooded theworld ofgovernment, business, medicine and engineering whh unprecedented volumes ofstored data. These databases provide the raw material for information supply but havebeen largely impenetrable as potential sources ofexport knowledge. Computer-orientedtechaiques caa now be used, however, in integration with est&blished methods fromclassical statistics to generate rule-structured classiHers which not only make a betterjob of classifying new data sampled from the same source but also possess the qualityofclear explanatory structure. New developments in the computer induction of decisionrules have contributed to two areas, multivariate data a.nalysis and computer assistedsoftware engineering. Practical connections between the two are thereby coming to light.This paper reviews some of the more significant of these developments.

1 Introdliction vised and demonstrated an algorithm, ID3 (Iter-ative Dichotomizer 3) [25], capable of extracting

Computer induction of decision rules from sample complete logical descriptions from large files ofmultivariate data was already known a quarter multivariate noise-free data, adverse to analysisof a century ago. But Hunt, Marin and Stone's for reasons of logical rather than statistical com-CLS (Concept Learning System) initially aroused plexity. The data-analytic era of rule inductioninterest only among cognitive psychologists [11]. was consolidated with the pioneering work Clas-In recent years, however, new developments have sification and Regression Trees by Breiman, et al.contributed to two applied disciplines, namely [5].(1) multivariate data analysis and (2) computer-assisted software engineering. Practical connec- T h e °P e n i n š P a s s a S e o f t h e a b o v e t r e a t i s e c o n "tions between (1) and (2) are also thereby coming t a i n s t h e s t a t e m e n t : ' A n important criterion forto light. This paper reviews some of the more a g ° o d cl^sification procedure is that it not onlysignificant of these developments. * produces accurate classifiers but that it also pro-

T n „„_ r, . , . , , ,, . ^ , vides insight and understanding into the predic-In 1977 Fnedman independently mcorporated . & , , , , T , , <• , ,

,, i- i r/~IT n • i vi -J. LI r tive structure oi tne data. In the last few decadestne essentials ol CLb m an algorithm smtable for . . . . , , . .. , . . • . . r , .. . • , . /. the nse ol computiner and telecommunicationsmducing decision trees from statistical-type (l.e. , n , , , , , , ,t • >\ J A n , i - i_ • , j i • . . has nooded the worlds oi business, government,noisy ) data. For this he mtroduced an lmportant . . . , . . . , ' , , '

'. ,. „ ,, ^ , medicine and engineennff with unprecedented vol-mnovation, automatically prunmg the tree s more . , , , , ,,. , , , , ,, , , f ,. j umes of stored data. These databases providedistal nodes under the control of a user-supplied , . i r . r • ! , ,

. A, , . ,, ,. „ . . . tne raw matenal lor lnlormation supplv, but haveparameter. At about the same time Oumlan de- , . , , . ,

been largely lmpenetrable as potential sources ofiQ. ., rr., .... ... , j . ., expert knowledge. However, computer-orientedbimilar paper with the same title was pubushed ln the r ° •> r

Computer Journal. Permission for reprint obtained by the techniques can now be used in integration withBritish Computer Society. established methods from classical statistics to

Informatica 17 (1993) 3-12 D. Michie

generate rule-structured classifiers which not onlymake a better job of classifying new data sampledfrom the same source, but also possess the qual-ity of clear explanatory structure emphasised inthe above-quoted passage. Applicability has beenshown both to data derived from databases (realcase histories) and from simulators (model casehistories), as in the follovving knowledge-intensiveapplication areas.

Synthesis from simulation data- aerospace- instrumentation- manufacturing- pharmaceuticals- electronics trouble-shooting- interpretation of biomedical monitoring- generating software from specifications

Synthesis from captured data- nuclear engineering- gas and oil processing- circuit fault diagnosis- steel and chemical process industries- seismic measurement interpretation- clinical diagnosis- credit control- stockmarket assessment

In what follows the essentials of rule-based tech-niques drawn from machine learning are sum-marised in the context of earlier approaches.

2 Nature of the Problem

Given. A 'training set', or estimation sample,of case-descriptions, each in the form of a listof attribute-values (e.g. age, duration of preg-nancy, number of previous births, number of pre-vious pregnancies, marital status, etc), togetherwith a classification of each case into, say, YES,NO, or more generally class\,class2,.. • , c/ass,-,(e.g. Did patients with these characteristics electto have an amniocentesis test? Did loan appli-cants giving these questionnaire answers turn outto be creditworthy? Was this pattern of meteoro-logical measurements followed by thunderstorms?Which category of fault was identified from theobserved engine test results?)

Required. A classifier (i.e. some formula or rule

defmed over the attributes) which can classify newcases sampled from the same population.

2.1 Two Approaches

We demand of a classifier: (1) that it should pre-dict with high accuracy; (2) that it should be sim-ple and easy to understand.

Decision formulae derived from standard mul-tivariate statistics, such as discriminant analy-sis or 'naive Bayes' methods, have the form ofsets of positive and negative weights, scoringattribute-values for their individual contributions(assumed independent) to an ACCEPT-versus-REJECT preference for each decision class. Suchlists of numbers mean less to the user than theydo to the machine.

Logical decision formulae ('rules') get awayfrom the arithmetic. Instead of the operators ad-dition, multiplication etc, decision-tree rules forexample use the logical operator 'if...then...else'for combining relevant subsets of attributes intoclassifying expressions. The above-referencedtreatise by Leo Breiman and his colleagues openswith the following illustration:

At the University of California, SanDiego Medical Center, when a heart at-tack patient id admitted, 19 variablesare measured during the first 24 hours.These include blood pressure, age and 17other ordered and binary variables sum-marising the medical symptoms consid-ered as important indicators of the pa-tient's condition.The goal of a recent medical study wasthe development of a method to iden-tify high risk patients (those who willnot survive at least 30 days) on the ba-sis of the initial 24 hours.

The tree-structured classification rule whichthese authors obtained from their data is below.

if the minimum systolic blood overthe initial 24-hour period < 91

then risk is HIGHelse if age < 62.5

then risk is NOT-HIGHelse if sinus tachycardia is present

then risk is HIGHelse risk is NOT-HIGH

METHODOLOGIES FROM MACHINE LEARNING... Informatica 17 (1993) 3-12

The reason for using the term 'tree-structured'is clear from the graphical representation of thissame rule, as shown in Fig. 1.

Breiman and his colleagues comment on thisrule that 'its simplicity raises the suspicion thatstandard statistical classification methods maygive dassification rules that are more accurate.When these were tried, the rules produced wereconsiderably more intricate, but less accurate.'During the six years which have elapsed sincethen, dedsioix-tree induction from data has beensubjected to field trials in various countries byboth academic and industrial groups. The resultshave been in conformity with the above-cited ob-servation. The chief advantages have been that:

1. the amount of calculation is much smaller;

2. the dassifiers produced are easier to under-stand;

3. filtering out irrelevant attributes is done au-tomatically;

4. decision-tree classifiers induced from data inthe style of Breiman and his colleagues havebeen found in practice to be usually moreaccurate than dassifiers formed by adding updiscriminant scores.

2.2 Reasons for Improved Accuracy

When applied to the kinds of data which make dif-ficulties for standard statistical analysis decision-tree methods gain improved accuracy in two ways.

1. They can handle both numerical and non-numerical attributes with equal ease.

2. They do not suffer any loss of discriminantpower when some of the attributes violatethe simplistic assumption of mutual indepen-dence. Consider the four items of dedsiondata (cases) in Table 1 which, if read as state-ments from an 'oracle' instead of passively asdata, collectively define the exdusive-or rela-tion between two binary attributes.

Linear scoring systems such as discriminantanalysis are powerless to find scoring coeffidentsfor al and a2 such that they can be multipliedby al's and a2's values, added up, and comparedwith a threshold in order reliably to distinguishtrue and false cases. If the reader cares to try it

Case al a2 Class1 +1 +1 False2 +1 - 1 True3 - 1 +1 True4 - 1 - 1 False

Table 1: A simple example of a data-set whichgives trouble to linear numerical estimation meth-ods. For rule-induction procedures (indudingstandard Boolean simplification procedures) theproblem is trivial

he will find that it cannot be done. But in com-mon with other logic-based induction algorithms,such as those routinely employed by electrical en-gineers for drcuit simplification, dedsion-tree in-duction trivialises the problem:

if al = +1then if a2 = +1

then falseelse true

else if a2 = +1then trueelse false

To construct from such data a classifying ex-pression using numerical multivariate analysiswould require an excursion into non-linear re-gression equations, or (equivalently, see Angus)[1] into multilayer neural networks. Moreover,dedsion-rule methods are not limited, as areBoolean simplification techniques, to problemswhere attributes are guaranteed to be of logi-cal type. Modern rule-induction techniques areequally at home with inputs which indude numer-ical attribute-values. For this, the standard ap-proach converts input values to logical type by au-tomatically splitting numerical ranges into inter-vals according to an entropy-minimisation prind-ple. Among other sources, the book by Breimanand his colleagues can be consulted for details.

3 Machine Learning As DataAnalysis

Multivariate statistical methods were developedfrom mathematical and sdentific foundations by

6 Informatica 17 (1993) 3-12 D. Michie

Isage >62.5?

Issinus tachy-cardia present ?

Is the minimum systolic bloodpressureover the inifial 24hour

period > 91 ?

Figure 1: Decision tree induced from heart-attack data corresponding to if-then-else rule (see text).'G' stands for high risk, 'F' for lower risk.

pioneers such as Francis Galton, Karl Pearson andRonald Fisher. These men made it possible togive precise answers to questions phrased vvithinthe statistical paradigm. 'Within what limits oferror can this or that classifier be expected toperform?' or 'How much will the error be nar-rowed by a given increase in the size of estimationsample?' Studies in the computational theory oflearning are directed towards building scientificfoundations for the machine-learning approach asan extension of classical probability and statistics(in the case of decision-tree induction see, for ex-ample Refs 7, 16, 17 andl8). The connection be-tween inductive learning and statistical data anal-ysis can be explained as follows.

Approximately two decades of exposure to dataturns a baby into a mentally capable adult. Ev-idently the developing brain extracts somethingof continuing and incremental value. We call theprocess of extraction 'learning'. For the some-thing which is extracted, stored, refined, built

upon and exploited, we have variations on no-tions of 'knowledge'. We speak about behaviour-patterns, habits, skills, hypotheses, beliefs, mod-els, descriptions, concepts, theories. These struc-tures have one thing in common: they all actas classifiers. Empirical scientists here recognisesomething familiar. They too are concerned withextracting theories from data, alternating withthe theory-guided sampling of new data. In thiscycle of extraction and testing the scientist com-monly calls on the aid of a special breed of nu-merical craftsman, the data analyst.

According to one rather undemanding defini-tion, the statistical data analyst's fitting of mod-els to data would qualify as a form of machinelearning. This definition says:

a learning system uses sample data (the trainingset)to generate an up-dated basisfor improved classificationof subsequent data from the same source.

METHODOLOGIES FROM MACHINE LEARNING... Informatica 17 (1993) 3-12

Notice that the definition, although phrasedstrictly in terms of classification, logically extendsto acquisition of improved performance on taskswhich do not look at all like classification. Itera-tive situation-action tasks come to mind such asriding a bicycle, solving an equation, or parsinga sentence. The extension becomes obvious whenfor the decision dasses we choose names whichrefer to partitions of the space of situations as'suitable for action A', 'suitable for action B', etc.

4 Machine Learning AsProblem Solving

To illustrate the way in which iterative situation-action problems can be coded as classification,consider the task of solving simple equations inschoolroom algebra (see Ref. 20). Thus 3x+l — 2is transformed by an appropriate action ('collectlike terms on same side of the equation') into3x — 2 - 1 , which is in turn transformed into3x = 1 (by 'combine like terms') and thence intothe final 'situation', x = | (by 'divide by the co-efficient of the unknown'). This is the solution or'goal'.

Using the attributes/classes format, theproblem-description is given in Table 2. Line 1gives the number of attributes, lines 2-7 describethe attributes and the last two give the numberand names of the classes.

The key to the six attribute names is:

al Does the equation have a common factor?(comfact)

a2 Are there like terms on opposite sides?(likeopp)

a3 Are there any bracketed terms? (bracket)

a4 Does either side have like terms? (likesam)

a5 Is exactly the same present on both sides?(sametrm)

a6 Is there only one unknown term and is itscoefficient equal to one? (xcoeffi)

The seven class names given earlier have thefollowing interpretations:

1. Divide the equation by its common factor (di-vef).

comfact logical yes nolikeopp logical yes nobracket logical yes nolikesam logical yes nosametrm logical yes noxcoeffi logical yes no7divef collect multbr combinecancel divex stop

Table 2: Problem description for inductive deriva-tion of an equation-solving rule

2. Collect like terms on the same side of theequation (collect).

3. Multiply out bracketed terms (multbr).

4. Combine like terms (combine).

5. Cancel out a term that appears on both sides(cancel).

6. Divide by the coefRcient of the unknown (di-vex).

7. Stop because the equation is solved (stop).

The four lines of examples given at the outset,namely

3x + 1 = 2,3x = 2 - 1,3x = 1, x = -

o

vvould appear in an induction file as:

ID no al a2 a3 a4 a5 a6 Class1 no yes no no no no collect2 no no no yes no no combine3 no no no no no no divex4 no no no no no yes stop

From the output of a skilled human equation-solver on a number of example cases of the abovekind, an induction system acquires expertise.From sequences of illustrative solution steps, asolver is built. Application of the solver to ac-tual problems is iterative. In each cycle the out-put action is applied to the current state of the


equation to yield a modified state. This is thenfed back to the classifier as a new input, andso on until the output 'stop' appears. Althoughseemingly remote from the task of synthesizing acontroller from a recorded behavioural trace of askilled human operator (see Ref. 19 for a vvorkedexample), the logic is the same. Successive 'snap-shots' of the state of the dynamical system corre-spond to successive lines of symbols in equation-solving. The STOP action is of course normallydropped to allow for instabilities and perturba-tions from the goal state, normally absent fromsyrnbol-maiupulation tasks.

5 A More DemandingDefinition Of Learning

A more demanding definition of learning is nowcoming from applied artificial intelligence:

a learning system uses sample datato generate an up-dated basisfor improved classificationof subsequent data from the same sourceand expresses the new basis in intelligible sym-

bolic form.

According to this more demanding definition,not only improved performance results from thelearning process, but also an explicit set of rules.Decision-tree induction in the algebra domainmeets not only the less demanding but also themore demanding criterion. An intelligible sym-bolic form obtained by the ACLS algorithm [23]from a training set of equation solutions is shownin Fig. 2. Such a form constitutes an operationaltheory, a prescription which can be followed byany agent able to interpret it, vvhether human ormachine.

Michie et aii. [20] found that use of a versionof ACLS by high-school children could be an ef-fective mode of self-instruction. Children wereasked to teach the machine equation-solving froman elementary algebra text-book by selecting andsupplying example solvings. The children's sub-sequent grasp was tested. against that of a groupof classmates who had been exposed instead to aconventional CAI algebra package.

if exactly the same term occurs on both sidesthen cancel the same term from each sideelse if the equation has a common factorthen divide by the common factorelse if there is a bracketed termthen multiply out the bracketselse if there are like terms on opposite sidesthen collect like terms on thesame sideelse if one side has more than one like termthen combine like termselse if the unknown term has a coefRcient notequal to onethen divide by the coefficient of the unknownelse stop; the equation is solved.

Figure 2: Induced rule as an operational theorydiscovered form supplied data

6 Machine Learning AsTheory-Construction

Scientists of the past have been content if theautomated procedures of data analysis satisfiedthe undemanding definition only, taking on them-selves the responsibility of abstracting from theanalysis the desired explanatory or predictive the-ories. Al-based machine learning, combining as itdoes both logical and statistical idioms, followsthe more demanding definition. It thus entersdirectly into the theory-building process, as wasfirst shown in 1976 by groups working in chem-istry at Stanford, USA, [6] and in plant pathol-ogy at the University of Dlinois [8]. Subsequentstudies in USA, Australia, Slovenia and Britainestablished the tree-structured paradigm of com-puter induction as the dominant form for rule-based data analysis. The following were the chiefadvances.

(1) In 1977 J.H. Friedman introduced waysof deriving tree structures from data in thestyle of Earl Hunfs 1966 scheme but constrainedby criteria for pruning unproiitable branches[9]. Decision-tree induction was thus generalisedto noisy data, yielding trees with confidencemeasures associated with their leaves (outcomenodes).

(2) In 1979 J.R. Quinlan published the firstof a series of papers describing the ID3 series ofdecision-tree algorithms, the efficiency and ver-satility of which led to their widespread adop-

METHODOLOGIES FROM MACHINE LEARNING... Informatica 17 (1993) 3-12 9

tion as the basis of commercial inductive softwareengineering [25]. For large problems (the 1979paper describes a rule-based solution of a prob-lem which in unreduced form constituted some 2 |million records) extracted decision trees were efR-cient at run-time but unstructured and hence ob-scure - the automated equivalent of hand-crafted'spaghetti code'.

(3) A.D. Shapiro and T.Niblett [28] elaboratedthe Quinlan paradigm with a method known as'structured induction', subsequently studied indepth by Shapiro [27]. To supplement the initiallygiven primitive attributes they added separatelyinduced procedural attributes. Structured induc-tion confers on inductive data analysis the bene-fits of top-down problem decomposition, as in thediscipline of structured programming. In its orig-inal form, it can only be applied where noise isabsent and where the problem can be fully spec-ified in terms of the primitives. Today the struc-turing steps can in suitable cases themselves beautomated (see (7) below).

(4) In 1987 J.R. Quinlan adapted his C4 algo-rithm for inducing trees from noisy data so as togenerate solutions in the form of compact sets oflogic rules [26]. After pooling the branches har-vested from the trees separately induced from thesame data set, the program winnows out redun-dancy and delivers a compact and intelligible localtheory of the data.

(5) Quinlan's current version, C 4.5, has beenused to recover from the recorded behaviour ofsimulator-trained human subjects sets of 'produc-tion rules' of the type postulated by cognitivepsychologists [19]. Such productions, in someneurally encoded form, are believed to under-lie learned decision skills. As compared withthe original expert behaviour, induced rules ex-hibit a 'clean-up efTect', having shed some of theinconsistencies and noise which even the mosthighly trained nervous system introduces into therecognise-act cycle.

(6) D. Michie derived, [17] and with A.Al-Attar,partly tested, a formulation of decision-tree induc-tion from the axioms of Bayesian probability [18].Main gains obtained from this approach are: (i)reunion of rule induction with statistical decisiontheory; (ii) a practical way of extending Shapiro-Niblett structured induction into the analysis ofnoisy data.

(7) Full automation of structured induction re-quires algorithms for generating new proceduralattributes, rather than depending for this on theknowledge-based insight of human domain spe-cialists. Advances by Muggleton and Buntine andby Bain and Muggleton at the Turing Institute,Glasgow [2,22] have yielded needed algorithmswithin the framework of first-order predicate logic(see also Ref. 21).

7 Inductive ProgramGeneration

The work by A.D. Shapiro referred to above ex-tended data-oriented induction into the realm ofcomputer-assisted software engineering (CASE),and recent extensions to first-order level have es-tablished points of contact with the formal meth-ods school. On the practical side, a path has fi-nally been found to circumvent what has some-times been termed the 'bottleneck problem' facingknowledge acquisition from experts.

A few years ago it was hoped that experts-doctors, engineers, etc. - would be able to teachtheir skills directly to computers, which wouldthen be able to carry out much of their routinediagnostic work. Faults or symptoms would befed into a computer, which vvould then give adiagnosis of the problem. Unfortunately, it wasfound that if explicit how-to-do-it rules are re-quired from them, experts cannot efFectively feedtheir own decision-making processes into com-puters. The soya bean specialist, the analyticalchemist or the cardiologist largely reacts 'intu-itively' to data, in ways he or she cannot fully ex-plain. In the realm of evaluating credit-worthinessin the finance industry. L. Sterling and E. Shapirogive a telling account of the phenomenon [30].

The major difficulty was formulating therelevant expert knowledge. Our expertwas less forthcoming with general rulesfor overall evaluation than for rating thefinancial record, for example. He hap-pily discussed the profiles of particularclients, and the outcome of their creditrequests and loans, but was reluctant togeneralise.

The observation that specialists transmit theirinarticulate skills to trainees by example, rather


than by using explicit rules, led in the mid-1980sto the development of commercial programs toexploit 'teaching by showing' in the style illus-trated with schoolroom algebra. The machinelearns 'how to do it' from experts who supply ex-amples. A number of large corporations, notablyBritish Petroleum [29], have begun using this ap-proach as a cost-effective way of building large ap-plications. The resultant programmer productiv-ities exceed current industry standards by an or-der of magnitude. The world's largest expert sys-tem, BMT, for configuring fire-detection equip-ment, was built by the German company Brain-ware using the RuleMaster and lst Class induc-tive shells [10]. BMT consists of 150,000 lines ofinductively generated C code and is in routine useby the client organisation. The total figure of 9man-years expended on the project includes man-agement, support staff and domain specialists aswell as programmer time. Reductions of softwaremaintenance overheads have also been impressive,as can be seen from the summary results set outin Table 3.

The reader can appreciate the connectionwith CASE methods by thinking of expert sup-plied examples as statements in a requirements-specification language: 'in situations of this kindwe require the system to do x; in situations ofthat kind we require it to do y, etc ' In this con-text an induction routine can be thought of asa kind of compiler, translating from a specifica-tion consisting of example cases into a programconsisting of if-then-else expressions. A mass ofpartly redundant and partly incomplete specifi-cations becomes a structured set of efficient ex-ecutable rules. Permissiveness tovvards redun-dancy and incompleteness in the user's tabulationof cases marks the sole but significant departureof rule-induction programming from mainstreamdecision-table methods [12, 13]. Rule inductionin a CASE context is indeed little more than therebirth of structured decision tables, with this dif-ference: that specification tables, as we may hereterm them, are initially partial, and are developedincrementally in successive cycles of generate-and-test. Herein lies the key to the extraordinaryproductivities, illustrated in Table 3, where num-bers of the order of 100 lines of installed code perprogrammer-day are the norm. The even moreremarkable gains in software maintain-ability can

be similarly explained. Modern inductive shellsautomatically flag not only incomplete parts ofthe rule-base but also those which have been de-rived from generalisation steps in the inductionprocess. The user can then interactively validateflagged passages, editing specification tables as re-quired, re-inducing at each stage from the edited'spec' [15, 16]. General implications for softwaremanufacture were discussed a few years ago in theauthor's Royal Society Technology Lecture [14],and have more recently been elaborated in thespecific context of interactive validation and ofinductive logic programming [16, 21].

8 Automated Synthesis OfOperational Knowledge

The possibility had long been foreseen of usinginductive inference to derive operational models(a formal equivalent of 'skills') from deep models(a formal equivalent of 'understanding'). Thusfrom a numerical model of an aero-engine a qual-itative model can in principle be prepared, fromwhich all possible combinations of engine faultsmay automatically be listed, together with thecorresponding sensor readings and test responses.From such a giant tabulation, efficient rules foruse in future fault diagnosis may be machine-induced. The wild card in the foregoing is thephrase 'principle'. But Slovenian work using acomputer model of the human heart resulted inthe discovery by machine of a corpus of new clin-ical rules for interpreting electrocardiogram pat-terns [3, 4]. Work at Glasgow's Turing Instituteon diagnosing electronic faults in a space satel-lite confirmed the methodology as fully practica-ble [24]. These synthetic rule-bases comprise newdiagnostic knowhow beyond the achievements ofhuman specialists. Their means of constructionalso automatically guarantees completeness andcorrectness with respect to the formal modelsused to generate the exhaustive datasets. Sincethe latter are logically equivalent to the descrip-tive specifications from which they were derived,they too may be viewed as formal specifications,in the form of very long sentences written in aground-level data-description language. This in-sight, however, logically irrefutable, seems strangeto some who approach via formal methods, vvhereconciseness in a specification is of the essence.

METHODOLOGIES FROM MACHINE LEARNING... Informatica 17 (1993) 3-12 11

ApplicationMYCIN

XCON

GASOIL

BMT

No.ofrulesMedicaldiagnosisVAX computerconfigurationHydrocarbonseparationsystemconfigurationConfigurationof fire-protectionequipment inbuildings

Develop-mentman-yrs400

8,000

2,800

>30,000

Main-tenanceman-years/year100

180

1

9

InductivetoolsN/A

30

0.1

2.0

N/A

N/A

ExpertEaseand Extran 7

1 st Class andRuleMaster

Table 3: Tabulation from Ref. 29, with 1990 data on BMT added

The Bratko and Pearce syntheses do indeed startfrom concise (intensional) specifications, but findthe path to automated synthesis via an intermedi-ate product, namely a complete extensional formwhich is its logical equivalent, with the addedproperty of inductive transformability into a set,again concise, of operational rules.

References

[1] J.E. Angus, Oa the coanection between neu-ral network learning &nd multivariate non-linear least squares estimation. InternationalJournal of Neural Networks 1, 42-46 (1989).

[2] M. Bain and S. Muggleton, Non-monotomclearning. In Machine Intelligence 12, editedJ.E.Hayes,D.Michie and E.Tyugu), OxfordUniversity Press, Oxford (1991).

[3] I. Bratko, I. Mozetic and N. Lavrac, Auto-matic synthesis and compression of cardio-logical knowledge. In Machine Intelligence II,edited J.E.Hayes, D.Michie and J.Richards,p.p. 435-454. Oxford University Press, Ox-ford (1988).

[4] I. Bratko, I. Mozetic and N. Lavrac, Kardio:a Study in Deep and Qualitative Knowledgefor Expert Systems. MIT Press, Cambridge,MA (1989).

[5] L. Breiman, J.H. Friedman, R.A. Olshenand C.J. Stone, Classification and Regres-sion Trees. Wadsworth and Brooks, Mon-terey, CA (1984).

[6] B.G. Buchanan, D.H. Smith, W.C. White,R.J. Gritter, E.A. Feigenbaum, J. Lederbergand C. Djerassi, Applications of Artificial In-telligence for chemical inference. XXII. Au-tomatic rule formation in mass spectrometryby means of the meta-DENDRAL program.Journal of the American Chemical Society98, 6068-6078 (1976).

[7] L. Cestnik and I. Bratko, On estimatingprobabilities in tree pruning. In Proceed-ings of the Sixth European Working Ses-sion on Learning (EWSL-91), Porto, Portu-gal (1991).

[8] R. Chilausky, B. Jacobsen and R.S. Michal-ski, An application of variable-valued logic toinductive learning ofplant d/sease diagnosticrules. Proceedings of the Sixth Annual Inter-national Symposium on Multi-variable Logic,Utah (1976)

[9] J.H. Friedman, A recursive partitioning de-cision rule for nonparametric classification.IEEE Transactions on Computers, C-26,404-408 (1977).


[10] J.E. Hayes-Michie (ed.) Pragnmtica. (Inter-national Bulletin of the Induclive Program-ming Special Interest Group), No. 1. TuringInstitute Press, Glasgow (1990).

[11] E.B. Hunt, J. Marin and P.Stone, Exper-iments in Induction. Academic Press, NewYork (1966).

[12] L.S. Levy and H.T. Stump, Inverted deci-sion tables and their app/ication: automatingthe translation of specifications to peograms.AT&T Technical Journal 64, 533-558 (1985).

[13] A. Lew, Proof otcorrectness ofdecision tableprograms. The Computer Journal 27, 230-232 (1984).

[14] D. Michie, The superarticulacy phenomenonin the context ofsoftware manufacture. Pro-ceedings of the Royal Society of London A405, 185-212. Also reproduced in The Foun-dations of Artificial Intelligence, edited D.Partridge and Y. Wilks), pp. 411-439. Cam-bridge University Press.

[15] D. Michie, Problems of automatic conceptformation. In Applications of Expert Sys-tems, vol. 2, edited J.R. Quinlan), pp. 310-333. Addison-Wesley, London (1989).

[16] D. Michie, Human and machine learning ofdescriptive concepts. In ICOT Journal (ed.Hiroshi Hara), March issue, pp. 2-20. Insti-tute for New Generation Technology, Tokyo(1990).

[17] D. Michie, Personal models of rationality. InJournal of Statistical Planning and Inference(special issue on foundations of statistics andprobability), 21, 381-399.

[18] D. Michie and A. Al-Attar, Use of sequen-tial Bayes with class probability trees. Inmachine Intelligence 12, edited J.E. Hayes,D. Michie and E. Tyugu. Oxford UniversityPress, Oxford (1991).

[19] D. Michie, M. Bain and J.E. Hayes-Michie,Cognitive models from subcognitive skills.In Knowledge-based Systems in IndustrialControl, edited M. Grimble, S. McGhee andP. Mowforth. Peter Peregrinus, Stevenage(1990).

[20] D. Michie, A. Paterson and J.E. Hayes-Michie, Learning by teaching. Second Scan-dinavian Conference on Artificial Intelligence(SCAI 89).

[21] S. Muggleton, Inductive logic programming.In Algorithmic Learning Theory, Springer,Heidelberg (1990).

[22] S. Muggleton and W. Buntine, Machine in-vention of first-order predicates by invertingresolution. Proceedings of the Fifth Interna-tional Conference on Machine Learning, pp.339-352.

[23] A. Paterson and T. Niblett, ACES UserManuai. Intelligent Terminals Ltd, Glasgow(1982).

[24] D. Pearce, The induction of fault diagno-sis system from qualitative models. In Pro-ceedings of the Seventh National Conferenceon Artificial Intelligence (AAAI-88) St Paul,Minnesota, pp. 353-357 (1989).

[25] J.R. Quinlan, Discovering rules by inductionfrom large collections of examples. In ExpertSystem in the Microelectronic Age, editedD. Michie, pp. 168-201. Edinburgh Univer-sity Press, Edinburgh.

[26] J.R. Quinlan, Generating production rulesfrom decision trees. International Joint Con-ference on Artificial intelligence 1987 (IJCAI-87), pp. 304-307. Kaufmann, Los Altos, CA(1987).

[27] A.D. Shapiro, Structured Induction in Ex-pert Systems. Turing Institute Press andAdisson Wesley, London (1987).

[28] A.D. Shapiro and T. Niblett, Automatic in-duction of classiHcation rules for a chessendgame. In Advances in Computer Chess,vol. 3 edited M.R.B. Clark, pp. 73-92 (1982).

[29] S. Slocombe, K. Moore and M. Zelouf, En-gineering expert system applications. Pre-sented at the BCS Annual Conference (De-cember 1986).

[30] I. Sterling and E. Shapiro, The Art ofProlog.The MIT Press, Cambridge, MA (1986).

Informatica 17 (1993) 13-20 13

ANALYTICAL FORM SOLUTION OF THE DIRECTKINEMATICS OF A 4-4 FULLY IN-PARALLELACTUATED SIX DEGREE-OF-FREEDOMMECHANISM

C. InnocentiDIEM - Facolta di Ingegneria - Universita di BolognaViale Risorgimento, 2 - 40136 Bologna -ITALYTel 051-6443450 - Fax 051-6443446e-mail:[email protected]. Parenti-CastelliIstituto di IngegneriaFacolta di Ingegneria - Universita di FerraraVia Scandiana, 21 - 41100 Ferrara - ITALYTel +39-532-65150 - Fax +39-532-740983

Keywords: robots, kinematics, parallel mechanisms, closure equations

Edited by: Jadran LenarčičReceived: February 10, 1993 Revised: March 4, 1993 Accepted: March 15, 1993

The paper presents the direct position analysis in analytical form of a six-degree-of-freedom 4-4 fully-parallel mechanism. For a given set of actuator displacements themechanism becomes a structure and the analysis finds all the possible closures of thestructure. The analysis is performed in two steps. First, the two closures of thetetrahedron-like subchain ofthe structure are found. Then, for each tetrahedron closure,two transcendental equations are determined that represent the closure of the remain-ing part ofthe 4-4 structure. The two equations can be reduced to algebraic equationsand, after eliminating the unwanted unknowns, a final 8th order equation in only oneunknown is obtained. Hence, the maximum number of possible real closures of the 4-4structure is sixteen. Numerical examples are reported which illustrate and confirm thenew theoretical result.

1 Introdliction ies (base and platform) connected to each otherby six adjustable-length legs. The leg ends are

The requirement of improving the performances connected to the base and the platform by spher-of robotic systems seems to find a promising an- ical pairs. Several connection patterns are possi-swer in adopting parallel structures which, in- ble since two or three spherical pairs can coalescedeed, experience high payload to link weight ra- to form multiple spherical pairs, consequently dif-tios, high stiffness and a better position accuracy ferent mechanisms can be devised. The actuatedwith respect to the widely used serial manipula- legs provide the platform vvith six degrees of free-tors. A wide bibliography for the research on par- dom relative to the base.allel mechanisms and manipulators can be found The direct position analysis (DPA) of parallelin [!]• mechanisms asks for position and orientation (lo-

Parallel mechanisms are closed chains with one cation) of the output link (platform) when a setor more loops where only some pairs are actively of actuator displacements is given. It represents acontrolled. The basic arrangement of a fully in- challenging problem for the non-linear equationsparallel actuated mechanism consists of two bod- involved. Indeed, many solutions are possible and

14 Informatica 17 (1993) 13-20 C. Innocenti & V. Parenti-Castelli

numerical methods, currently adopted to this pur-pose, prove difficulty to find all solutions. On thecontrary the analytical form solution, if feasible,would provide all the possible solutions and addinsight into the kinematics of the mechanism.

The analytical form solution is represented, ingeneral, by a system of equations in echelon form,that is, if the equations are solved in the appro-priate order, each equation can be regarded asone equation in only one unknown. By referringto algebraic equations, the closed form solutionis feasible when the order of the equations is lessor equal than four; for higher orders the solutionsmust be determined numerically and the DPA issaid to be solvable in analytical form. In general,the first equation to be solved is of high orderwhile the remaining are linear (in the unknownto be solved for). Thus, in this case, the order ofthe first equation represents the maximum num-ber of possible real solutions.

Only few parallel mechanisms have been solvedin analytical form, as reviewed in [2], and some ofthem have been solved after mechanism geomet-rical simplification such as base and platform pla-nar [3-7] and symmetrically shaped [8] have beenintroduced. Mechanisms with general geometry,although more difficult to be solved in analyticalform, would give, however, more freedom to themechanism design, also aJlowing simpler hardwareconstructions.

This paper presents the DPA in analyticalform of the fully-parallel mechanism schemati-cally shown in Fig. 1. The six legs meet boththe base and the platform at four points, centersof spherical pairs. Namely, two legs meet singlythe base and the platform, while the remainingfour meet in pair both the base and the platform,forming with these a tetrahedron- like pattern.Consequently, either single and double sphericalpairs occur at the connection points. The free-dom each leg has to rotate about its axis does notaffect the gross motion of the mechanism, how-ever it can be eliminated by a suitable hardvvaredesign (For instance, by substituting a universaljoint for one spherical pair). The four connectionpoints does not necessarily belong to the sameplane. It is worth noting that the base and theplatform have a symmetric topological role. Themechanism is denoted as 4-4 parallel mechanism.When the leg lengths are frozen the mechanism

Figure 1: The 4-4 fully-parallel mechanism.

becomes a statically determined 4-4 structure.Still maintaining the 4-4 pattern, different

mechanisms can be obtained by devising variousleg connection arrangements to the base and plat-form. Several of these have been studied in [4]where the DPA analytical form solution of threecases (and one subcase of these) has been pre-sented. In their study both base and platformhave been considered as planar.

The 4-4 mechanism presented in this paper isa new case that has not been treated in [4]. Itcompletes the 4-4 parallel mechanism dass, dis-regarding arrangements with three leg ends thatcoalesce.

The analytical form solution of the 4-4 mecha-nism is obtained in two steps. The first step solvesfor the two possible closures of the tetrahedron-part of the 4-4 structure. For each of them, thesecond step provides two transcendental equa-tions that represent the closure of the remainingpart of the 4-4 structure. The two equations canbe reduced to algebraic form and, after eliminat-ing one unknown, a final 8th degree polynomialequation in only one unknown is obtained. Thus,in the complex field, the number of possible solu-tions for the 4-4 mechanism is sixteen. Finally anumerical example is reported that supports thenew theoretical result.

ANALYTICAL FORM SOLUTION OF Informatica 17 (1993) 13-20 15

Figure 2: The 4-4 structure. Figure 3: The tetrahedron structure.

2 Direct Kinematics

In this section, the kinematic model of the 4-4structure, schematically shown in Fig. 2, is devel-oped.

The base points Aj, j = 1,4, and the platformpoints Bj, j = 1,4, the centers of spherical pairs,and their positions are known in reference systems

p,the base while system Wp is fbced to the platform.It can be recognized that points Ai, A^, B\ andB-i are the vertices of a tetrahedron, which can beregarded as a rigid body when the leg lengths Lj,j = 1,4, are given (see Fig. 2). It is also recog-nized that the tetrahedron itself can be assembledin different ways.

The closure of the structure is thus performedin two steps. First the closures of the tetrahe-dron regarded as a separate rigid body are deter-mined. Then, for each of them, the closures of the4-4 structure are found. Henceforth, the positionvector of a point P in reference system Wk is de-noted by (P)k, and the components of a vector uin reference system Wk are denoted by (u)k-

2.1 Tetrahedron's assembling

With reference to Fig. 3, system Wt fixed to thetetrahedron is chosen with origin in A\, positive

direction of x axis from A\ to A2, y axis suchthat point B\ lies in the plane (x,y) with positivey component, and z axis according to the right-hand rule.

Position vcctor (A2)t is:

and Wp, respectively. System Wb is fixed to where

(1)

(2)

The coordinates of point B\ in Wt are determinedas follows. The angle /? € [O,TT] of vector (B\ —A\)t forms with x-axis of Wt is given by:

cos p = E2)/{2 (3)

If |cos^| > 1 the tetrahedron cannot be assem-bled in the real field and the whole 4-4 struc-ture could not be assembled either. Supposing| cos/?|si, it stems:

where2 ,O-|isin/3 = +[l-cos2/3]2

(4)

(5)

The position of point Bthe following conditions:

can be determined by


(B2 - A2)2 = L\

(B2-Bx)f=F2

(6) (12)

where:(7)

It is worth considering the following position:

(B2-Ax)t = 6-(A2-Ax)t + (8)H-(B1-Ai)t +a • (A2 - Ax)t • (Bx - Ax)t

where 6, [i and a are quantities to be determined.The system of equation (6) can then be rewrit-

ten in the form:

(B2-AX)2 = L2

[(B2-Ax)t-(A2-Ax)t]2 = L2

[(B2-Ax)t-(Bx-Ax)t]2 = F2

(9)

which gives:

(10)(B2 - Ax)2 = L\

L2 + (A2-AX)2-2(A2-Ax)f-(B2-Ax)t =

L2 + (BX-AX)2-2(Bx-Ax)J-(B2-Ax)t =

where, for obtaining the second and the thirdequation, the first has been taken into account.

By substituting in (10) the position (8) for(B2 — Bx)t, the following system is obtained:

[(A2-Ax)J-(Bx-Ax)t}2^0

is satisfied, i.e., if points Ax, A2 and B2 are notaligned.

Let condition (12) be verified. Thus 6 and /Jcan be determined and, from equation (11), thevalue of a2 can be determined. The tetrahedroncan be assembled only if:

> 0 (13)

In this case, indeed, the two real solutions for ecan be obtained:

(14)

and the tetrahedron can be assembled in two dis-tinct ways.

Once the tetrahedron is assembled, it can beconsidered as a binary rigid body connected tobase and platform by means of two revolute pairswhose axes are defined by points Ax, A2 andpoints Bx, B2 respectively (henceforth, axes AXA2

and B\B2 for brevity).In order to perform the second step of the

DPA, the description of the tetrahedron's shape interms of Denavit-Hartenberg parameters is of in-terest. The shortest distance between axes AXA2

and B\B2 is represented by line AQBQ (see Fig.3). The position of points AQ and BQ can be de-termined as follows.

With the positions:

(11)[(A2 - Ax) - Ax)t}

2) =

2-S-fi-(A2-Ax)J-(Bx-Ax)t

6-(A2-Ax)2

t+»- (A2 - AX)J • (Bi - Ax)t =

6-(A2- AX) (Bx - Ax)2 =

= l-(A2-Ax) (15)

= m-{B2-Bx)

where / and m are scalar quantities to be deter-mined, the geometrical conditions:

(A2-Al)T--(Bo-Ao) =

(B2-Bx)T-(Bo-Ao) =

(16)

that represent the orthogonality conditions of theEquations (11) represent a linear system in S and line A<\RQ with the axes AXA2 and B\B2 can bcfi that admits a unique solution if the condition \vritten as:

ANALYTICAL FORM SOLUTION OF . . . Informatica 17 (1993) 13-20 17

(17)

With condition (12) still verified, linear system(17) can be solved for 1 and m, after all vectorsare measured in reference system Wt.

By equation (15), position of points AQ in Wband BQ in Wp can be determined respectively.

The D-H parameter a, distance between axesAiA2 and B\B2, is given by:

a = \{A2-Ax)tx{B2-Bx)t\(18)

The D-H parameter a, skew angle of axes AXA2

and B\B2, can be obtained by: Figure 4: The auxiliary structure.

(A2-cos a = -(B2 - Bx)t

sina =

F1 • F\{A2-Ax)tx{B2-Bx)t]

(19)

- Ax)t

E-F-a

The two possible closures of the tetrahedron aretwo mirrored closures, thus the parameters /, m,a, and cosa have the same values for both ofthem, while parameter sin a has opposite values.

2.2 4-4 structure closure

Let the tetrahedron be assembled in one of thetwo possible closures and, recalling the previouslyreported consideration, consider the tetrahedronas a binary link T connected to base and platformby two revolute pairs. Hence, the 4-4 structure(see Fig. 2) can be represented by the kinemati-cally equivalent structure shown in Fig. 4.

A reference system WA fixed to the base is cho-sen with origin in Ao and z axis directed fromA\ to A2. Let RA be the 3 x 3 rotation matrixfor the (coordinate) transformation from WA toWb. Similarly a reference system WB fixed to theplatform is chosen with origin in BQ and z axisdirected from B\ to B2. Let RB be the 3 X 3 ro-tation matrix for the (coordinate) transformationfrom WB to Wp.

A careful inspection of Fig. 4 shows that the lo-cation of the platform (with respect to the base)

can be uniquely parametrized by the angles <f>iand <fo, which represent respectively the angularposition of link T with respect to base and plat-form. Conversely, for a given location of the plat-form with respect to the base, angles <f>i and <j>2

result uniquely determined.

Let R be the 3 x 3 rotation matrix for the co-ordinate transformation from WB to WA- MatrixR is a function of angles <f>i and <f>2 for a givengeometry of the 4-4 structure.

By imposing the constraints due to the legsA3B3 and A4B4, whose lengths are respectivelyL5 and LQ, the closure equations of the equiva-lent 4-4 structure (of Fig. 4) can be written as:

[R • (B3 - B0)B + (Bo - A0)A-

(A3 - A0)A}2 = L\

(20)

(A4 - A0)A}2 = L\

where

R=C\C2

S\C2

— US\S2

— UC\S2

V\S2

~Cl«2 -SiS2 -

vc2

US\S2

UC\C2

VS\

— VC\

u(21)


(Bo - AQ)A =a • C\

0(22)

and u — co sa , v = s i n a , Cj = cos <f>j, Sj = sin <f>j,with j = 1,2.

By considering that (Bo — Ao)2 = a2, system(20) can be rewritten as follows:

(A3 - A0)T

AR(B3 - B0)B+

(A3 - AofA(Bo - A0)A -(B3 - B0)

TBRT(B0 - A0)A =

[(A3 - A0)A2 + (B3 - B0)B2 + a2 - L}]/2(A4 - A0)

TAR{BA - B0)B+

{A4 - Ao)l(Bo - A0)A -(B4-B0)lR

T(B0-A0)A =[{A4 - A0)A2 + (B4 - B0)B2 + a2 -

(23)

where

R Q — AQ)A =a • c2

—a • s2

0(24)

Relation (24) puts into evidence that equations(23) are linear in the cosine and sine of angles <j>\a n d <f>2.

Equations (23), after the well known expres-sions:

(25)

where tj = tan(mj/2), are substituted for sineand cosine, can be rearranged in the followingform:

eirt[-t}2 = (26)

«",J=0,2

•\J=O,2

where coefficients e^j and / tJ (i, j = 0,2) are func-tions of both the geometry of the 4-4 structureand the assembly configuration of the tetrahe-dron.

Equations (23) are two algebraic equations inthe unknowns t\ and t2. One unknown, t2 for in-stance, can be eliminated from (23), and one finalequation in unknown ti only is thus obtained.

The elimination of t2 leads to the following con-dition:

= 0 (27)

GQ G\ G2 00 GQ G2 G2

HQ HI H2 0O JJ TT TT

iio -"l il2

that, after developing the 4x4 determinant, leadsto:

(G0H2 - G2HQ)2+ (28)

= 0

t=0,2

where the following positions have been consid-ered:

(29)

i=0,2

Indeed, the determinant in (27) is the eliminantof system (26) and its vanishing represents thenecessary and sufficient condition for equations(26) to have the same solutions for t2.

Equation (28) is an 8th order algebraic equationin the unknown t\, which has eight solutions inthe complex field.

Determination of t2. For each solution t\ = t\k(k = 1,8) of equation (28), left-hand sides ofequations (26) are polynomials in the variable t2.They generally admit a first-order greatest com-mon divisor whose vanishing provides the com-mon root t2 = t2k.

Thus, the position of the platform with respectto the base can be determined by means of rela-tions (25), (22) and (21). In particular, the position of points Bj of the platform in referencesystem Wb is given by:

(Bj)b = RA[R{Bj - BQ)B

+(A0-Ob)b

(Bo - Ao)A]

(30)

where (AQ — Ob)b is the position vector of point AQin Wb, and both matrix R and vector (ô - AQ)A

are computed for t\ = t\k and t2 = t2k.In conclusion, one closure of the 4-4 structure

can be obtained for every solution of equation

ANALYTICAL FORM SOLUTION OF Informatica 17 (1993) 13-20 10

1

2

3

4

5

6

7

8

B,3.919000004.75758710

3.31765037

3.91900000

3.06371316

4.92494677

3.919000004.88683211

3.12414963

3.919000005.41905454

-2.06767670

3.91900000

-5.79059278

0.33237673

3.91900000-5.79709458

-0.18743919

3.91900000

0.84555973

5.73815892

3.919000000.964740205.71932823

B24.90000000

3.95490931-3.05756315

4.90000000

4.84015608

-1.25015562

4.900000003.82908511

-3.21373726

4.90000000

-0.30361999

-4.98977103

4.90000000

-1.21931066

4.84801831

4.90000000

-1.64849734

4.71937035

4.90000000

4.93718069

0.78373898

4.90000000

4.952415490.68086767

B31.030313787.97888183-0.59115408

-0.19569624

5.91293201

1.93308669

7.413169510.58749265

1.30578233

8.58081974

2.15808595-0.78958414

5.87188051

-0.49562467

-1.13376461

1.67644660

-0.41612749

-0.31451165

3.45743097

-0.96582475

0.21494968

-0.05781944

1.485368431.48685219

B<10.143525996.833809045.03362784

8.77551303

6.805501357.82551607

-2.732213333.57714394

3.33880768

-1.866823943.92495221

1.14042236

-2.78349578

-6.51139250

1.07869976

0.41347608

-10.52353198

3.18481399

-1.51222561

3.85844994

8.46291228

3.79421623

5.6088353410.66066503

9

10

11

12

13

14

15

16

Bi3.919000000.74090168-5.75260843

3.91900000

5.02835369

-2.89086460

3.919000005.26323603

-2.43716751

3.91900000

4.63028153

-3.49312638

3.91900000

-4.36007315

-3.82507531

3.91900000-5.60975474

1.47380147

3.91900000

4.74653541

3.33344275

3.919000004.238810563.95903069

B24.90000000

4.922O8620-0.87353730

4.90000000

3.67370583

3.39026334

4.900000003.36111971

3.70038840

4.90000000

4.066409422.90763039

4.90000000

2.02193037

-4.57184838

4.90000000

-2.65814506

-4.23370581

4.90000000

-1.51797840

4.76295513

4.90000000

-2.163357294.50664900

B30.102702991.18131426-1.36604783

0.62022597

7.46465198

1.25431822

0.078204486.55728872

1.75493429

1.11843459

8.076111750.28617083

1.89922819

-0.53643883

0.08648990

6.27569737

-0.28244892

1.21730553

0.51056606

0.58605307

1.08111846

8.257330441.611642381.08207460

B49.31884867-3.24071401-4.75844002

2.15722486-1.17611044

-4.98866659

3.40584298-0.85848049

-5.31140770

1.25479252-1.50044306-4.64039027

10.38741880

-6.29873348

-3.19170784

8.09930050

-10.879646860.60585867

9.78766240

5.98020127

0.16534076

5.10272935

5.9867910810.40444371

Table 1: Coordinates of points Bj, j = 1,4, in reference system Wj, for all 16 closure s of the 4-4structures.

(28). Moreover, taking into account of the twopossible assembly configurations of the tetrahe-dron, the 4-4 structure admits sixteen closures inthe complex field.

3 Case Study

The direct position analysis of a 4-4 mechanismsis reported. For a given set of actuator displace-ments, which is characterized by a given set of leglengths Lj, j = 1,6, the mechanism becomes the4-4 structure shown in Fig. 2.

Positions of points Aj, j = 1,4, and Bj, j =1,4, are known respectively in W& and Wp. Arbi-trary length unit are considered. The coordinatesof Aj in Wb are: Ax = (0,0,0), A2 = (5,0,0),A3 = (4,4,0), A4 = (5.5,-2,4.4), and the co-ordinates of Bj in Wp are B^ = (0,0,0), B2 -(6.5,0,0), B3 = (3,5,0), B4 = (-1,-3,6). Theleg lengths are: Lx = 7, L2 = 5.9, L3 = 7, L4 = 5,L5 = 5, L6 = 10.

According to the procedure presented in the pa-per, the two possible assembly configurations ofthe tetrahedron have been found each providing

eight real solutions. All solutions have been veri-fied to perform the same set of leg lengths.

The sixteen solutions are reported in Table 1,in terms of the coordinates in Wb of points Bj,(j — 1,4), of the platform.

4 Conclusions

The paper presented the direct position analysisof one type of 4-4 Stewart platform mechanismin analytical form. That is, for a given set ofactuator displacements, all the possible ways ofassembling the 4-4 structure can be determined.The geometry of the mechanism is quite general,that is, neither the base nor the platform are nec-essarily planar.

The analysis can be solved in two steps. Firsta quadratic algebraic equation provides the twopossible ways of assembling a tetrahedron-likesubstructure of the 4-4 structure, then an 8th or-der algebraic equation provides the closures of theremaining 4-4 structure. Thus, in the complexfield, the closures of the 4-4 structure resulted tobe sixteen.

20 Inf. rmatica 17 (1993) 13-20 C. Innocenti k V. Parenti-Castelli

The new theoretical result has been confirmedby a numerical example that has been reported inthe paper.

Acknowledgements

The financial support of the Ministero dellaRicerca Scientifica e Tecnologica under MPI 40%and 60% funds is gratefully acknowledged.

References

[1] Merlet J.P., "Les robots les", Hercnes, Paris,1990.

[2] Innocenti C , Parenti-Castelli V., "BasicIdeas and Recent Techniques for the An-alytical Form Solution of the Direct Posi-tion Analysis oi Fully-Parallel Mechanisms",Int. J. of Laboratory and Automation, Vol.4,1992, pp. 107-113.

[3] Griffis M., Duffy J., "A Forward Displace-ment Analysis of a Class of Stewart Plat-form", J. of Robotic Systems, Vol. 6, No. 6,1989, pp. 703- 720.

[4] Lin W., Duffy J., Griffis M., "Forward Dis-placement Analysis of the 4-4 Stewart Plat-forms", 21st ASME Mechanisms Conference,DE-Vol. 25, Sept. 16-19, 1990, Chicago, pp.263-269.

[5] Nanua P., Waldron K.J., Murthy V., "DirectKinematic Solution of a Stewart Platform",IEEE Trans. on Robotics and Automation,Vol. 6. No. 4, August 1990, pp. 438-444.

[6] Zhang C , Song S., "Kinematics of ParallelManipulators with a Positional Subchain",1990 21th ASME Mechanisms Conference,16-19 Sept., 1990 Chicago, IL, DE-Vol. 25,pp. 271-278.

[7] Zhang C , Song S., "Forward Kinematicsof a Class of Parallel (Stewart) PlatformsWith Closed-Form Solutions", 1991 IEEEInt. Conference on Robotics and Automa-tion, Sacramento, California, April, 1991, pp.2676- 2681.

[8] Merlet J.P., "An Algorithm for the For-ward Kinematics of General Parallel Manip-ulators", 5th Int. Conference on AdvancedRobotics, ICAR'91, June 19-22, 1991, Pisa,Italy, pp. 1136-1140.

Infprmatica 17 (1993) 21-34 21

TOWARDS A MATHEMATICAL THEORYOF NATURAL-LANGUAGE COMMUNICATION

Vladimir A. FomichovDepartment "Math. Provision of Information-Processing and Control Systems",Moscow Institute of Electronics and Mathematics,Bolshoy Vuzovsky pereulok, 3/12,109028 Moscow, Russia.E-mail: [email protected] (to Fomichov)Fax: 007/095/227 28 07 (to Fomichov)

Keywords: communication, formalization, integral formal semantics, knowledge, linguocybernetics,mathematical linguistics, natural language processing

Edited by: A.P. Železnikar


The need ofeffective mathematical means for solving a number oftasks associated withthe design of natural-language—processing systems (NLPSs) is argued. The basic ideasof cogDitive science concerning natural language (NL) understanding are stated. It isshown that the main popular approaches to the formalization of NLsemantics do notsatisfy the requirements of cognitive science and computer science as concerns manyimportant aspects of NL-communication.

The interdisciplinary problem of developing a mathematical theory of NL-communication (as a collection of models based on common mathematical means ofdescribing knowledge and texts' structured meanings and being useful for the design ofNLPSs) is posed.

Several principles of a new approach to the mathematical study of NL-communicationcalled Integral Formal Semantics are set forth. This approach provides, in particular,the definition of a new class of formal languages called standard K-languages. Somenew opportunities afforded by standard K-languages for modeling NL-communicationare characterized.

The conclusion is drawn that the premises have been created already for developing inlarger scope than before the researches aimed at vrorking out a mathematical theoryof NL-communication. It is noted that such a theory will be, in essence, cognitivemathematical linguistics and may be called also mathematical Iinguocybernetics takinginto account the peculiarities ofits methods and models.

1 Introdliction relational databases to computer-aided design ofcomplex technical systems with the aid of their

The work on constructing the natural-language- natural-language specifications. The great in-processing systems (NLPSs) has been carried out Ventory of diverse applications of NLPSs can befor forty years already, beginning with the first fOUnd, for instance, in Hahn (1989).systems of machine translation. A considerableprogress has been achieved during this time, and The attained progress permitted to start aNLPSs are being built now for use in a large spec- number of projects on creating highly complicatedtrum of applications, from communication with NLPSs, first of all, full-text databases (DBs),

22 Inforraatica 17 (1993) 21-34 V.A. Fomichov

telephoiies-interpreters, and computer systemsenriching and updating knowledge bases (KBs)of artificial intelligence systems (AISs) by meansof extracting information from scientific papers,text-books, patents, etc.

The first attempts in the fifties to build sys-tems of machine translation led to the discoverythat the designers of such systems knew very lit-tle about natural language (NL) and were ableto formalize only a few of its numerous regular-ities. This situation together with the need ofhigh-level algorithmic languages and translatorsfrom such languages caused the emergence andquick development of the theory of formal gram-mars, languages, and translators.

It seems that an analogical situation takes placenowadays. The complexity of some tasks like thecreation of full-text DBs is so great that many re-searchers acutely feel the necessity of developingeffective formal tools for constructing such sys-tems. In particular, Hajičova (1989) notes thatthe projects like automatic compilation of KBscannot be realized with a brute force method,but require the development of theories formal-izing intricate regularities of NL-comprehensionand the use of NL in communication. Sgall (1989)expresses a similar opinion, pointing out the im-portance of elaborating formal and implementabledescriptions of NL-semantics for solving tasks likeautomatic enriching and updating KBs.

Many aspects of creating effective mathemati-cal methods for the design of NLPSs, especially oftheir semantic components, are analyzed in Fomi-chov (1992).

At the very beginning of the nineties the mainpopular approaches to the formalization of NL-semantics were Montague Grammar and its ex-tensions, Situation Semantics, Discourse Repre-sentation Theory, Theory of Generalized Quan-tifiers (the references can be found, in particu-lar, in Fomichov (1993)), and Dynamic PredicateLogic (Groenendijk & Stokhof, 1989). All theseapproaches stem from mathematical logic and areunderlain by model-theoretic semantics.

Unfortunately, there exists a great distance be-tween the possibilities of mentioned approachesand the demands of cognitive science and com-puter science. This is shown in sections 2 and 8of Fomkhov (1993) and in sections 2 and 3 of thispaper.

Most important is that the expressive powerof each of these approaches is insufficient in or-der (a) to describe structured meanings of realdiscourses—abstracts, scientific papers, patents,etc; (b) to represent knowledge about the real-ity, in particular, to build formal descriptions ofnotions; (c) to represent goals of intelligent sys-tems; (d) to reflect in models the activity of intel-ligent systems in the course of natural-languagecommunication: such systems can pose questions,carry out various operations, etc.

Nevertheless, the stock of mathematical meansuseful for the design of NLPSs is now rather rich.The reason is that new, practically effective ap-proaches to the formal study of NL-semanticsand NL-pragmatics were elaborated beyond theframeworks (partially or completely) of enumer-ated most popular approaches.

One of such new approaches was developed un-der the LILOG-project funded by the IBM Ger-many (Herzog & Rollinger, 1991). The other ap-proach called Integral Formal Semantics (IFS) hasbeen created in Russia from the end of the sev-enties on. IFS provides powerful formal meansto describe structured meanings of sentences anddiscourses, represent knowledge and goals, buildmodels of NLPSs and of NLPSs' subsystems (seeFomitchov (1984), Fomichov (1992, 1993)).

It seems that these new approaches and theworks of some other researchers have created a"critical mass" of scientific results permitting toraise the interdisciplinary problem of developinga mathematical theory of NL-communication.

This problem is formulated in section 4 ofthe present paper. In section 5 several prin-ciples of IFS are set forth, and some impor-tant new opportunities afForded by IFS (moreexactly, by standard K-languages) for modelingNL-communication are described.

The posed problem is discussed in section 6.

2 The Context of CognitivePsychology and CognitiveLinguistics for the FormalStudy of Natural Language

The problems and achievements in the field ofconstructing NLPSs, on the one hand, and greatdifficulties on the way of formalizing regularities

TOWARDS A MATHEMATICAL THEORY... Informatica 17 (1993) 21-34 23

of NL-comprehension, on the other hand, haveevoked an increasing interest of many psycholo-gists and linguists to investigating such regulari-ties.

In linguistics a new branch is formed called cog-nitive linguistics and being a part of cognitive sci-ence. Cognitive linguists consider language "as aninstrument for organizing, processing, and con-veying information The formal structuresof languages are studied not as if they were au-tonomous, but as reflections of general conceptualorganization, categorization principles, processingmechanisms, and experiential and environmentalinfluences" (Geeraerts, 1990, p. 1).

The obtained results permitted to formulate thefollowing now widely accepted principles of NL-comprehension.

1. The meaning of a natural-language text(NL-text) is represented by means of aspecial mental language, or a languageof thought (MePcuk & Zolkovskij, 1970;Chafe, 1971; Schank, 1972; Apresyan, 1974;Fodor, 1975, 1988; Lakoff & Johnson, 1980;Wierzbicka, 1980; Johnson-Laird, 1983; Fau-connier, 1985; Lakoff, 1987; Langacker, 1987,1990; Caron, 1989).

2. People build two different (though interre-lated) mental representations of a NL-text.The first one is called by Johnson-Laird(1983) the propositional representation (PR).This representation reflects the semantic mi-crostructure of a text and is close to the text'ssurface structure.

The second representation being a mentalmodel (MM) is facultative. The MM of atext reflects the situation described in thetext. Mental models of texts are built onthe basis of both texts' PRs and diverseknowledge—about the reality, language, dis-cussed situation, and communication partic-ipants (Johnson-Laird, 1983).

3. A highly important role in building the PRsand MMs of NL-texts is played by diversecognitive models accumulated by people dur-ing the life—semantic frames, explanationsof notions' meanings, prototypical scenar-ios, social stereotypes, representations of gen-eral regularities and area-specific regulari-

ties, and other modols determining, in par-ticular, the use of metaphors and metonymy(Minsky, 1975; LakofF & Johnson, 1980;Johnson-Laird, 1983; Fauconnier, 1985; Fill-more, 1985; Seuren, 1985; Johnson, 1987;Lakoff, 1987).

4. The opinion that there exists syntax as an au-tonomous subsystem of language system hasbecame out of date. Syntax should dependon descriptions of cognitive structures, on se-mantics of NL.

Natural language understanding by peopledoesn't include the phase of constructing thepure syntactic representations of texts. Thetransition from a NL-text to its mental rep-resentation is carried out on the basis of var-ious knowledge and is of integral character(Thibadeau et al, 1982; Johnson-Laird, 1983;Seuren, 1985; LakofF, 1987; Langacker, 1987,1990; Caron, 1989; Fisher et al, 1991).

5. Semantics and pragmatics of NL are insepa-rably linked and should be studied and de-scribed by the same means (Schank et al,1985; Sgall et al, 1986).

A significant role in formulating the enumer-ated principles was played by the researches ondeveloping computer programs capable to carryout the conceptual processing of NL-texts. Thisapplies especially to the works which can be at-tributed to the semantics-oriented (or semanti-cally driven) approaches to natural language pars-ing (for more details see Hahn (1989)).

It appears that the set of principles statedabove may serve as an important reference-point for the development and comparisonof approaches to mathematical modeling NL-understanding.

24 Tnformatica 17 (1993) 21-34 V.A. Fomichov

3.1

The Restrictions of MainPopular Approaches to theFormalization ofNatural-LanguageSemantics

The standpoint of philosophy,cognitive psychology, andcognitive linguistics

The shortcomings of main known approaches tothe formal study of NL-semantics were felt in theeighties by many philosophers, psychologists, andlinguists.

Basic philosophical ideas of model-theoretic se-mantics were criticized, in particular, by Putnam(1981), Johnson-Laird (1983), Fillmore (1985),Seuren (1985, 1986), and LakofF(1987).

The main approaches to the formalization ofNL-semantics popular in the eighties—MontagueGrammar and its extensions, Situation Seman-tics, Discourse Representation Theory, and The-ory of Generalized Quantifiers are strongly con-nected with traditions of mathematical logic, ofmodel-theoretic semantics and do not provide for-mal means permitting to model the processes ofNL-comprehension in correspondence with enu-merated above principles of modern cognitive sci-ence.

In particular, these approaches do not affordeffective formal tools to build (a) semantic rep-resentations of arbitrary discourses (e.g., of dis-courses with references to the meaning of frag-ments being sentences or larger parts of texts),(b) diverse cognitive models, for instance, expla-nations of notions' meanings, representations ofsemantic frames, (c) descriptions of sets, relationsand operations on sets.

Besides, these approaches are oriented towardsregarding assertions. However, it is important tostudy also goals, promises, advises, commands,questions.

The dominant paradigm of describing surfacestructure of sentences separately from describ-ing semantic structure (stemming from the pio-neer works of Montague (1970, 1974a, 1974b))contradicts to one of key principles of cogni-tive linguistics—the principle assuming the de-pendency of syntax on semantics.

Highly emotionally the feeling of dissatisfac-

tion with the possibilities of the main popularapproaches to the formalization of NL-semanticswas put into words by Seuren (1985, 1986). Inparticular, P. Seuren expressed the opinion thatthe majority of studies on the formalization ofNL-semanlics was carried out by researchers in-terested, first of all, in demonstrating the use offormal tools possessed by them, but not in de-veloping forrnal means permitting to model themechanisms of NL-comprehension.

As it is known, ecology studies the living beingsin their natural environment. In Seuren (1986)the need of new, adequate, ecological approachesto studying the regularities of NL-comprehensionis advocated. Many reasonings and observationsuseful for working out ecological approaches tothe formalization of discourses' semantics can befound in Seuren (1985). In this monograph a pe-culiar attention is given to the questions of ex-pressing and discerning the presuppositions of dis-courses, and the so called Presuppositional Propo-sitional Calculus is suggested.

In the second half of the eighties a number ofnew results concerning the formalization of NL-semantics was obtained. Let us mention herethe approach of Saint-Dizier (1986) motivatedby the tasks of logic programming, the results ofCresswell (1985) and Chierchia (1989) on describ-ing sentences' structured meanings, the theoryof situation schemata (Fenstad et al, 1987), Dy-namic Semantics in the forms of Dynamic Pred-icate Logic (Groenendijk & Stokhof, 1989) andDynamic Montague Grammar (Groenendijk &Stokhof, 1990; Dekker, 1990).

Unfortunately, the restrictions pointed above inthis section apply also to these new approaches. Itshould be added that Chierchia (1989) describesstructured meanings of some sentences with in-finitives. But the expressive power of semanticformulae corresponding to such sentences is verylittle in comparison with the complexity of realdiscourses from scientific papers, text-books, en-cyclopedic dictionaries, legal sources, etc.

Thus, the approaches mentioned in this sectiondo not provide efiective and widely-applicable for-mal tools for modeling NL-understanding in ac-cordance with stated principles of cognitive psy-chology and cognitive linguistics.

The lack of such means for modeling NL-understanding can be seen also from the most


complete text-book on mathematical methods inlinguistics (Partee, ter Meulen, & Wall, 1990).

3.2 The standpoint of computerscience

One can't say that all approaches to the formal-ization of NL-semantics mentioned in the prece-dent subsection are not connected witli the prac-tice of designing NLPSs. There are publica-tions, for example, on using for the design ofNLPSs Montague Grammar (MG) in modifiedforms (ClifFord, 1988; Hirst, 1988; Scmbok fcvan Rijsbergen, 1990), Situation Semantics (Ya-sukawa et al, 1988), Discourse RepresentationTheory (Herzog & Rollinger, 1991).

The language of intensional logic provided byMG is used also in Generalized Phrase Struc-ture Grammars (Gazdar, Klein, Pullum, & Sag,1985) for describing semantic interpretations ofsentences. Such grammars have found a numberof applications to natural language processing.

Nevertheless, these and other approaches men-tioned in this section possess a number of im-portant shortcomings as concerns applying formalmethods to the design of NLPSs and to develop-ing the theory of NLPSs.

The demands of diverse application domains tothe means of formal describing natural languagemay difFer. That is why let distinguish for fur-ther analysis the following groups of applicationdomains:

1. Natural-language interfaces to databases,knowledge bases, autonomous robots.

2. Full-text databases; computer systems au-tomatically forming and updating knowl-edge bases of artificial intelligence systems bymeans of extracting information from scien-tific papers, text-books, e tc , in particular,automatic abstracting systems.

3. Such subsystems of automatized program-ming systems which are destined for trans-forming the NL-specifications of tasks intothe formal specifications for the further syn-thesis of programs; such similar subsys-tems of CAD-systems which are destinedfor transforming the NL-specifications of de-signed technical objects into the formal spec-ifications.

Obviously, the enumerated application domainsrepresent only a part of all possible domains,where the development and use of NLPSs are ac-tual. Much larger list of such domains can befound, in particular, in Hahn (1989).

However, for our purpose it is sufficient to con-sider only mentioned important domains of ap-plying NLPSs. The analysis of formal means forthe study of NL needed for these domains will al-low us to get a rather complete list of demands tothe formal theories of NL which should be satis-fied by useful for practice and widely-applicablemathematical tools of studying NL-semantics andNL-pragmatics.

Let us regard for each distinguished group ofapplications the most essential restrictions of enu-merated approaches in this section to the formalstudy of NL-semantics.

3.2.1 Group 1

Semantics-oriented, or semantically driven NL-interfaces work in the following way (for moredetails see Hahn (1989)). They transform a NL-input (or at first its fragment) into a formal struc-ture reflecting the meaning of this input (or themeaning of some inpufs fragment) and called se-mantic representation (SR) of the input or inpufsfragment. Then the SR is used (possibly, aftertransforming into a problem-oriented representa-tion) for working a plan of the reaction to theinput with respect to a knowledge base, and af-ter this some reaction is produced. The reactionsmay be highly diverse: AISs can pose questions,fulfill calculations, search required information,transport things, etc.

For constructing NL-interfaces in accordancewith these principles, the following shortcomingsof MG and its extensions', including DynamicMontague Grammar, of Situation Semantics, Dis-course Representation Theory, Theory of Gener-alized Quantifiers, Dynamic Predicate Logic, andof other approaches mentioned in this section areimportant.

1. The effective formal means of describingknowledge fragments, the structure of KBsare not provided.

2. There are no sufficiently powerful and flexibleformal means to describe surface and seman-

26 Informatica 17 (1993) 21-34 V.A. Fomichov

tic structures of questions and commands ex-pressed by complicated NL-utterances.

3. There are no sufficiently powerful and flexi-ble formal means to represent surface and se-mantic structures of intelligent systems' goalsformed by complicated NL-utterances.

The possibilities of intelligent systems to un-derstand the goals of communication partici-pants and to use the information about thesegoals for planning the reaction to a NL-inputare not modeled.

4. The enumerated approaches do not givethe flexible and povverful means of for-mal describing structured meanings of NL-discourses (including real discourses from sci-entific papers, legal sources, patents, etc).The means of describing structured meaningsof discourses are extremely restricted and un-satisfactory from the viewpoint of practice.In particular, discourses with references tothe meaning of sentences and larger frag-meiits of texts are not considered.

5. The existence of sentences of many typeswidely used in real life is ignored. For in-stance, the structure of the following sorts ofsentences is not studied; (a) containing ex-pressions built out of descriptions of objects,sets, notions, events, etc. by means of logicalconnectives ("Yves has bought a monographon mathematics, a text-book on chemistry,and a French-Russian dictionary"), (b) de-scribing the operations on sets ("It will beuseful to include Professor A. into the Ed-itorial Board of the journal B."), (c) withthe words 'notion' or 'term' (the latter inthe sense "a notion"), (d) with the words 're-spectively' or 'correspondingly' ("Ljubljana,Oslo, and Bratislava are capitals of Slovenia,Norway, and Slovakia, respectively").

6. The models of the correspondences betweentexts, knowledge about the rea!ity, and texts'semantic representations are not built, andthe adequate means for developing models ofthe kind are not provided.

7. The inputs of NLPSs may be incompletephrases, even separate words (e.g., the an-swers to questions in the course of a dia-logue). The interpretation of such inputs

is to be found in the context of precedentphrases and with respect to the knowledgeabout the reality and about the concrete dis-cussed situation.

However, such a capability of NL-interfacesisn't studied and isn't formally modeled.

8. The structure of metaphors and incor-rect, but understandable expressions frominput texts, the correspondences betvveenmetaphors and their meanings are not inves-tigated by formal means.

9. The same situation takes place relatively tothe formal study of metonymy—the phe-nomenon which often manifests in input textsof applied intelligent systems. As Lakoff(1987, p. 77) notes, "metonymy is one ofthe basic characteristics of cognition. It isextremely common for people to take onewell-understood or easy-to-perceive aspectof something and use it to stand either forthe thing as a whole or for some other aspectorpartofit".

10. Wilks (1990, p. 348) writes that manyNLPSs (in particular, systems of machinetranslation) do not work so as it is ex-plained by the "official" theories in publica-tions about these systems and function "insuch a way that it cannot be appropriatelydescribed by the upper-level theory at all,but requires some quite difFerent form of de-scription".

The approaches mentioned in this section donot afford the opportunities to describe ade-quately the main ways of processing informationby semantic components of NLPSs.

3.2.2 Group 2

Obviously, the restrictions 1, 3-6, and 8-10 areimportant also from the viewpoint of solving thetasks like the development of full-text DBs.

The restriction 7 should be replaced by a sim-ilar restriction, since fragments of discourses per-taining to business, technology, science, etc. maybe incomplete, elliptical phrases.

The following restriction is to be pointed outadditionally: the semantic structure of discourseswith promises (protocols, contracts often include

TOVVARDS A MATHEMATICAL THEORY... Informatica 17 (1993) 21-34 27

such discourses), interrelations betvveen surfaceand semantic structures are not studied and mod-eled.

Let us make some remarks. The possibility tobuild semantic representations (SRs) of compli-cated goals is necessary, in particular, for the de-velopment of algorithms permitting to find refer-ents of such expressions as "this success", "thisfailure", etc.

Yet twenty years ago Wilks (1973, p. 116)noted that "any adequate logic must contain adictionary or its equivalent if it is to handle any-thing more than terms with naive denotationssuch as 'chair' ".

However, all approaches to the formalizationof NL-semantics enumerated in this section donot take into account the existence and roles ofvarious semantic dictionaries. Because of this,in particular, reason there is no opportunity tomodel the correspondence between texts, knowl-edge about the reality, and SRs of texts.

At first sight, the demands to the means of de-scribing structured meanings of discourses and tothe models of the correspondences between texts,knowledge, and SRs of texts are much higher forthe second group of applications than for the firstone.

Nevertheless, it is not excluded that the jointfuture work of philosophers, linguists, specialistson computer science, and mathematicians willshow that such demands are in fact very simi-lar or the same for these two groups of NLPSs'applications.

3.2.3 Group 3

Additionally to the shortcomings important forthe groups 1 and 2, the following restrictionsshould be mentioned.

1. There are no efFective formal means to rep-resent structured meanings of NL-discoursesdescribing algorithms, methods of solving di-verse tasks. In particular, there are no ade-quate formal means to describe on a semanticlevel the operations with sets.

2. The opportunity to build semantic represen-tations of complicated notions' descriptions(from encyclopedic dictionaries, etc.) is notafforded.

It appears that the collection of restrictionsstated above provides a useful reference-pointfor further enriching the stock of means andmodels for the mathematical study of NL-communication.

4 The InterdisciplinaryProblem of Developing aMathematical Theory ofNatural-LanguageCommunication

In situation when known formal methods ofstudying NL-semantics proved to be ineffectivefor solving many actual tasks of designing NLPSs,a number of researchers in diverse countriespointed out at the necessity to search new math-ematical ways of modeling NL-communication.

According to the opinion of Peregrin (1990),additional efForts are to be undertaken in orderto develop powerful formal means for describingthe regularities of NL-understanding. For this afull-fledged linguistic analysis of NL-phenomenashould be carried out.

Habel (1988) notes the actuality of creatingadequate mathematical foundations of computa-tional linguistics and underlines the necessity tomodel the processes of NL-communication on thebasis of formal methods and theories of cognitivescience.

Fenstad & Lonning (1990, p. 70) posed the taskof working out adequate formal methods for Com-putational Semantics—"a field of study which liesat the intefsection of three disciplines: linguis-tics, logic, and computer science". Such methodsshould permit, in particular, to establish interre-lations between pictorial data and semantic con-tent of a document.

A. P. Ershov, the prominent Russian theoreti-cian of programming, raised the problem of de-veloping a formal model of Russian language (Er-shov, 1986).

It is very interesting how close the ideas ofSeuren (1986) about the need of ecological ap-proaches to the formal study of NL are to thefollowing words of A. P. Ershov published in thesame 1986: "We want as deeply as it is possibleto get to know the nature of language and, in par-ticular, of Russian language. A model of Russian


language should become one of manifestations ofthis knowledge. It is to be a formal system whichshould be adequate and equal-voluminous to theliving organism of language, but in the same timeit should be anatomically prepared, decomposod,accessible for the observation, study, and modifi-cation" (Ershov, 1986, p. 12).

The need of good formal models for good engi-neering in the field of designing NLPSs is arguedin (Joshi, 1989).

Thus, the idea of developing new formal meth-ods destined for the study of NL-semantics andadequate to the complexity of NL has been find-ing gradually supporters in various countries.

A number of such formal methods useful forthe design of NLPSs was developed during theeighties. Let us mention now only three examples.

Appelt & Kronfeld (1987) elaborated a formaltheory of referring within the framework of a gen-eral theory of speech acts and rationality.

A large volume of studies aimed at creatinga formal theory of natural-language question-answering systems was carried out in Germanyunder the known project LILOG funded by theIBM Germany. One of the goals of this projectwas to determine formal languages permitting toreflect on a semantic level a wide range of NLphenomena and serving as target languages in or-der to represent (in a logical form) the informa-tion extracted from texts in German (Herzog &Rollinger, 1991).

A new, powerful approach to the mathematicalstudy of NL-semantics and NL-pragmatics calledIntegral Formal Semantics (IFS) has been devel-oped in Moscow from the end of the seventies on.This approach provides, in particular, the highlyefFective mathematical means to describe struc-tured meanings of sentences and discourses, torepresent knowledge and goals of intelligent sys-tems (see the next section).

It seems that the demands of practice, on theone hand, and the results obtained in mathemat-ical linguistics, computer science, and cognitivescience, on the other hand, permit and force us togo to a new level of mathematical studying natu-ral language and to raise the problem of develop-ing a mathematical theory of NL-communication.

It is the problem of building a collection ofmathematical models based on common math-ematical means and useful for the design of

NLPSs. The collection of such models should per-mit to reflect and to study all the regularities ofNL-communicat.ion which were mentioned in theprecedent section.

Besides, the new theory should help to solveat least the tasks pertaining to three groups ofapplication domains distinguished above.

Naturally, such a new mathematical theory isto be developed jointly by philosophers, psychol-ogists, linguists, specialists on computer science,and mathematicians. Hence the posed problem isinterdisciplinary.

5 Shortly About IntegralFormal Semantics

The analysis carried out in the sections 2 and 3may create the impression that a long way shouldbe gone yet in order to get mathematical toolspermitting to overcome stated restrictions and todevelop complicated models useful for the designof NLPSs and for representing hypotheses in cog-nitive psychology and cognitive linguistics.

Fortunately, the real situation is not so pes-simistic, and in fact powerful mathematical meansfor modeling NL-communication were elaboratedyet in the eighties.

Such means are provided by an originalapproach to the mathematical study of Nl-communication called Integral Formal Semantics(IFS). This term is introduced, in particular, inFomichov (1993).

The first version of IFS was developed in 1978- 1983. By 1984 a great volume of theoretical re-sults was obtained. A part of these results is re-flected in Fomichov (1978,1980,1981), Fomitchov(1983, 1984). One can find additional referencesin Fomichov (1992).

It is interesting that approximately in the sametime, in 1981-1983, the first publications on theTheory of Generalized Quantifiers (Barwise &Cooper, 1981), Discourse Representation The-ory (Kamp, 1981), Lexical-Functional Grammars(Kaplan & Bresnan, 1982), and Situation Seman-tics (Barvvise & Perry, 1983) appeared.

Strong impulses to the development of IFS weregiven by the well known works of T. Winograd onthe program understanding NL, W. A. Woods onthe system LUNAR, R. Schank on the ConceptualDependency theory, Y. Wilks on machine trans-

TOWARDS A MATHEMATlCAL THEORY... Informatica 17 (1993) 21-34 29

lation published in the first half of the seventies,by the paper of D. Bobrow and T. Winograd onthe knowledge representation language KRL, andby the main ideas of the known theory of linguis-tic models "Meaning-Text" developed in Moscow(MePcuk & Zolkovskij, 1970; Apresyan, 1974).

The set of basic philosophical ideas of IFS in-cludes, in particular, the following principles.

1. The chief task of researches on the formal-ization of NL-semantics is to be the elabora-tion of formal models of NLPSs and of suchsubsystems of NLPSs which belong to the socalled semantic components of NLPSs.

The structure of such models of a number ofkinds is suggested in Fomichov (1992); theadditional references can be found in Fomi-chov (1993).

2. The studies are to be oriented towards con-sidering not only assertions, but also com-mands and questions which may be inputs ofNLPSs.

3. The basis of researches is to be a formalmodel reflecting many peculiarities of seman-tic structures of sentences and discourses ofarbitrary great length and providing the de-scription of some class of formal languagesconvenient for building semantic representa-tions (SRs) of NL-texts in a large spectrumof applications and on different levels of rep-resentation.

Two main variants of such a model are de-veloped. The first variant is provided by thetheory of free S-models, S-calculuses and S-languages, T-calculuses and T-languages (theSTCL-theory). This theory afforded already in1981 - 1983 a really ecological approach to theformalization of NL-semantics providing powerfuland convenient for linguists mathematical meansfor representing both structured meanings of NL-texts and knowledge about the reality (Fomichov,1981; Fomitchov, 1983, 1984).

The expressive power of S-languages and T -languages is very great (see the section 5 of Fomi-chov (1992)) and essentially exceeds, for example,the expressive power of Discourse RepresentationTheory.

The basic ideas of the STCL-theory were re-ported at the First symposium of the Interna-tional Federation of Automatic Control (IFAC)

on Artificial Intelligence which was held in 1983and were published in the proceedings of thissymposium (it should be noted that the paperFomitchov (1984) is a considerably abridged ver-sion of Fomitchov (1983)).

The STCL-theory became the starting pointfor developing the theory of K-calculuses, al-gebraic systems of conceptual syntax, and K-languages (the KCL-theory) being nowadays thecentral component of IFS. The foundations of theKCL-theory are stated in Fomichov (1988a) (itis simultaneously a monograph and a text-book)and also, in particular, in Fomichov (1988b, 1992,1993).

The KCL-theory provides the definition of aclass of formulae permitting (a) to describe struc-tured meanings of complicated sentences and dis-courses and (b) to build the representations ofdiverse cognitive structures.

For instance, this theory allows us to describein a mathematical way structured meanings of:

a. the expressions "a group of seven students","a man in the age from 21 to 27 years beinga chemist or a biologist" (Fomichov, 1992);

b. sentences "Namur and Lyon are the cities ofBelgium and France, correspondingly", "Na-mur, Leuven, Gent belong to the cities ofBelgium, and London does not belong tothe cities of Belgium, or France, or Swe-den", "There exists a country in Europewith the number of cities greater than 15","Belongs Gent to the cities of Belgium?""To whom did P. Carpenter phoned at 4:30p.m.?" (Fomichov, 1992), "The terms 'cy-tosine', 'ischemia', and 'insulin' are used ingenetics, cardiology, and endocrinology, re-spectively" (Fomichov, 1993);

c. discourses

"The chemical action of a current consists in thefollowing: for some solutions of acids (salts, alka-lis), by passing an electrical current across sucha solution one can observe isolation of the sub-stances contained in the solution and laying asidethese substances on electrodes plunged into thissolution.

For example, by passing a current across a solu-tion of the blue vitriol (CuSO^) a pure copper willbe isolated on the negatively charged electrode.


One use this to obtain pure metals" (Fomichov,1992),

"An adenine base on one DNA strand links onlywith a thymine base of the opposing DNA strand.Similarly, a cytosine base links only with a gua-nine base of the opposite DNA strand" (Fomi-chov, 1993).

The referenced papers contain semantic repre-sentations of these NL-expressions built with thehelp of the so called standard K-languages.

Besides, standard K-languages afford large op-portunities to represent structured meanings ofnotions' definitions. In particular, semantic rep-resentations of the following notions' definitionsare built in Fomichov (1993):

"Thrombin is an enzyme which helps to convertfibrinogen to fibrin during coagulation",

"Sphygmomanometer is instrument destined tomeasure blood pressure",

"A genotype is a collection of all genes locatedin chromosomes of an organism",

"Type A blood group are persons who possesstype A isoantigen on red blood cells and anti-Bagglutinin in plasma",

"Messenger RNA is molecule which is formedfrom DNA and transfers the genetic code to thecytoplasm where protein synthesis occurs",

"Tympanic membrane is a membrane betweenthe outer and middle ear. Surgical evacuation ofthe purulent material through the tympanic mem-brane prevents hearing loss and mastoiditis. Thisprocedure is called a myringotomy".

It should be mentioned that SRs of all texts ad-duced here (except the texts with the word 'cor-respondingly' or with the expression 'the terms')can be constructed also by means of S-languagesof type 5 provided by the STCL-theory and char-acterized in (Fomitchov, 1983, 1984).

Hence the expressive power both of standardK-languages and of S-languages of type 5 consid-erably exceeds the expressive possibilities of otherapproaches to the formalization of Nl-semanticsdiscussed above.

The integral character of IFS manifests in af-fording opportunities to describe both structuredmeanings of texts and knowledge about the real-ity.

Some important possibilities of S-languages torepresent knowledge are described in Fomitchov(1983, 1984). K-languages also provide simi-

lar possibilities (see the section 9 of Fomichov(1992)).

In the subsection 9.4 of Fomichov (1992) someopportunities of recording NL-communication bymeans of standard K-languages are explained.I.e., it is shown how it is possible to representin a formal manner the actions carried out by in-telligent systems in the course-of communication.

The subsection 9.5. of Fomichov (1992) is de-voted to describing with the aid of standard K-languages semantic-syntactic information associ-ated with words and fixed word combinations.

It may be added that IFS gives also two vari-ants of a complicated and useful for practicemathematical model of the correspondence "Text(phrase or discourse) + Knowledge about reality—> Semantic representation of atext". Therefer-ences and brief information about this model canbe found in Fomichov (1992). This model satis-fies the demand of cognitive linguistics about thedependency of syntax on semantics.

6 Discussion

It should be noted that the problem of creat-ing a mathematical theory of NL-communicationwas raised in Fomichov (1978, 1980, 1981) andFomitchov (1983, 1984). These publications ap-pear to be the first ones (or belong to the firstones), where this problem is formulated.

In Fomichov (1988b, 1992) the conclusion isdrawn that the KCL-theory provides the premisesfor starting in more large scope than before thedevelopment of a mathematical theory of NL-communication and gives good chances to workout such a theory.

Taking into account the experience obtained inthe framework of Integral Formal Semantics, itappears to be worth-while to base a mathemati-cal theory of NL-communication on the powerful,widely-applicable or universal formal means ofdescribing knowledge and discourses' structuredmeanings.

Such means, on the one hand, can be used forbuilding models of knowledge bases and bases ofgoals. On the other hand, formal representationsof texts' structured meanings can be used for de-scribing, firstly, surface structures of texts and,secondly, the correspondences "Text + Knowl-edge —• Semantic representation of a text".

TOWARDS A MATHEMATICAL THEORY;.. Informatica 17 (1993) 21-34 31

The idea of describing surface structures con-sists in getting these structures by means of in-deterministic transformations of semantic struc-tures (transformations are to be based on varioussemantic-syntactic dictionaries).

It is the idea absolutely contradictory to the ap-proach of Montague Grammar, but correspondingto the principle of cognitive linguistics about thedependency of syntax on semantics. It seems thatthis idea points out at the only effective way ofdescribing surface structures of scientific papers,patents, etc. and is in a good agreement withthe principles of designing the semantics-orientedNLPSs stated by Hahn (1989).

A mathematical theory of NL-communicationshould be based on the ideas of cognitive linguis-tics. Hence such a theory will be, in essence, andmay be called cognitive mathematical linguistics.

Since this hypothetical new theory will describethe activity of intelligent systems having knowl-edge bases and goals, the new theory may becalled also mathematical linguocybernetics.

Železnikar (1988a, 1988b, 1989) expresses theopinioii that computer systems will be graduallyreplaced in the future by information machines.Natural languages are the most important meansof storing and conveying information.

That is why it appears that the creation ofa domain-independent and realization-indepen-dent mathematical theory of NL-communicationwill contribute essentially to developing a theoryof information machines.

7 Conclusions

It may be hoped that this paper will help tojoin the efforts of philosophers, psychologists, lin-guists, specialists on computer science, and, nat-urally, mathematicians and to work out a for-mal theory of natural-language communication(or cognitive mathematical linguistics, or math-ematical linguocybernetics).

Quite good premises for speeding-up the stud-ies in this direction creates Integral FormalSemantics—a new approach to the mathemat-ical investigation of NL-semantics and NL-pragmatics developed in Russia. The task ofdemonstrating all these premises goes far beyondthe scope of the present paper and will be thesubject of the future research work.

It seems that the creation of such a new theorywill be of high significance for the development ofInformatics.

References

[1] Appelt, D. & Kronfeld, A. (1987). A compu-tational model of referring. In Proc. of theTenth Int. Joint Conf. on Art. Intelligence.Milan, 640-647.

[2] Apresyan, Yu. D. (1974). Lexical Semantics.Synonymic Means of Language. Moscow:Nauka Publ. (in Russian).

[3] Barwise, J. k Cooper, R. (1981). General-ized quantifiers and Natural Language. Lin-guistics & Philosophy, 4 (2), 159-219.

[4] Barwise, J. & Perry, J. (1983). Situations andAttitudes. Cambridge, MA: The MIT Press.

[5] Caron, J. (1989). Precis de Psycholin-guistique. Paris: Presses Universitaires deFrance.

[6] Chafe, W. L. (1971). Meaning and the Struc-ture of Language. Chicago and London: TheUniversity of Chicago Press.

[7] Chierchia, G. (1989). Structured meanings,thematic roles and control. In Chierchia, G.,Partee, B. H., & Turner, R. (Eds.), Proper-ties, Types and Meaning, 2. Semantic Issues.Dordrecht: Kluvver Acad. Publ., 131-166.

[8] ClifFord, J. (1988). Natural language query-ing of historical data bases. ComputationalLinguistics, 14 (4), 10-34.

[9] Cresswell, M. J. (1985). Structured Mean-ings: the Semantics of Propositional Atti-tudes. Cambridge, MA, London: The MITPress.

[10] Dekker, P. (1990). The scope of negationin discourse: towards a flexible dynamicMontague grammar. In Quantification andAnaphora 1. DYANA. Deliverable R2.2.A.July 1990. Edinburgh, 79-134.

[11] Ershov, A. P. (1986). Machine fund of Rus-sian language: An external task statement.In Karaulov, Yu. N. (Ed.), Machine Fund

32 Informatica 17 (1993) 21-34 V.A. FomKhov

of Russian Language: Ideas and Opinions.Moscow: Nauka Publ., 7-12 (in Russian).

[12] Fenstad, J. E., Halvorsen, P.-K., Langholm,T., van Benthem, J. (1987). Situations, Lan-guage and Logic. Dordrecht: D. Reidel.

[13] Fenstad, J. E. & Lonning, J. T. (1990). Com-putational semantics: steps towards "Intelli-gent" Text Processing. In Studer, R. (Ed.),Natural Language and Logic. Berlin, e tc :Springer-Verlag, 70-93.

[14] Fillmore, C. (1985). Frames and the seman-tics of understanding. Quaderni di Seman-tica, 6, 222-253.

[15] Fisher, C , Gleitman, H., & Gleitman, L. R.(1991). On the semantic content of subcate-gorization frames. Cognitive Psychology, 23,331-392.

[16] Fodor, J. A. (1975). The Language ofThought. New York: Crowell.

[17] Fodor, J. A. (1988). Psychosemantics: TheProblem of Meaning in the Philosophy ofMind. 2nd print. Cambridge, MA, London:The MIT Press.

[18] Fomichov, V. A. (1978). A mathematicalmodel of natural-language texts' seman-tic analysis. In Theses of reports of theSecond sc.-techn. conf. of young special-ists of VNTITSentr. Moscow: All-UnionScientific-Technical Information Centre (VN-TITSentr), 58-60 (in Russian).

[19] Fomichov, V. A. (1980). Some principles ofmathematical describing natural language'ssubsets. In A. K. Ailamazyan (Ed.), Reportsof lst and 2nd sc.-techn. conf. of young spe-cialists. Informational Processes and TheirAutomatization. Moscow, VNTITSentr, 125-136 (in Russian).

[20] Fomichov, V. A. (1981). To the theory oflogic-algebraic modelling the speech-formingmechanisms of conceptual level. 1. Taskstatement and the idea of an approach toits resolving. Moscow Institute of ElectronicEngineering, 85 pp. - Paper deposited inAUISTI of Ac. Sc. USSR (AU-Union Insti-tute of Scientific and Technical Information;

in Russian - Vsesoyuzny Institut Nauch-noy i Teknicheskoy Informatsii, or VINITI)27/10/81, No. 4939-81 Dep. - The abstract2B1386DEP in the Abstracts Journal "Math-ematika" of AUISTI, 1982, No. 2 (in Rus-sian).

[21] Fomitchov, V. A. (1983). Formal systemsfor natural language man-machine inter-action modelling. In International Sympo-sium on Artificial Intelligence IFAC, Vol. 1.Leningrad: Ac. Sc. USSR, 223-243 (in twovariants - in Russian and in English). Paperpresented at the First Symposium of the Int.Federation of Automatic Control on Artifi-cial Intelligence.

[22] Fomitchov, V. A. (1984). Formal systems fornatural language man-machine interactionmodelling. In V. M. Ponomaryov (Ed.), Ar-tificial Intelligence. Proceedings of the IFACSymposium. Leningrad, USSR, 4 - 6 October1983 (IFAC Proc. Series, 1984, No. 9). Ox-ford, New York, e tc : Pergamon Press, 203-207.

[23] Fomichov, V. A. (1988a). Representing In-formation by Means of K-calculuses: Text-book. Moscow: The Moscow Institute ofElectronic Engineering (MIEE) Press (inRussian).

[24] Fomichov, V. A. (1988b). About themeans of constructing natural-language-communication mathematical theory. In V.N. Markov (Ed.), Mathematical provision ofcalculating, informational, and control sys-tems. Moscow: The MIEE Press, 21-25 (inRussian).

[25] Fomichov, V. (1992). Mathematical mod-els of natural-language-processing systems ascybernetic models of a new kind. Cybernet-ica (Belgium), XXXV (1), 63-91.

[26] Fomichov, V. A. (1993). K-calculuses and K-languages as powerful formal means to designintelligent systems processing medical texts.Cybernetica, XXXVI (1).

[27] Gazdar, G., Klein, E., Pullum, G., & Sag, I.(1985). Generaiized Phrase Structure Gram-mar. Oxford: Blackwell.


[28] Geeraerts, D. (1990). Editorial Statement.Cognitive Linguistics, 1 (1), 1-3.

[29] Groenendijk, J. & Stokhof, M. (1989). Dy-namic Predicate Logic, tovvards a composi-tional, non-representational semantics of dis-course. The University of Amsterdam. TheITLI Prepublication Series, LP-89-02.

[30] Groenendijk, J. & Stokhof, M. (1990). Dy-namic Montague Grammar. In Qu.antifica-tion and Anaphora I. DYANA. Dyiiamic In-terpretation of Natural Language. ESPRITBasic Research Action BR3175. DeliverableR2.2.A. July 1990, 1-35.

[31] Habel (1988). Kog-nitionswissenschaft als Grundlage der Com-puterlinguistik. In Computerlinguistik undihre theoretischen Grundlagen. Berlin e tc :Springer-Verlag, 204-209.

[32] Hahn, U. (1989). Making understanders outof parsers: semantically driven parsing as akey concept for realistic text understandingapplications. International Journal of Intelli-gent Systems, 4 (3), 345-385.

[33] Hajičova, E. (1989). The introduction to thespecial issue of the journal. Computers andArtificial Intelligence, 8 (5).

[34] Herzog, 0 . & Rollinger, C.-R. (1991) (Eds.).Text Understanding in LILOG. IntegratingComputational Linguistics and Artificial In-telligence. Final Report on the IBM Ger-many LILOG-Project. Berlin e t c : Springer-Verlag.

[35] Hirst, G. (1988). Semantic interpretation andambiguity. Artificial Intelligence, 34, 131-177.

[36] Johnson, M. (1987). The Body in the Mind:The Bodily Basis of Meaning, Imagination,and Reason. Chicago: The University ofChicago Press.

[37] Johnson-Laird, P. (1983). Mental Models.Cambridge, MA: Harvard University Press.

[38] Joshi, A. K. (1989). Formal and computa-tional models. The Prague Bulletin of Math-ematical Linguistics, 51, 5-9.

[39] Kamp, H. (1981). A theory of truth and se-mantic representation. In J. Groenendijk, T.Janssen, & M. Stokhof (Eds.), Formal Meth-ods in the Study of Natural Language. Part1. Amsterdam: Mathematical Centre, 227-322.

[40] Kaplan, R. M. & Bresnan, J. (1982). Lexical-Functional Grammar: a formal system forgrammatical representation. In J. Bresnan(Ed.), The Mental Representation of Gram-matical Relations. Cambridge, MA: TheMIT Press, 173-281.

[41] LakofF, G. & Johnson, M. (1980). MetaphorsWe Live By. Chicago: The University ofChicago Press.

[42] Lakoff, G. (1987). Women, Fire and Dan-gerous Things: What Categories Revealabout the Mind. Cliicago: The University ofChicago Press.

[43] Langacker, R. W. (1987). Foundations ofCognitive Grammar. Vol. 1. Theoretical Pre-requisites. Stanford, CA: Stanford UniversityPress.

[44] Langacker, R. W. (1990). Concept, Image,and Symbol. The Cognitive Basis of Gram-mar. Berlin, New York: Mouton de Gruyter.

[45] Mel'cuk, I. & Zolkovskij, A. (1970). Towardsa functioning meaning-text model of lan-guage. Linguistics, 57, 10-47.

[46] Minsky, M. (1975). Framevvork for represent-ing knowledge. In The Psychology of Com-puter Vision, Winston, P. H. (Ed.). N. Y.:Mc Graw-Hill book comp.

[47] Montague, R. (1970). Universal Grammar.Theoria, 36, 373-398.

[48] Montague, R. (1974a). English as a formallanguage. In Formal Philosophy. Selected pa-pers of Richard Montague, Thomason, H.(Ed.), New Haven and London: Yale Uni-versity Press, 188-221.

[49] Montague, R. (1974b). The proper treatmentof quantification in ordinary English. Ibid.,247-270.


[50] Partee, B. H., ter Meulen, A., & Wall, R.E. (1990). Mathematical Methods in Linguis-tics. Dordrecht e tc : Kluwer.

[51] Peregrin, J. (1990). On a logical formaliza-tion of natural language. Kybernetika, 26 (4),327-341.

[52] Putnam, H. (1981). Reason, Truth, andHistory. Cambridge: Cambridge UniversityPress.

[53] Saint-Dizier, P. (1986). An approach tonatural-language semantics in Logic Pro-gramming. Journal of Logic Programming, 3(4), 329-356.

[54] Sembok, T. M. T. & van Rijsbergen, C. J.(1990). A simple logical-linguistic documentretrieval system. Information Processing &Management, 26 (1), 111-134.

[55] Seuren, P. A. M. (1985). Discourse Seman-tics. Oxford: Basil Blackwell.

[56] Seuren, P. A. M. (1986). Formal theory andthe ecology of language. Theoretical Linguis-tics, 13, 1-18.

[57] Schank, R. (1972). Conceptual Dependency:A theory of natural language understanding.Cognitive Psychology, 3 (4).

[58] Schank, R., Birnbaum, L., & Mey, J.(1985). Integrating semantics and pragmat-ics. Quaderni di Semantica, VI, No. 2.

[59] Sgall, P., Hajicova, E., & Panevova, J.(1986). The Meaning of the Sentence in ItsSemantic and Pragmatic Aspects. Pragueand Dordrecht: Academia and D. Reidel.

[60] Sgall, P. (1989). The tasks of semantics andthe perspectives of computers. Computersand Artificial Intelligence, 8 (5), 403-421.

[61] Thibadeau, R., Just, M. A., & Carpenter,P. A. (1982). A model of the time course andcontent of reading. Cognitive Science, 6, 157-203.

[62] Wierzbicka, A. (1980). Lingua Mentalis. TheSemantics of Natural Language. Sydney, e tc :Academic Press.

[63] Wilks, Y. (1973). An artificial intelligenceapproach to machine translation. In Schank,R. & Colby, K. (Eds.), Computer Models ofThought and Language. San Francisco: Free-man, 114-151.

[64] Wilks, Y. (1990). Form and content in se-mantics. Synthese, 82 (3), 329-351.

[65] Yasukawa, H., Suzuki, H., & Noguchi, N.(1988). Knowledge representation languagebased on Situation Theory. In K. Fuchiand L. Kott (Eds.), Programming of Fu-ture Generation Computers II: Proc. of the2nd Franco-Japanese Symp. Amsterdam e tc :North-HoUand, 431-460.

[66] Železnikar, A. P. (1988a). Principles of Infor-mation. Cybernetica, XXXI (2), 99-122.

[67] Železnikar, A. P. (1988b). Information De-terminations I. Cybernetica, XXXI (3), 181-213.

[68] Železnikar, A. P. (1989). Information Deter-minations II. Cybernetica, XXXII (1), 5-44.

Informatica 17 (1993) 35-40 35

ALPHA AXP OVERVIEW

Jurij Šilc+, Borut Robičt and Jože Buh*

t Jožef Stefan InstituteLaboratory for Computer Architectures, Jamova 39, Ljubljana, SloveniaPhone: +38 61 159 199, Fax: +38 61 161 029E-mail: [email protected] or [email protected]

* EuroComputer SystemsDigital Authorized Representative, Vojkova 50, Ljubljana, SloveniaPhone: +38 61 182 500, Fax: +38 61 181 005

Keywords: Alpha AXP architecture, RISC architecture, multiple instruction issue, shared-memorymultiprocessing, overview

Edited by: Marco Botta


The paper describes the Alpha, AXP architecture and some existing implementations. Itis a true 64-bit BJSC architecture supporting multiple instruction issue, shared-memorymultiprocessing, and several today's leading operating system environments. The firstAlpha AXP microprocessor DECchip 21064 and several hardware products using it arealso briefly described.

1 IntroductionAs the 20th Century draws to a clone, more andmore computer power is being needed to drive ourextremely complex applications. The applicationof computing took us from mainframes and 16-bit minicomputers of the early 1970s to the 32-bit microprocessors of the mid-1980s, in the ageof desktop computing with windowing user inter-faces. What will stimulate the next technologi-cal leap that brings us applications of the future?One of the answers might be an advanced 64-bitRISC architecture.

In late 1970s DEC introduced the VAX ar-chitecture with one hardware product (the VAX11/780), one operating system (VAX VMS), onenetwork (DECnet) and one high-level language(Fortran).

In early 1990s it introduced Alpha AXP ar-chitecture with several hardvvare products, threeoperating systems, multiple networking protocols,multiple languages, and so forth.

In this paper, we shall present a short overviewof the Alpha project. We will start with descrip-tion of main project goals and proceed to basicarchitectural features such as multiple instruc-

tion issuing and the possibility of multiprocessing.The first implementation of the Alpha AXP archi-tecture is DECchip 21064. The chip and severalhardware products using it are briefly described.Finally, we give a short overview of the operatingsystems supported by Alpha AXP architecture.

2 Project overview

Alpha was the largest engineering project inDigitaTs history, spanning more than 30 engi-neering groups in ten countries [8]. It startedwith a task force chartered to define a high-performance RISC architecture for the 1990s andbeyond. Even before the architecture definitionwas complete, work began on implementing ahigh-performance microprocessor. The work wasdone in the summer 1991 when a product-levelchip DECchip 21064 was fabricated [4]. However,a prototype chip was fabricated in late 1990 andwas used in an experimental multiprocessor sys-tem called ADU (Alpha demonstration unit) [9].This system was of great benefit to software de-velopers since it allovved them to boot the firstAlpha AXP operating systems early in 1991.

36 Informatica 17 (1993) 35-40 J. Šilc et al.

3 Project goals

The Alpha AXP architecture1 project startedwith a small list of goals:

• High performance and longevity. In cur-rent architectures, a primary limitation is the32-bit memory address. Therefore, the projectadopted afull 64-bit architecture (with a mitiimalnumber of 32-bit operations for backward com-patibility). It was estimated that it would be rea-sonable for raw clock rates to improve only by afactor of 10 over the coming 25 years. If the clockcannot be made faster, more work should be doneper clock tick to obtain increase in performance.Alpha AXP architecture was therefore designed toencourage multiple instruction issue implementa-tions eventually sustaining about 10 new instruc-tions starting every clock cycle. Additional per-formance improvements are to be expected frommultiple processors. Hence, Alpha AXP architec-ture project early focused on multiple processors,and designed a multiprocessor memory model andmatching instructions from the very beginning.

• Run several operating systems. Under-pinnings were placed for interrupt delivery andreturn, exceptions, context switching, memorymanagement, and error handling, all in a set ofprivileged software subroutines called PALcodes.By having different sets of PALcode for differentoperating systems, neither the hardware nor theoperating system is burdened with a bad interfacematch, and the architecture is not biased towarda particular computing style.

• Easy migration from other architecturecustomer bases. To run an existing (old archi-tecture, such as VAX and MIPS) binary versionof a complex application, the idea of binary trans-

1 Computer architecture is defined as the attributes andbehavior of a computer as seen by a machine languageprogrammer, while implementation is defined as the ac-tual hardvvare structure, logic design, and data-path or-ganization of a paiticulai embodiment of the architectuie.The architecture therefore carefully describes the behaviorthat a machine language programmer sees, but does notdescribe the means by which a particular implementationachieves that behavior.

lation was adopted [7]. It allows a user to get ap-plications up and running immediately, with min-imal porting effort.

4 Alpha AXP architecture

4.1 Its approach to RISC architecture

Alpha is a 64-bit load/store RISC architecture de-signed with particular emphasis on clock speed,multiple instruction issue, and multiple processors[1]. Its architects examined and analyzed currentand theoretical RISC architecture design elementsand developed high-performance alternatives forthe Alpha architecture.

4.2 True 64-bit architecture

All registers are 64 bit in length and all operationsare performed between 64-bit registers. Hence,it is not a 32-bit architecture that was later ex-panded to 64 bits. There are 32 integer registersR0..R31 and 32 floating-point registers F0..F31.

The basic unit of data is 64-bit quadword.There are three fundamental datatypes: inte-ger (32-bit longvvord, 64-bit quadword), IEEEfloating-point (32-bit S-floating point, 64-bit T-floating point), and VAX floating-point (32-bit F-floating point, 64-bit G-floating point).

Table 1: MIN and MAX Values for the Floating-point Data Formats

Data Format MIN MAXF-floating 0.294e-38 1.70e38G-floating 0.56e-308 0.899e308S-floating 1.175e-38 3.40e38T-floating 2.225e-308 1.798e308

Each of 168 Alpha instructions is 32 bits inlength. There are four major instruction for-mats (PALcode, Branch, Memory, Operate) andall have 6-bit opcode.• PALcode instructions. These instructionsspecify one of few dozen complex operations fromPrivileged Architecture Library2 to be performed.

2A Privileged Architecture Librarjr is a set of subrou-tines that is specific to a paiticular Alpha operating-systemimplementation.

ALPHA AXP OVERVIEW Informatica 17 (1993) 35-40 37

• Branch instructions. Conditional branch in-stmctions can test a register for positive/negativeor for zero/nonzero. They can also test integerregisters for even/odd. Unconditional branch in-structions can write a return address into a reg-ister. There is also a calculated jump instructionthat branches to an arbitrary 64-bit address in aregister.• Memory instructions. Memory instructionsare used for loads, stores, and a few miscellaneousoperations.• Operate instructions. There are fivegroups of register-to-register operate instructions:integer arithmetic, logical, byte-manipulation,floating-point, and miscellaneous.

4.3 Multiple instruction issue

Alpha implementation will issue multiple instruc-tions in a single cycle. To improve the odds ofmultiple-issue, compilers should choose pairs ofinstructions to put in aligned quadwords. Pickone from type A and one from type B but only atotal of one Load/Store and Branch per pair:

TypeA TypeBInteger Operate Floating OperateFloating Load/Store Integer Load/StoreFloating Branch Integer BranchInteger Operate Floating Operate

Unconditional BranchBranch to SubroutineJump to Subroutine

To avoid any mechanism that would hindersuch implementations, all special or hidden pro-ccssor resources were avoided [6]. Therefore, thereare:

• No branch delay slots. Branch delay slotsrequire exactly one following instruction to be ex-ecuted after a conditional branch. This, hovvever,does not scale well to a multiple-way issue chipwith a multiple-cycle instruction cache where sev-eral instructions will be needed in the delay slot.

• No suppressed instructions or skips.When execution of one instruction conditionallysuppresses or skips a following one (found in someother RISC architectures) the suppression bitsrepresent a nonreplicated hidden state. Hence,

it is difficult to multi-use more than one potentialsuppressor.

• No precise arithmetic exceptions. Re-porting an arithmetic exception (such as over-flow and underflovv) means that instructions sub-sequent to the one causing the exception mustnot be executed. This, however, becomes diffi-cult in a pipelined multiple issue implementation.Alpha architecture uses the Trap Barrier instruc-tion which stalls instruction issuing until all priorinstructions are guaranteed to complete vvithoutincurring arithmetic traps. A code-generationdesign was documented by Alpha project whichneeds one trap barrier per branch to give precisereporting.

• No single-byte writes to memory. Thebyte load/store instructions found in some otherRISC architectures can be a performance bottle-neck because they require an extra byte shifterin the speed-critical load and store paths, andthey force a hard choice in fast cache design.Therefore, in the Alpha AXP architecture, a byteload is done as an explicit load/shift sequence;a byte store as an explicit load/modify/store se-quence. Instructions in these sequences can bemulti-issued with other computation.

Moreover, there are no condition codes, noglobal exception enables, and no multiplier-quotient or string registers.

4.4 Shared-memory multiprocessing

An Alpha system consists of a collection of pro-cessors and shared coherent memories that areaccessible by all processors. (There may also beunshared memories.) There are several types ofaccesses3 that a processor may generate to sharedmemory locations (I-stream access, D-stream ac-cesses, and barriers).

Writes to shared data must be synchronized bythe programmer.

The basic multiprocessor interlocking primitive

3Instruction fetch by processor i to location x, returningvalue a.Data read by processor i to location x, returning value a.Data write by processor i to location x, returning value o.Memory barrier instruction issued by processor i.I-stream memory barrier instruction issued by processor i.

38 Informatica 17 (1993) 35-40 J. Šilc et al.

Table 2: Competitive Position

VendorDeviceMax freq. (internal)No. chips requiredPeak MIPSPeak MFLOPSBase arch. design

Digital21064

200MHz1

400200

64-bit

MIPSR4000

lOOMHz1

10050

64-bit

Sun/TIViking50MHz

115050

32-bit

IBMRIOS

50MHz7-9200100

32-bit

HPPA-4

66MHz2

132132

32-bit

Inteli860XP50MHz

1150100

32-bit'

Motorola88110

50MHz1

150100

32-bit

Table 3: CMOS Technology Roadmap

CMOSMfg YearMin features sizeChip power supplyMax /xP chip sizeGate oxideEfFective L# of wiring levels# of transistors

119852.0jxm5.0V0.9cm2

30nml.Zfim2200,000

219871.5/zm5.0V1.2cm2

22nm1.0/zm2400,000

319891.0/zm3.3V1.5cm2

15nm0.7îm3800,000

419910.75/zm3.3V2.2cm2

lOnm0.5/im31,680,000

is a RISC-style load-locked, in-register modify,store-conditional sequence of instructions.

If the sequence runs without interrupt, excep-tion, an interfacing write from another processor,or a PALcode instruction, then the conditionalstore succeeds - an atomic update was in factperformed. Otherwise, the store fails and the pro-gram eventually must branch back and retry thesequence until it succeeds.

This style of interlocking scales well with veryfast caches, and makes Alpha a suitable architec-ture for building multiprocessor systems. Thereis no strict multiprocessor read/write ordering,whereby the sequence of reads and writes issuedby one processor is delivered to all other proces-sors in exactly the order issued. The strict or-dering can be specified when needed by insertionof Memory Barrier instruction. This instructionguarantees that all subsequent loads or stores wUlnot access memory until after all previous loadsand stores have accessed memory, as observed byother processors.

5 An implementation:DECchip 21064

DECchip 21064 microprocessor represents thefirst implementation of the Alpha AXP architec-ture [4]. It is a super-scalar super-pipelined pro-cessor, using dual instruction issue, that has sam-pled up to 200MHz cycle time. Super-pipelinedmeans that an instruction is issued to the func-tional unit at every clock tick and the results arepipelined. The integer pipeline is seven stagesdeep, where each stage is a 5 ns clock cycle. Thefirst four stages are associated with instructionfetching, decoding, and scoreboard checking ofoperands. Pipeline stages 0 through 3 can bestalled. Beyond 3, however, all pipeline stagesadvance every cycle. Most ALU operations com-plete in cycle 4 allovving single-cycle latency, withthe shifter being the exception. Primary cacheaccesses complete in cycle 6, so cache latency isthree cycles. The floating-point pipeline is identi-cal and mostly shared with the integer pipeline instages 0 through 3, however, the execution phaseis three cycles longer.

Table 2 shows competitive position of DECchip21064 microprocessor.

ALPHA AXP OVERVIEW Informatica 17 (1993) 35-40 39

Table 4: Alpha AXP System Comparison Chart

SystemNo. ofprocessorsCPUClock speedMax memory capacityMax disk capacityMax I/O throughput

3000/4001

DECchip 21064133 MHz512 MB9.5 GB

90 MB/s

3000/5001

DECchip 21064150 MHz

1GB11.6 GB

100 MB/s

40001 or2


2GB56 GB

160 MB/s

7000up to 6


14 GBover 10 TB400 MB/s

10000up to 6


14 GBover 10 TB400 MB/s

The CMOS process is a .75 micron and 1.4- by1.7-cm chip incorporating 1.68 million transistors.DECchip 21064 include 8KB instruction cache,8KB data cache and two associated translationbuffers, a four-entry 32B/entry write buffer, apipelined 64B integer execution unit with 32-entry register file, and a pipelined floating-pointunit with an additional 32 registers. The bus in-terface unit handles all communication betweenthe chip and environment. The CMOS technologyused to manufacture the DECchip 21064 evolvedfrom three previous generation used to producevery high-performance microprocessors (Table 3).

6 Hardware products

At the time of writing, there are several hardwareproducts available spanning desktop through datacenter: DEC 3000/400 AXP (Desktop Worksta-tion or System), DEC 3000/500 AXP (DesksideWorkstation or System), DEC 4000 AXP (Dis-tributed/Departmental System), DEC 7000 AXP(Data Center System), and DEC 10000 AXP(Mainframe-Class System). Three more productsare due in the next few months [3]. See Table 4.

7 Operating systems

Alpha AXP supports today's leading operatingsystem environments: OpenVMS, UNIX, andWindows NT (to be announced soon).

• OpenVMS AXP.VMS, now known as OpenVMS, supports open

industry standard interfaces4, system purchasing

flexibility, and licensing. This combination is sonew to the industry that the VMS operating sys-tem has been renamed to OpenVMS. It runs onboth VAX and Alpha AXP hardvvare platforms.OpenVMS Alpha AXP provides the same featuresas OpenVMS VAX, enabling users and applica-tions to easily move from one system to another.The benefits of VAXclusters will be made avail-able on OpenVMS Alpha AXP, allowing Open-VMS VAX and OpenVMS Alpha AXP systems toco-exist in the same cluster. The main purpose ofmoving the OpenVMS system to the Alpha AXParchitecture was to deliver the performance ad-vantages of RISC to OpenVMS applications [5].

• DEC OSF/1 AXP.DEC OSF/1 AXP is a true native UNIX. It

implements the common definition agreed uponby UNIX Systems Labs (System V) and the OpenSoftware Foundation (OSF/1):- OSF Application Environment Specification;- Systems V Interface Definition;- OSF/Motif User interface;- Distributed Computing Environment support;and- Distributed Management Environment supportcommited.

DEC OSF/1 AXP supports several key stan-dards in the area of the operating and windowsystems.s It adds a range of enhanced program-

4The OpenVMS operating environment complies withIEEE POSIX and OSF/Motif standards and is X/OpenXPG3 BASE Branded. Future plans also call for Open-

VMS compliance with OSF Distributed Computing Envi-ronment, and support for XPG4.

SIEEE POSIX 1003.1 (1990), 1003.2 (partial), 1003.4a(threads), and 1004.3 Dll (threads); FIPS 160 (ANSI C);X/Open Portability Guide 3; System V Interface Defini-tion 2; System V Release compatibility (all SVID3 Baseand Kernel Extensions with the exceptions of streams, sig-nals, and counters); 4.3 BSD; Applications Environmentspecification (AES); MITs X Window System, XII Release5; Motif version 1.1.3; and ISO 9660 (CDROM f.s.).

40 Informatica 17 (1993) 35-40 J. Šilcet al.

ming tools available with DEC OSF/1 DeveloperExtension package. These tools provide a com-plete softvvare development environment for pro-grammers and application developers. It also pro-vides ULTRDt compatibility, through standardsconformance, development tools networking, userinterfaces, data interoperability and compilationsystems, allovving ULTRIX customers to move toDEC OSF operating system on the Alpha AXParchitecture in one step.

To support realtime, DEC OSF/1 offers:

- A pre-emptive kernel, to ensure that externalrealtime events get immediate attention- Fixed priority scheduling, to ensure that real-time applications aren't delayed by backgroundactivity;- Clocks a,nd timers, to provide the increasedfunctionality and granularity needed for realtimeapplications;- Process memory locking, to prevent system pag-ing and svvapping that could cause the system torespond unpredictably;- Asynchronous I/O, that enables application be-tween realtime processes;- Semaphores, for fast, reliable communicationbetween realtime processes;- Shared memory, for fast data sharing betweenprocesses or applications.

• Windows NT AXP.Alpha AXP systems will ruh Microsoffs Win-

dows NT. Through Windows NT, users who useDOS and Window-based applications today cancontinue to use them tomorrow on Alpha AXP-based systems that run many times faster thantoday's fastest PCs.

For these operating systems, the Noveraber1992 edition of the Alpha AXP Software Appli-cation Listing [2] provides a compendium of in-formation on over 1500 software products.

References

[1] Digital Equipment Corporation (1992) AlphaArchitecture Handbook. EC-H1689-10.

[2] Digital Equipment Corporation (1992) AlphaAXP Soft.ware Application Listing. First Edi-tion, EC-.J2088-10, November 1992.

[3] Digital Equipment Corporation (1993) AlphaAXP Systems Summary. EC-F2183-46.

[4] Dobborpnlil D. W. et al. (1992) A 200 MHzfiiM Dual Issue CMOS Microprocessor. IEEETrans. Solid State Circuits, 27, 11, p. 1555-1567.

[5] Kronenberg N. et al. (1993) Porting Open-VMS from VAX to Alpha AXP. Comm. ACM,36, 2, p. 45-53.

[6] Sites R. L. (1993) Alpha AXP Architecture.Comm. ACM, 36, 2, p. 33-44.

[7] Sites R. L. et al. (1993) Binary Translation.Comm. ACM, 36, 2, p. 69-81.

[8] Supnik R. M. (1993) DigitaTs Alpha Project.Comm. ACM, 36, 2, p. 30-32.

[9] Thacker C. P., Conroy D. G. & StewartL. C. (1993) The Alpha Demonstration Unit:A High-Performance Multiprocessor. Comm.ACM, 36, 2, p. 55-67.

The follovving tiademarks are used in this paper: AXP, Al-pha AXP, DEC, DECchip, DECnet, Digital, OpenVMS,ULTRIX, VAX, VAXcluster, VMS are trademarks of Dig-ital Equipment Corporation, OSF/1 and OSF/Motif areregistered trademarks of the Open Software Foundation,Windows NT is a trademark of Microsoft Corporation, andUNIX is a registered trademark of UNIX System Labora-tories, Inc.

Informatica 17 (1993) 41-51 41

OPEN SECURE MODEL AND ITS FUNCTIONALITYIN NETWORKS WITH VALUE-ADDED SERVICES

Borka Jerman-BlažičJožef Stefan Institute, Jamova 39, Ljubljana, SloveniaPhone: +38 61 159 199, Fax: +38 61 158 058

Keywords: security mechanisms, kryptography, notworks, value-added services

Edited by: Rudi Murn

Received: December 12, 1992 Revised: February 9, 1993 Accepted: March 1, 1993

The contribution gives an overview of the security functions employed h\ the ValueAdded Networks technology. The overvievr represents brieHy the user requirements forsecure communication, the basic threats to which communication systems are exposedand the framework ofOpen Secure Model defining the security functions and services tobe used in an open network. The security mechanisms used for provision of the secuntyservices a,nd functions are briefly described. Several applkations used in Value AddedNetworks with iabuilt security functions are brietiy introduced at the end.

1 Introduction

Information gains value once it is exchanged orconsumed. To be exchanged or consumed it mustbe transferred or delivered. The need for effectiveand safety means of carrying this process is todaygrowing faster then the growth of information andelectronic data processing. Nowadays, informa-tion interchange and data communication is anintegral part of any modern information system.Information interchange is a process taking partin the services offered through networks knownas Value Added Networks or VANs. These net-works usually interconnect communicating userswith various information services.

Every VAN is difFerent in its structure and theservices it ofFers. Some of the services are verygeneralized and some are extremely specialised. Acomprehensive VAN is likely to have the followinggeneral components [1]:

basic network,

generic services,

transaction relay,

application enabling,

information databases,

network management and help desk.

Generic services are general purpose servicesneeded by a wide range of customers rather beingapplication or industry specific. The main exam-ples are electronic mail, bulk data transfer, Elec-tronic Data Interchange (EDI) and informationservices support for managers or professionals.

VANs do not have to own a proper wide areanetwork, but they generally do. At its simplestthey might just provide point-to-point packetswitching. At a more advanced level they mayalso offer protocol conversion to support a widerange of different equipment. Overall, their aimis to provide full connectivity between all theequipment and systems that their customers need.Connectivity and security are inherently contra-dictory requirements. However, with speciallyadded security services it is possible to built open,fully connected network with any required level ofsecurity.

This contribution gives an overview of the se-curity functions taken as a part of globally in-terconnected network. The overview deals withthe user requirements for secure communication,the basic threats to which communication sys-tems are exposed and the framevrork of the modelwhich defines the security functions and servicesin an open netvrork. The security mechanismsemployed for the provision of the security servicesare briefly described. Several applications used in

42 Informatica 17 (1993) 41-51 B. Jerman-Blažič

Value Added Networks with inbuilt security func-tions are briefly introduced at the end.

2 Open networks and threats

Today, separate networks are integrated into aglobal connected network known as Global Inter-net [2] consisting of a number of interconnectednetworks. Individual computers and work sta-tions are connected to fast Local Area Networks(LAN) spanning office building or parts of them.These LANs are usually connected into fast areabackbone networks interconnecting one building,building complex or campus area at a speed com-parable to that of LANs (100 mb/s). Metropoli-tan Area Networks are emerging. They will spanentire cities with speeds above 100 Mb/s. Un-like LANs, MANs will be an integral part ofthe modern public network infrastructure beingowned and operated by teleoperators and sharingthe addressing and network management schemesof other public network. Wide Area Netvvorksare used for long-haul services. Modern WANsare based on fibber optics transmission systemsin the Gb/s speed range. All these networks to-day are not anymore separate physical networksbut rather virtual networks, i.e collections of Net-work Service Access points (NSAPs) forming oneworld-wide logical network. One NSAP can si-multaneously belong to any number of such log-ical networks. The protocol suites used in theseinterconnected logical network are mainly the In-ternet protocol i.e TCP/IP suite defined on theRequest for Comment standards [3] and the OSIprotocols developed within OSI Reference Model[4]. Upper layer ISO protocols can be run onthe top of TCP/IP and vice versa, enabling theconnected networks with different technology toprovide global connectivity. Recently, the Con-nectionless Network Service and Protocol devel-oped within ISO was adopted as an RFC standardand that among the other developed interworkingtechniques can be considered as step forward to-wards better coexistence of these technologies andprovision of global connectivity.

Connectivity and security are inherently con-tradictory requirements. However, openness as isunderstood today does not mean lack of securitybut it means interconnectivity and the ability tointeroperate betvveen systeras in different organi-

zations and from different manufacturers. Whenan open distributed system is built up it becomesessential to define the user requirements regard-ing the security of communication. The users,depending on which service of the communicatingsystern are they using may require difFerent levelof security. Usually, users are concerned with thefollovving:

- the identity of the other communicatingparty,

- that nobody else can listen to the session,

- that nobody can undetect delete from,change or add to the information they areinterchanging with other party,

- that commitments made during the sessioncan beyond reasonable doubt, afterward beprovided to an impartial judge.

The user apprehensions come out from the factthat the communication systems and resourcesconnected are usually targets of different threats.

The threats can be oriented towards the com-munication network itself or towards unautho-rized access to local system where the commu-nication network is used only as a medium of ac-cess. So, three categories of assets within a glob-ally interconnected network can be identified, themanipulation of which is a serious threat:

- the resources in the network,

- the informations conveyed

- and partner relations.

Local systems are resources accessed throughthe communication system and they must be pro-tected. The communication system itself is a re-source and must be protected too. The users ofcommunication systems expect the communica-tion system components to be present and to func-tion and in that sense, the availability of servicesand stability of services are also the assets of thecommunication system and need protection.

Informations are the actual content of commu-nication. Unauthorized access to informations,both by eavesdropping and by damaging, can de-stroy the value of information. Informations heldlocally and accessible through communication me-dia also belong to that category.

OPEN SECURE MODEL AND ITS FUNCTIONALITY... Informatica 17 (1993) 41-51 43

The relation between communicating partnersis another basic asset of communication. With-out trust in the authenticity of a communicatingpartner all communication with is worth noth-ing. Trusted partner relations are characterisedby: the trust in the identity of the partner andthe trust in the actions of communications.

All the assets of the communication systemare exposed to two fundamentally different kindsof threats i.e intentional or not intentional. Inthe classical security technology only one type ofthreats is considered i.e the intentional threatsrepresented by the act of espionage and sabo-tage. Espionage comprises all passive intentionalthreats such as to get unauthorized knowledge ofconfidential or classified information. Sabotagecomprises all active intentional threats i.e all kindof unauthorized manipulation of data, access tothe resources to the communication system etc.

The other potential accidents which are also re-garded as security relevant in the communicationnetworks are accidental threats such as bad main-tenance leading to an interruption of the networkservices. From the point of view of the users itdoes not make any difference if this is caused bya malicious or by an unable administrator.

The various threats and attacks in an openenvironment are classified within the frameworkdocument of ISO (International Standard Orga-nization) [5]. This document (ISO -7498 part 2)identify five difFerent attacks to the open commu-nicating system i.e :

— masquerade,

— repudiation of action or service or,

— denial of service.

— data interception

— data manipulation

a. Masquerade can happen during the mu-tual validation of the message transfer agent(MTA is an entity which transfer/exchangemessages in the electronic mail service) isby the exchange of the MTA names in plaintext. An unknown MTA (for example in test-ing procedure) may be interconnected withsome operational MTA by sending one of theknown MTA names. This is a typical mas-querade of identity with the intention to steal

vrorking resources or information. The mas-querade of user identity is possible also bytricky handling of routing oriented addresses[6].

b. Repudiation of action or service: repudiationof origin, submission, or delivery of informa-tion is extremely painful if contracts or otherbusiness documents are considered. How totrust to an invoice received by an EDI serviceif no evidence of the sender identity can beprovided?

c. Denial of services: denial of services can hap-pen due to accidental interruption caused bylocal system failures or by nonconformantcomponents in cooperating systems, such aserroneous entries in address routing or namemapping tables. Intentional interruptions arenormal for maintenance purposes.

d. Datainterception: the breach of confidential-ity is the most common attack in the exist-ing networks. It is impossible to guess thenumber of intentional espionage by systemadministrator or other unauthorised personsable to read data on their own or on othersystems. Data may be intercepted also nonintentionally in case of misrouted messagesetc.

e. Data manipulation: is any kind of unautho-rized modification of data and thus violatestheir integrity. The managing of electronicmail addresses is also in some sense a vi-olation of integrity, accidentally caused bybad maintenance. This is obviously a case ingatewaying, electronic message get loss or cutof their bodies. This type of vulnerability ofthe communication system includes also ma-nipulation of a message contents in the orig-inator's local store after non-repudiation ofsubmission and/or manipulation of messagecontents in the recipienfs store after non- re-pudiation of delivery of the message.

The situation that communication services notbeing provable and that different security fail-ures can happen in a globally interconnected net-work was acknowledged on many forums. Some ofthem spent a lot efforts to develop security func-tions and to provide security models. ISO has


addressed the security issue in several documentsdefining the Security Services or more properlythe Security Functions in an Open Environment.General overview of the Open Secure Architectureis given in the Security Frameworks document [7].

3 Security functions andservices

The Security Framework is intended to addressthe application of security services in an OpenSystem Environment, where the term "Open Sys-tems" is taken to include areas such as DataBases, Distributed Applications, OfRce Docu-ment Processing and Communication Netvvorks.This framework defines the means of providingprotection for systems and objects within thesystems and with interactions between systems.The framework address both information and se-quence of operations which are used to obtain spe-cific security services. These security services mayapply to the communication systems as well as tothe information exchanged between systems andto the local resources or data managed systems.The term security in the ISO framework is de-fined as "a mean of minimizing the vulnerabilitiesof assets and resources". Security is therefore un-derstood as a system preventing the attacks andprotecting the assets from the threats. Threatsare therefore, encountered by security servicesimplemented at different layer of communicatingnetworks or within the user interfaces. Securityservices are implemented by employing securitymechanisms. Some mechanisms prevent attacks,other detect attacks, some of the latter providerecovery of an unmanipulated state. They are:

Authentication: Many open systems applica-tions have security requirements which de-pend upon correctly identifying the princi-ples involved. Such requirements may in-clude the protection of assets and resourcesagainst unauthorized access, for which anidentity based access control mechanismmight be used, and/or for accounting andcharging purposes. The process of corrob-orating an identity is called authentication.

Access control: Many open systems applica-tions have security requirements which de-mand that resources be only used in a man-

ner consistent with the prevailing securitypolicy. The process of determining whetherthe use of resources within an open systemenvironment is permitted and subsequentlypreventing such use is called access control.

Non-repudiation: the non-repudiation servicesensures the proper collection and mainte-nance of information consisting of the originor delivery of data in order to protect an orig-inator against the false denial of a recipientthat the data has been received or to pro-tect a recipient against the false denial by anoriginator that the data has been sent.

Data Integrity: the maintenance of data valueis actually its integrity. Many open sys-tem applications have security requirementswhich depend upon the integrity of informa-tion. Such requirements may include the pro-tection of information used in the provisionof other security services such as authentica-tion, access control, confidentiality, audit andnon-repudiation, that, if an attacker couldmodify them could reduce or nullify the ef-fectivness of those services.

Data Confldentiality: Many applications haverequirements vvhich depend upon the se-crecy of information. Such requirements mayinclude the protection of information usedin the provision of other security servicessuch as authentication, access controls or in-tegrity, that if known by an attacker, couldreduce or nullify the efFectiveness of those ser-vices. The maintenance of the secrecy of datais called confidentiality.

Audit: A security audit is an independent reviewand examination of system records and activ-ities. The purpose of a security audit is anindependent review and examination of sys-tem records and activities. The security au-dits: tests the adequacy of system controls,confirm compliance with established securitypolicy, recommend any indicated changes incontrols, policy and procedures, assists in theanalysis of the attacks, and hence recom-mend damage control procedures. A securityaudits requires the collection and recordingof security related events in a security au-dit trail. A security audit itself involves the


analysis and reporting of the information col-lected by the security audit trail.

Key management: In communication and in-formation systems there is an ever increas-ing need for data to be protected againstunauthorized disclosure or manipulation us-ing cryptographic mechanisms. The securityand reliability of such mechanisms is directlydepended on the protection afforded to a se-curity parameter, called the key. The pur-pose of the key management is to provideprocedures for handling cryptographic keyingmaterial to be used in symmetric or asym-metric cryptographic mechanisms. Key man-agement includes key generation, key distri-bution, key installation, key storage and keydeletion. A fundamental problem in a keymanagement is to establish keying materialwhose origin, integrity, and in the case of se-cret keys, coniidentiality can be guaranteed.

The placement of particular security functionin the Open Architecture is not exactly defined.Security services or functions may be providedby different layers and by difFerent protocols de-pending of the user and application requirements.Some applications are more and some are less vul-nerable. The protection of particular applicationdepends also of the adopted security policy andon the technology used. It can be said, that nouniversal model exists and that the placement ofparticular function is chosen after the features andthe requirements of particular application regard-ing security are identified and by pragmatic con-siderations.

For example, in connection oriented end to endservices the transport connection is dedicated toserve one end to end instance of communicationand for that reason any security function can beplaced at the transport layer or between the trans-port and the network layer. There are severalprotocolsv that implement several security func-tions on that level i.e NLSP, SDNS, EESP andSP4 [7]. The placement of the function dependsaJso on particular user requirement such as forexample traffic flow confidentiality. This securityfunction can only be reliable implemented at lay-ers 1 through 3.

In the case of connection-less protocols such asIP the label techniques known as IPSO (IP Se-curity Option) is used. Such labels (sensitive,

unclassified, top-secret etc) are usually accompa-nied with encrypted data. If the data are sentto a trusty communication system (the deliveryof data is guaranteed to be to a authorized localsystem) then the label could be satisfactory pro-tection but in the case of untested netvrork i.e apublic data network then the packets of data areencrypted.

The placement of Authentication, Integrity andConfidentiality functions in the higher layers ordirectly in the Application Processes such is elec-tronic mail is straight forward solution which ispragmatic but not optical. Placement of the se-curity functions and mechanisms for each applica-tion (i.e for Virtual terminal, for File Transfer, forDirectory services etc) separately requires exces-sive development and duplication of functional-ity. This approach also contradicts the principlethat security should be an integral part of thewhole communication system and services pro-vided. However, the practice has shown that thisapproach is much more used today due to thecomplexity of the Interconnected networks anddifferent requirements for security in different ap-plications.

The security functions and services in the net-works are provided by employment of securitymecharilsms. The security mechanisms are alsodeiined in the Open Framework [5]. They arebriefly described in the chapter that follovvs.

4 Security mechanisms

Mechanisms and algorithms providing differentsecurity functions and services are all called secu-rity mechanisms. In fact these mechanism form ahierarchy:

Higher level mechanisms, such as security pro-tocols and semantic message contents,

Lower level mechanism, such as cryptosystems,forming parts of the above mentioned higher levelmechanisms,

Physical mechanisms, such as encryption chipsand pieces of program code, implementing theabove mentioned mechanisms.

The application of these mechanisms dependmainly on the security functions required and onthe complexity of the system to be protected.The security of a local system can, to a greatextent, can be ensured by physical security ar-


rangement. However, in a global open network itis impossible to guarantee security of communi-cations by means of physical safeness and therecryptographic techniques are applied.

4.1 Cryptography

Cryptography is a long-established way to keepinformation secret but nowadays the crypto-graphic mechanisms are specially developed andused to protect transfer of data and information.There are many cryptographic mechanisms butas basic ones used in globally connected networksthe following are considered: encryption and tech-niques for providing integrity and authenticationof the messages. For more details see [8,9]

An encryption mechanism is used to convert acleartext message into a cryptogram. An encryp-tion mechanism is based on a public algorithmand at least one key whose value is randomly cho-sen from larger set. A symmetric cryptosystemhas two functions i.e encrypt and decrypt. A mes-sage encrypted with key K can be only decryptedwith the same key [10].

In asymmetric encryption mechanisms the keyis divided into two parts, the encryption key andthe decryption key, in such a way that the en-cryption key specifies the encryption transforma-tion and the decryption key determines its leftinverse mapping decryption. The receiver of dataholds a secret key with which he can decifer buta different key is used by the sender to encipherand this can be made public vvithout in any waycompromising the system. This system providessecure communication in only one direction. Itis known as asymmetric or public key encryptionmechanism.

If it is unfeasible to derive the encryption keyfrom the decryption key then the system is calledpublic key signature mechanism. In that case ad-ditional information is required to check the dig-ital signature. An asymmetric encryption mech-anism provides complete confidentiality (only thelegitimate recipient in possession of the secret keycan decrypt the message) but no authenticationof the sender (anybody with access to the recipi-ent's public key could generate the message) butif the technique of digital signature is applied thenauthentication can be provided too. Unlike a nor-mal signature on a document, the value of themessage i.e the whole plaintext of the message is

transformed. In order to check the signature, thereceiver applies the encipherment function usingthe public key of the sender. If the result is aplaintext message, the signature is considered tobe valid. The argument for its validity is thatonly by possessing the secret key could anyoneproduce the transformed message which encipherswith the public key to generate a valid plaintext.Tliere are other ways of forming digital signaturein wliich the signature is not transformation ofthe message itself but an additional and separatevalue tlial goes along with the plaintext. In thatcase also origin and integrity of the message maybe claimed.

Data Integrity mechanisms provides means forthe sequence of messages to stay intact. Thismeans that no message have, undetected beenomitted or duplicated and that the original order-ing of the messages is preserved. Data integrityprovides detection of the changes in the data be-ing transferred, optionally recovering from thechanges, when possible and reporting the caseswhere recovery is not possible. Usually the tech-nique of a Checksum [11] or, preferable Cyclic Re-dundancy Check [12] is used to detect changes inthe data stream. Both represent a special fieldin the message which content guarantee that thedata were not changed.

Authentication is today used by use of pass-words which are concerned as very vulnerable andthat sort of Authentication is known as weak Au-thentication. Strong Authentication can be basedon symmetric or public key cryptography. Theprocedure of Strong Authentication and key ex-change is described in the CCITT Recommenda-tion X.5O9 [13] or ISO 9594 [14]. With symmet-ric cryptosystems a mutually agreed pairwise keybelonging to the appropriate security context isused for Strong Authentication between any pairof parties A and B. Public key signature mech-anisms have several advantages over symmetriccryptosystems when used for authentication. Keymanagement is greatly simplified by the fact thatonly public keys of the pairs need to be sharedand only one key pair is needed for each party.

A rapidly growing area in that field is that ofzero-knowledge techniques [15]. In these tech-niques, the secret authentication information ofeach party plays very much the same role as thesecret key in the public key cryptographic system


but it can not be used for data encryption, onlyfor data authentication and possible digital sig-nature. Some existing zero-knowledge techniquesmake key management very simple by completelyabolishing the need for user dependent public key.The draw back of these techniques is that they re-quire the secret keys to be generated by a trustedthird party and they can not be used for confi-dentiality.

In the case of cryptography, the physical mech-anism at the bottom of the hierarchy that areneeded to actually perform the cryptographicfunctions employed can be pieces of software run-ning on a piece of hardware or hardware. Withthe transmission speeds offered by current datanetworks, the efficiency of the physical mecha-nisms used is becoming a major issue of sys-tem design. The choice betvveen various physicalmechanism is trade-ofF between economy, flexibil-ity and performance. For details see [16].

4.2 Known cryptosystems

The most commonly used today cryptosystemsare the Data Encryption Standard (DES) whichdevelopment was initiated by US National Bu-reau of Standards but later resulted in manycommercial applications [17]. The Rivest-Shamir-Adleman (RSA) algorithm is the most commonlyused and probably most usable Public Key Cryp-tosystem today [18]. The Diffie-Hellman schemefirst proposed as the first published "public keyalgorithm" is still concerned as one of the bestmethods for secretly sharing pairvvise symmet-ric keys [19]. The algorithm is based on pub-lic "half-keys" and secret values associated withthem. From their public half-keys the commu-nicating parties can determine a pairwise session,which remains secret from other parties. This keycan then be used for mutual authentication andor exchanging secret information.

5 Security policy

An integral part of the Open Security Frameworkis the Security Policy. A security policy is a set ofrules which constrain one or more sets of securityrelevant activities of one or more sets of elements.Secure policy need not apply to all activities andelements in a communication system. This means

that its specification must inchide a specificationof the activities and the elements to which thepolicy applies. The rules for each security serviceare derived from the security policy.

Security policies are conventionally divided intoIdentity-Based and Rule-Based policies; Identity-Based security policies are based on privileges orcapabilities given to users and/or Access ControlLists associated with data items and other re-sources. In a Rule-Based security policy, SecurityClasses are normally used for determining what isauthorized behaviour. In identity-based systems,the users traditionally identifies himself by pre-senting to the system something he knows (e.g apassword). This is often called "need to know"policy.

It is only after an explicit security policy hasbeen stated that security becomes an engineeringproblem and every organization seriously inter-ested in security should have one. The enforce-ment of the adopted security policy and monitoryof security related events lies in the domain ofengineering.

A body responsible for the implementation of asecurity policy is called Security Authority. Secu-rity Authority may be a composite entity but suchentity must be always identifiable. A security do-main is a set of elements under a given policyadministered by a single authority for some spe-cific security relevant activities. An activity in-volves one or more elements such as: connectionsbetween difFerent layers in the protocol suite, op-eration relating to a specific management func-tion, non-repudiation operations involving a no-tary etc. The enforcement of the adopted secu-rity policy usually goes through generation of se-curity control information. One of them is a Se-curity Label. A security label is a set of securityattributes that is bound to an element, communi-cation channel or data item. A security label alsoindicates, either explicitly or implicitly, the au-thority responsible for creating the binding andthe security policy applicable to the use of thelabel. A security label can be used to supporta combination of security services. Examples ofsecurity labels are: indication of sensitivity i.e un-classified, confidential etc, to indicate protection,disposal and other handling requirements.

Another very important Security Control Infor-mation (SCI) is the Certificate. A certificate con-


tains SCI relating to one or more security services.Certificates are issued by Certification Authority.It is used to convey SCI from an authority to en-tities which require this information to performa security function. In general a certificate maycontain SCI for all security functions.

The security mechanism described in the chap-ter above involve the exchange of SCI either be-tween two communicating parties or betwoen thesecurity authority and the interacting parties.There are two common forms of protected secu-rity information that are used by the d< scribedmechanisms. One is called security lokeu, usedto protect security information that is passed be-tween interacting parties. The other is called asecurity certificate, used to protect security infor-mation obtained from an authority for use by oneor more of the interacting parties.

The Security Framework [5] does not define themethods and the procedures for implementing theSecurity Policy and related SCI. This is left to bedeveloped by particular organization and system.The security models and techniques developed forVANs and globally interconnected netvvorks is rel-atively new and the number of publicly known im-plementations is relatively small. The followingchapter gives a brief overview of known applica-tions in the field.

6 Applications providingsecure functions in VANs

6.1 Kerberos

The most prominent strong authentication servicein wide use today is the Kerberos AuthenticationServer created in the Athena project at MIT [20].Kerberos is in everyday use in several major U.Suniversities and obviously has solved a number ofsecurity problems in them. In Kerberos, authen-tication is based on symmetric encryption whichprecludes the stronger service of non-repudiationand leads to the problems of key management.However, non-repudiation is not considered as aserious threat in university environment. Ker-beros works in limited environments and there-fore the number of shortcomings such as the pos-session of the all master keys by only one partyi.e Kerberos itself can become unfeasible to bemanaged one day when the number of users and

service grow.

6.2 Private Enhanced MailThe other forthcoming application within theInternet is PEM (Private Enhanccd Mail) [21].PEM provides security services for e-mail usersand is a result of the developm,ent efforts by BBNin Cambridge, U.S based on the RFC 1113-1115which have been developed by IETF (Internet En-gineering Task Force) Privacy and Security Re-search Group [22]. The services provided are thefollowing: confidentiality, data origin authentica-tion and connectionless integrity as defined byISO [5]. These services are bundled into twogroups:

1. default services meaning tliat all messagesprocessed via PEM incorporate the authen-ticity, integrity and non-repudiation supportfacilities and

2. optional services such as confidentiality.

For compatibility purposes PEM is designed tobe transparent for X.4OO message transfer agentsystems. In the recipienfs workstation PEM mes-sages may be retrieved also by Post Office Pro-tocol [22] or by IPMS protocol P7 as defined inX.4OO environment [23].

PEM message processing involves three steps:SMTP (Small Mail Transfer Protocol) canonical-ization needed for compatibility with the MTAs,computation of the message integrity code (MIC)and computation of the optional message encryp-tion code. The second step begins with the calcu-lation of MIC (similarly to DES message authen-tication code) then encryption follows if requiredby the originator. Message key, used exclusivelyto encrypt the particular message, is generatedspecially for that message. The encryption algo-rithm employed is specified in the Key-info fieldin the PEM header along with any parametersrequired by the algorithm. The message text isthen encrypted using this per-message key. Thethird and final processing step renders encryptedor MIC-only message into a printable form suit-able for transmission via SMTP or other messag-ing systems.

To provide data origin authentication and mes-sage integrity, and to support non- repudiationwith proof of origin, the MIC computed in step 2

OPEN SECURE MODEL.AND ITS FUNCTIONALITY... Informatica 17 (1993) 41-51 49

is padded and then encrypted using private com-ponent of the originator's public public key pair.This eflFects a digital signature on the message,which can be verified by any PEM user. If themessage is encrypted, this signature value is en-crypted using the secret, per message key, whichwas ernployed to encrypt the message itself. Theresulted value is 6-bit encoded and included in theMlC-Info field along with the identifiers of theMIC algorithm and digital signature algorithm.The MD2 hash function is employed as the MICalgorithm and the RSA algorithm is employed asthe digital signature algorithm [22]. A hash func-tion is a well defined function of a message whichappears to generate a random number.

The PEM specification recommend use of pub-lic key cryptography for message integrity and au-thentication and for distribution of message en-cryption keys. PEM uses the public key certifi-cates as defined in the CCITT X.5O9 Recommen-dation [13]. On the basis of X.5O9 definition ofcertificate handling an Internet Certification Hi-erarchical (ICA) scheme is envisaged in which dif-ferent Certification Policy's are employed. ICA isexpected to be developed in near future. For de-tails see [21].

6.3 SecuDe System

SecuDE is software package which consists of Se-curity Application Programmers Interfaces pro-viding support for the application of the Authen-tication Scheme and Certificate Handling, ThePrivacy Enhanced Mail Support, X.4OO Messagehandling and Key Management. The system pro-vides various cryptographic mechanisms such asDES, RSA, hash functions, key generation andgeneration and verification of digital signature[24]. The signature algorithm employed is a com-position of a hash function followed by an RSAfunction. The signer's public key which is usedfor the verification of the signature is certifiedby Certification Authority. For encryption, theDES algorithm is used and for the transfer itselfthe secret DES key is RSA encrypted. For everyuser, the pair of RSA keys used for encryptionand decryption is different from the pair of RSAkeys used for signature and verification. A specialmodule is provided for support of the functionsfor the generation and distribution of keys, cer-tificates and certification paths enabling the func-

tionality of the Certification Authority as envis-aged in X.5O9. Additional module is also devel-oped for support of PEM and secure X.4OO mail[25].

6.4 EDI and Inter-Bank paymentsystem

Other applications of the Security FrameworkServices are in the EDI environment and inthe Inter-Bank payment systems [26]. Some ofthem (i.e the system ETEBAC 5 developed inFRANCE [27] use the authentication mechanismas defined in X.5O9 and the C (Message Authenti-cation Code) computed on plaintext data. MACis defined in the document ANSI Financial Insti-titution Message Authentication and is a sort ofauthenticator). The MAC key is exchanged (en-crypted under RSA) for each session. The confi-dentiality is configured by another key drawn bythe sender. The non repudiation is based on theRSA algorithm. The digital signature comprisesthe MAC calculated on the Message Identifier andMAC calculated on the whole message. The se-cret key of the sender issued for computation ofthe signature.

The Holland AMRO-ABN bank implementa-tion comprises two modules: a one way hash func-tion which compresses the bulk payment to a codeof a fixed length. This module is based on theDES algorithm. The output code of the mod-ule 1 (hash function) is encrypted with an RSAdigital signature to provide currently authenticityand non-repudiation . Three main functions areessential: user identification, generation of digi-tal signature and verification of digital signature.The necessary key management is based on thegeneration of the keys by every user and the cer-tification of the keys by a Certification Authorityafter checking both the integrity and authenticityof the keys. Key generation is planned to be basedon CCITT X.509.The response messages are usedto provide non-repudiation of receipt.

Teletrust (Trustworthy Telematic Transac-tions) implements the public key mechanism andreliable hashing functions. The basic Teletrustdevice is a token.lt is a credit card size chip thatcan not be tampered. The token is protected by aPIN (Personal Identification Number, a sequenceof digits used for identification of the holder of abank card ) and is used as payment device. The


token is activated by the user and the mutualauthentication with the service provider is per-formed by exchange of the public keys with RSA.Once completed, a payment transaction may takeplace and the token is used to "sign" the payment(digital signature).The authentication requires aCertification Authority (in this case the centralbank). The token stores therefore the public keysand the signatures of the central bank used forthe authentication of the users's bank, which isused for authentication of the user and the user,which is used for the authentication of the trans-action.The digital signature is used for authenti-cation, integrity and non-repudiation service.

7 Conclusion

Security is a central consideration in the evolu-tion of Value Added Networks. Security servicesand functions are needed to protect the infrastruc-ture of the communication system, the local re-sources as well as to provide enough assurance tothe prospective users by guaranteeing safe trans-port of sensitive and high value information. For-tunately, today the fast progress of technical de-velopments is rapidly improving the security ofthe networks providing in the same time open-ness, connectivity and safety.

References

[1] R.Reardon (ed.) Future Networks, BlenheimOnline, London 1989.

[2] Internet: Getting started, SRI International,Menlo Park, CA, 1992.

[3] J.B. Postel, (ed.) IAB official protocol stan-dards, 1992 J.B. Postel, Internet Protocol,RFC 791, 1981.

[4] ISO, Information Processing Systems, OpenSystem Interconnection Reference Model,Patt-.l Basic Reference Model, ISO 7498-1,Geneva 1984.

[5] ISO, Information Processing Systems, OpenSystem Interconnection Reference Model,Part:l Security Architecture, ISO 7498-2,Geneva 1988.

[6] R.Grimm, Security on Networks: Do WE Re-ally Need it?, Comp.Netvrorks and ISDN Sys-tems, Vol.17, No 4&5, October 1989, p.315-321.

[7] A.T.Karila, Open System Security- an Architectural Framework, Espoo 1991,Helsinki.

[8] D.W.Davies and W.L.Price, Security forComputer Networks, Sc.ed.,J.Willey andSons, Chichester, 1989.

[9] S.Muftic, (ed.) Security Mechanisms forComputer Networks, Ellis Horwood Ltd,Chichester, 1989.

[10] C.Shannon, Communication Theory of Se-crecy Systems, Bell System Technical Jour-nal, Vol.28, 1949, p.656-715.

[11] ISO, Information Technology, Security Tech-niques, A Data Integrity Mechanism, ISODP 9797, Geneva 1990

[12] S.Walker, Network security: The parts of theSum, Proceedings of the 1989 IEEE Com-puter Society Symposium on Security andPrivacy, Oakland 1989, p.2-9.

[13] The Directory - Overview ofConcepts, Mod-els and Services, CCITT RecommendationX.5OO, Melbourne 1988, and The Directory,Part 8: Authentication Framework, CCITTRecommendation X.5O9 (Melbourne 1989,).

[14] ISO 9594, Information Processing Systems,OSI - The Directory, 9594 through parts 1 -8, Geneva 1989.

[15] ISO 10181, Information Technology, OSI Se-curity Model, Part 1 Security Framework,Part 2, Authentication framework A.Shamir,Identity-Based Cryptosystem and SignatureScheme, Advances in Cryptology: Proceed-ings of Crypto 84, Springer, Berlin, 1985,pp.47-53.

[16] J.Smith, Cryptographic Support in a Giga-bit Network, INET 92, Proceedings, InternetSociety Reston VA, Kobe, 1992, p.229-239.

[17] R.M.Davis, The Data Encryption Standardin perspective, Computer Security and theData Encryption Standard, National Bureau

OPEN SECURE MODEL AND ITS FUNCTIONALITY. Informatica 17 (1993) 41-51 51

of Standards Special Publication, 500-527,Washington, 1978.

[18] R:Rivest, A.Shamir, L.Adleman, A Methodfor Obtaining Digital Signatures and PublicKey Cryptosystems, Communication of theACM, Vol.21, No.2, 1978, p.120-126.

[19] W.Diffie,M.Hellman, New Direčtions inCryptography, IEEE, Transaction on Infor-mation Theory, Vol IT-22, No.6,1976, p.644-654.

[20] D.Otway, O.Rees, Efficient and Timely Mu-tual Authentication, ACM Operating Sys-tem Review, Vol 21, No.l, 1987, p.8-10.

[21] S.Kent, An Overview oflnternet Privacy En-hanced Mail, INET 92, Proceedings, InternetSociety Reston VA, Kobe, 1992, p.217-227.

[22] J.Linn, Privacy Enhancement for InternetElectronic Mail, Part III, Algorithms, SRINIC International, RFC 1115, Auvgust 1989.

[23] M.Rose, Post Office Protocol: Version 3, SRINIC International, RFC 1225, May 1991.

[24] R.Grimm, R.Nauster, W.Scheider, SecuDE,Pnndples of Security Operations, TechnicalReport, Version 2.0, GMD, Darmstadt, Ger-many, 1991.

[25] Message Handling Systems, Recommenda-tion X.4OO, CCITT, Blue Book Vol VII, Fas-cicle VII.7, Melbourne 1988.

[26] B.Jerman-Blažič, Security in Value AddedNetworks - security requirements for EDI,Comp.Stands, North Holland, Vol.12, 1991,p.23-33.

[27] The TEDIS Programme 1988-1989, ActivityReport, and TEDIS-EDI Security Workshop,Security in a Multi-Owner System, Brus-sels, June 20-21, 1989, Digital Signature inEDIFACT a TEDIS programme report pre-pared by CRYPTOMATHIC A/S, final ver-sion, Nov.29, 1990.

Taformatica 17 (1993) 53-58 53

REAL AI

Matjaž GamsJožef Stefan Institute, Jamova 39, Ljubljana, SloveniaPhone: +38 61 159 199, Fax: +38 61 158 058E-mail: [email protected]

Keywords: artificial intelligence, viewpoint

Edited by: Branko SoučekReceived: December 12, 1992 Revised: February 9, 1993 Accepted: March 1, 1993

In the first part, four viewpoints on AI are presented. It is proposed that a programexhibiting AI is one that can change as a result of interactions with the environment.While no program can be proclaimed as intelligent at tbe moment, intelligence may bejust an emerging property of successful information machines or beings. Jn the secondpart, ideas, problems and misconceptions about AI are analysed through grouping intothree categories: (1) facts - opinions that are supported by facts; (2) legends - opinions,based on facts but largely exaggerated; (3) myths - opinions not based on facts.

1 Introduction

Discussions about AI have attracted most of theresearchers inside the field as well as many out-side (Fox 1990, Hayes-Roth & Fikes 1991, Hayeset. al. 1992, Mettrey 1992, Schank 1991, Searle1982, Wilkes 1992). In particular, there have beenimportant shifts and modifications in world-wideopinion observed in recent relevant publications.These debates have motivated us to make an at-tempt to summarise them in a coherent way. *

First, let us analyse the notion of artificial in-telligence.

2 Viewpoints on AI

Thinking about the definition of AI, one shouldask 'Where's the AI?' (Schank 1991). There seemto be at least four prevailing answers to that ques-tion.

The first view sees AI as something magic thatemerges out of a computationally effective com-puter after you put in it enough things. Indeed,it is still often acclaimed that neural networksmimic the behaviour of human brains. That israther surprising since what they do at present is

*A similar but shorter paper (Gams 1992) has been pre-sented as the first papei in the AI section at the ERK'92conference - similar parts are reprinted with permission.

to mimic a numeric filter at best able to tune pre-defined parameters. Not surprisingly, computa-tionally more diverse and structurarily more com-plex statistical methods typically achieve betterclassification accuracy (Henery & Taylor 1992).

This approach has already contributed to thefirst dark age at the beginning of AI and is stillpresent in many subfields of AI. On the otherhand, one should not underestimate their advan-tages such as flexibility and robustness.

The second view sees AI as a superb infer-ence engine and therefore resembles knowledgeengineering. What one has to do is to find anexpert, encode his knowledge into lists of rules,add an inference engine with appropriate inter-face and there you are. This has led many peopleto think that AI means rule-based expert systems,and then they thought they understand them aswell as AI. And since they have also learned thelimitations of rule-based systems, they also thinkthat is the limitation of AI, not just of one com-ponent of AI (Hayes-Roth & Fikes 1991). Whileconnectionism as well as knowledge engineeringand inference engines are important parts of AIresearch and applications, labelling it intelligentor as AI itself is misleading (Schank 1991).

The third view maintains that, if no machineever did it before, it must be AI. For example,years ago research in computer chess was one of

54 Informatica 17 (1993) 53-58 M. Gams

the central themata in AI. People thought thatcomputer programs playing chess well would cer-tainly have to be intelligent. Now, nearly every-body agrees that these programs do not have anydeep intelligence at all. Luckily, there are severalexplanations of this contradiction. For one thing,this viewpoint seems to confuse getting a machineto do something intelligent with getting it to be amodel of human intelligence. More important, ifyou define AI in that way - to be one of frontiersof computer science, once the area that you arelooking at is understood, then it is no longer atthe frontiers of computer science and therefore nolonger AI, and so it is a no-win kind of situation(Schank 1991).

So, where is the intelligence in computer pro-grams? As mentioned earlier, many difficvdt prob-lems which had long been thought to requirereal intelligence, have been solved by rather un-intelligent methods. Intensifying this argument,even superb intelligent behaviour does not guar-antee that real intelligence and understandinghave been achieved. For example, Searle (Searle1982) has constructed a hypothetic Chinese roomin which a group of workers performs intelli-gent translation between two natural languages(English-Chinese) and each of them performs onlya subpart of the whole process on the basis of apredefined procedure. Although such a Chineseroom could pass Turing's test, the room (and no-body inside it) does not understand the wholeprocess and there is no real intelligence at all.

After a decade of quite intensive debate, therehas not been any definitive answer to this para-dox since it is actually a philosophical question:Is performance, i.e. a mechanicistic approach re-ally sufficient? From a practical point of view,most things in our world work in the mechanicis-tic mode. Of course, there are paradoxes and un-solved questions (e.g. are there other universes oris there only ours?) but people have successfullylived with them. Not to mention that other ap-proaches based on ideology or spiritism have notyielded similarly good results.

Therefore, it seems rather surprising that crit-icism is so strong in the AI area even if thesame arguments are being repeated over and overagain. A recent example of this kind could bethe article "Artificial Intelligence as the Year 2000Approaches" (Wilkes 1992). It provoked harsh

replies (Hayes et. al. 1992) in which several errorsand misconceptions were exposed. Nevertheless,in all this argumentatipn there are at least twopoints where the AI community still has to proveitself:

- if intelligence (in computers) were simple,fast and powerful computprs would have fa-cilitated it a long time ago, and

- many of the ideas in the AI field have pro-duced much more optimism than real im-provements.

Here we shall devote attentioji to the first argu-ment, and the second argument will be analysedin the second part of this paper.

For example, Wilkes (Wilkes 1992) claims thatintelligence may come from analogue circuitrysince, obviously, it has not come from digital com-puters so far. Searle (Searle 1982) claims thatdigital machines can not be intelligent as biologi-cal beings since they are essentially difFerent. Al-though Hayes (Hayes et. al. 1992) claims that noproof is given for such claims, the same is validalso for the reverse claim.

At this point we can only agree that real in-telligence in machines has not been achieved yet.Furthermore, we still do not have any good ideashow to make a true intelligent machine. However,two arguments seem plausible:

— Real human intelligence is very complex. Ifit were simple enough for us to understandit, than we would be too simple to perceivethat (as claimed by several authors).

— Intelligence may be just an emerging prop-erty of successful information machines or be-ings. There does not have to be any deepermotive or principle behind it. This approachis very close to the "artificial life" where com-putational models share many characteristicswith biological computation (Brooks 1991).

Furthermore, in computing there are goodfoundations and clear concepts like Turing's ma-chine or Church's thesis. There is also Turing'stest in which real intelligence is achieved whenhuman judges can not distinguish between theperformance of a computer and human. Sincecomputer programs are far away from achievingsuch a level, the contest area is often limited to

REAL AI Informatica 17 (1993) 53-58 55

a domain which still requires intelligence by hu-man counterpart. However, it is important to no-tice that the communication between the judgeand the contenders is an open one. Therefore,a computer program playing superb chess butunable to explain the motives of its moves cer-tainly would not pass the test while even a noviceplayer with normal explanation and reasoning ca-pabilities certainly would. In recent years slightlymodified tests, or competitions, are becoming an-nual events with rewards up to $100,000 (Epstein1992).

This leads us to the fourth view on AI. Trueintelligence, exhibited by computer programs,would have to have many or even most proper-ties of human intelligent behaviour regardless ofhow narrow the application area was. One of themain such properties is learning since intelligencefirst of all means getting better over time. In rela-tion to Turing's test, a computer program unableto learn from its mistake would certainly be ex-posed. Today, hardly any AI programs learn fromtheir mistakes, although - with very good reason,learning is the central area of AI at least in thelast decade.

There is some additional reasoning about intel-ligence:

- Intelligence is in size. It is hard to expect asmall program to display intelligence. Intelli-gence is neither simple nor easy understand-able.

- Intelligence is in complexity and heterogene-ity. This area is sometimes related to multi-strategy learning (Michalski & Tecuci 1992),a multiple-knowledge approach (Gams et. al.1991) and multi-agents (Minsky 1987).

- Intelligence is in the ability to perform wellreal-life tasks which require the use of knowl-edge. For example, Mathematica, a programfor symbolic computing is regarded as ap-proximately as intelligent as a numeric li-brary. Contrary to recent interest in logicprogramming, it is quite probable that in-telligence there will be at a similar level toMathematica until real-life knowledge is in-corporated into programs.

Furthermore, there are several aspects of intelli-gence each of which can be compared if not mea-

sured on a scale. For example, motional intelli-gence can be quite high in many animals. In an-other aspect, AI research can well be at the fron-tiers of computer science while AI applications fellinto an application area years away from scientificachievements. AI applications do not have to beintelligent, they have to be related to AI researchsimilar to other science/application relations.

In short, while agitated debates about AI raiseinterest and in both ways afFect funds, what reallymatters is what works and which new discoveriesare produced. It is not that AI needs definitions;it is more that AI needs substance (Schank 1991).Although general artificial intelligence has not yetbeen achieved, we know more and more about it.Some basic facts, legends and myths about AI willbe represented in the following sections.

3 Facts

The first AI concept is search. Most difficultproblems involve choosing between alternative so-lutions and evaluation processes in which the bestsolution is found. This basic search schema maynot be immediately observed in diverse subareasof AI such as scheduling, games, learning or ex-pert systems. Novice readers in AI might get dis-tressed by difFerent terminology and diverse tech-niques. However, even one of the oldest defini-tions of AI promotes it as a fight against combi-natorial explosion.

While faster computers certainly help, simplesearch techniques can not ever deal with the ex-ponential growth rate of the number of possibil-ities in a search tree. For example, in a singlefactory having 85 orders, 10 operations, and onlyone substitutional machine, one could create over10880 alternative schedules (Fox 1990) while thenumber of all atomic particles in our universe isestimated at 1080. Obviously, the key question ishow to reduce search space.

The second AI concept is knowledge rep-resentation. It is not that the knowledge rep-resentation concept is second to search; it is oneof the two. Knowledge is probably even more im-portant than search in biological systems. In reallife, response to a specific pattern is usually pre-stored - learned through experience. This resem-bles the fourth viewpoint on AI presented earlier.But from a practical point of view, computers as


well as AI were more successful in search than inknowledge representation.

Although there are many techniques from se-mantic nets to frames, the most successful AI ap-plications so far are expert systems. At times,it was thought that the expert-systems approachenables a uniform solution to knowledge represen-tation problems. It has led to overenthusiasm andoverselling the technological possibilities. Now weknow that expert systems are appropriate onlywhen problems are relatively small and stable orcan be decomposed into such subproblems, mean-ing that experts agree with each other upon aproposed knowledge base.

The main problem, how to represent differentkinds of knowledge, complex and heterogeneousknowledge, and combine them into one system hasnot been solved yet. As a consequence, success-ful learning from interactions with the environ-ment has not been, and quite probably can notbe achieved vvithout it.

AI copes with the search combinatoricexplosion by using knowledge. The use ofknowledge enables successful pruning of a searchtree. For example, in an expert system OPEX(Gams et. al. 1991) for generating appropriatemachining operation sequences, designed in coop-eration with researchers from the Faculty of Me-chanical Engineering and Jožef Stefan Institute,there are three levels of rules:(a) Rules for applying basic machining operation.Example:operation drilling : if

gdb:fc is-a-cylinder-in andgdb:dc included interval(3,40) andgdb:lc/max(gdb:dc) = 10 andgdb:nc subset interval( 11,12)

thenfc := is-a blank and dc := 0 and nc := unde-

fined(b) Rules that define various possibilities of link-ing basic machining operations within an individ-ual feature. Example:from boring to drilling if true end.(c) Rules for combining operation sequencesthat define which operation sequences should beadopted for a combination of features. Example:combination drilling and drilling if true end.

The task of OPEX is to design operation se-quences for a machine and a specified part, and

sort them according to predefined criteria. Naivecombining of operations quickly leads to combina-torial explosion, but through smarter selection ofpossibiUties, i.e. by utilising domain knowledge,the combinatorics is reduced to a feasible level.

As indicated by previous example, AI enhancessearch by reformulating problems, through theuse of opportunism, heuristics,.and by abstractionand differentiation of quantitative models. Thesean1 techniques behind the general principle of us-ing knovdedge to control search. In essence theyperform similar improvements of search as hierar-chical searcli or dynamic programming, however,the use of knowledge can greatly improve perfor-mance.

AI systems can increase productivity.Various reports estimate the number of AI sys-tems regularly in use to around 3000 with someof them being in use for more than 10 years.The main problem with such estimates is whereto put borders betvveen AI and non-AI applica-tions. For example, is Prolog interpreter an AIapplication or not? Clearly, there is no intelli-gence in it. On the other hand, Prolog as wellas Lisp and many other products were designedas a by-product in AI research. In our opin-ion, they should be included as AI applicationsas well as neural networks. Actually, marketplaceAI-software packages fall into at least four ma-jor categories: programming languages, program-ming environments, problem-solving shells (for aclass of problems), and appllcation shells (spe-cialised for a given domain).

For example, in our rather small country ofSlovenia, in two AI laboratories at Jožef StefanInstitute and the Faculty for Electrical Engineer-ing and Computing, around 60 applications weresuccessfully performed in recent years and 5 origi-nal programming systems with several thousandsof lines are in regular use (Urbančič & Križman1991).

4 Legends

AI systems are easy to build. Indeed, underspecific conditions, improvements in speed andproductivity are enormous when using AI sys-tems. For example, having stored a history ofevents, it is possible to design an expert systemwith the use of inductive learning tools in a couple

REAL AI Informatica 17 (1993) 53-58 57

of days. On the other hand, there are problemswhich take more or even much more time than byclassical methods.

Speciflcations and prototyping largelyenhance productivity. This is partially true.If the problem fits an application shell, knowledgegathered from experts can be put into a systemquickly and then tested. Rapid prototyping elicitsthe requirements and specifications of software forill-defined problems; in recent years it has beenincluded in software development approaches asanother example of AI product finishing in clas-sical computer science and applications. But thelimitations of the metliodology and conditions forsuccessfulness have also become known.

AI systems are easily verified and main-tained. Since AI systems rely on knowledge in-stead of formulae, e.g. expert knowledge in expertsystems, it is often propagated that these systemsare highly understandable and, therefore, easy tobe verified and maintained. For example, expertsystems provide explanation possibilities as a sortof rule tracker instead of 'trace' in conventionalprogramming languages. Practical experience hasshown that while it is an important improvementover classical methods, verification and mainte-nance remain time consuming phases.

5 Myths

Artificial intelligence approach doesnot need conventional program-engineeringand management techniques. This incorrectbelief is still quite common due in part to aca-demic ignorance of the requirements for buildingproduction-level systems.

Systems working on simple examples caneasily be upgraded to full-scale real-life sys-tems. Performing speech understanding for asmall vocabulary of, say 50 words, difFers greatlyfrom the same task but with thousands of words.Similarly, many problems are difficult only be-cause of their size. The myth of simple scalingis still very alive mainly due to an academic ap-proach where it is most important that idea isfresh and attractive (vrorking on a simple, care-fully designed problem). Literature reviews in AIshow that about half of all publications belong tothis category and only half of the systems actuallywork on non-toy problems. In the worst scenario,

some subareas of AI have for years attracted inter-est and funds without actually producing a pro-gram working on a realistic problem. There seemto be certain similarities to fashion movements inwhich a new direction promoted by famous peo-ple attracts global interest. After a critical massis obtained, the movement can sustain for severalyears without any realistic verification. The prob-lem is similar in several other sciences. The "pub-lish or perish" science tends to produce famouswriters instead of famous scientists, researchers,engineers or inventors. However slowly, in AI itis changing in favour of more strict verification ofresults. For example, there are several projectswhich for years have evaluated different methods(Henery & Tayk>r 1992). Even at our laboratorieswe have been testing all available inductive learn-ing systems for 5 years and making the resultspublic.

Small systems can exhibit full-scale hu-man intelligence. In serious AI circles it isknown that it is not possible to simulate full-scalehuman intelligence without huge and complex sys-tems and that searching for a genuine simple al-gorithm is similar to searching for perpetuum mo-bile.

If we have an expert, then we can cre-ate an expert system. Obviously, a lot more isneeded; first of all a feasibility study.

AI does not need business motivation toproduce valuable results. Several studies haveshown that those initiated by management havea better chance of returning profit.

AI tools can enable novices to developexpert systems. Inexperience and lack of skillcan not be compensated in any field.

Expert systems consist of expert sys-tems. Typically, in expert systems there is muchmore than that, including lots of classical pro-gramming.

Expert systems perform as specialised,stand-alone programs. Actually, they accessdatabases, conventional programming languages,operating systems etc.

All AI tools are the same. There are difFer-ent categories.

All expert systems are rule-based. Many,but there is much more.

Expert systems do not make mistakes. Inreal life there is no such thing.


AI replaces conventional approaches.Rather, they can both be useful depending onconditions, and are often combined together.

AI knowledge engineering is all we needto know about AI. The more you know the bet-ter. Again, AI consists of many diverse subareas.

AI tools are good only for AI appli-cations. AI softvvare supports qualitative andquantitative reasoning equally well.

In simple expert systems an exhaustivesearch can provide solutions. Yes, for toyproblems.

Tools equally support both forward andbackward chaining. At the expense of theother.

The more general the tool the better.Task specific tools are actually more productivebut on a more narrow area.

There exist universal algorithms for spe-cific subareas such as learning. In theory, notworking in practise.

Several subareas of AI have good theo-retical foundations. No true intelligence has itso far.

6 Conclusion

AI systems can work well under favourable con-ditions, and are neither panaceas nor research cu-riosities. AI is not (just) art or a fashion, it is firstof all a scientific discipline. At present, AI can im-portantly improve productivity and enhance theapplication areas of computers. As all other tech-nologies, it must be used with a certain precautionand especially when circumstances are favourable.Therefore, more knowledge about AI in general aswell as knowing about common legends and mythsabout AI may improve the success rate and ex-tend the number of AI applications.

References

[1] Brooks R. A. (1991) Intelligence Without Rea-son. Proceedings of the IJCAI'91 Conference,Sydney, Australia, p. 569-595.

[2] Epstein R. (1992) Can Machines Think; TheQuest for the Thinking Computer. AI Maga-zine, 13, 2, p. 80-95.

[3] Fox M. S. (1990) AI and Expert SystemMyths, Legends, and Facts. IEEE Expert,February 1990, p. 8-19.

[4] Gams M. (1992) Facts, Legends and Mythsabout AI. Proceedings ofthe ERK'92, Portorož,Slovenia, p. 139-142.

[5] Gams M., Drobnič M. k Petkovšek M. (1991)Learning from examples - a uniform view. Int.Journal for Man-machine Studies, 34, p. 49-89.

[6] Hayes P. J., Novak G. S. & Lehnert W. G.(1992) ACM Forum - In Defence of ArtificialIntelligence. it Communications of the ACM,35, 12, p. 13-14.

[7] Hayes-Roth F. & Fikes R. (1991) Interview.IEEE Ezpert, 6, 5, p. 3-14.

[8] Henery R. & Taylor C. (1992) Applications ofMachine Learning, Statistics and Neural Net-works in Prediction and Classification. Proceed-ings ofthe ISSEK Workshop'92, Bled, Slovenia,p. 22.

[9] Mettrey W. (1992) Expert Systems and Tools:Myths and Realities. IEEE Expert, 7,1, p. 4-12.

[10] Michalski R. S. & Tecuci G. (1991) Proceed-ings of the First International JVorkshop onMultistrategy Learning. Harpers Ferry, USA.

[11] Minsky M. (1987) The Society of Mind. NewYork: Simon and Schuster.

[12] Schank R. C. (1991) Where's the AI?. AImagazine, 12,4, p. 38-49.

[13] Searle J. R. (1982) The Chinese Room Revis-ited. Behavioral and Brain Sciences, 8, p. 345-348.

[14] Sluga A., Butala P., Lavrač N. & Gams M.(1988) An attempt to implement expert systemtechniques in CAPP. Robotics & Computer-Integrated Manufacturing, 4, 1/2, p. 77-82.

[15] Urbančič T. & Križman V. (1991) A Hand-book of AI Applications. IJS Press, Slovenia.

[16] Wilkes M. W. (1992) Artificial Intelligence asthe Year 2000 Approaches. Communications ofthe ACM, 35, 8, p. 17-20.

Informatica 17 (1993) 59-63 59

REGULAR GRAPHS ARE 'DIFFICULT' FOR COLOURING

Janez ŽerovnikUniversity of Maribor, Faculty of Technical Sciences,Smetanova 17, Maribor,Slovenia

Keywords: graph colouring, decision problem, NP-completeness, rcgular graphs

Edited by: Rudi Murn

Received: December 19, 1992 Revised: February 16, 1993 Accepted: March 1, 1993

Abstract: Let k be 3 or 4. In this two cases we prove that the decision problem ofk-colourability when restricted to A-regular graphs is NP-complete for any A > k + 1.

1 Introduction

In this note we consider the time complexity ofthe decision problem of (vertex) A;-colourabilityrestricted to regular graphs.

It is known that 'almost all /u-colourable graphsare easy to colour', namely the proportion of 'dif-ficult' graphs for the usual backtrack algorithmvanishes with growing problem size [9]. Know-ing this it is not surprising that there are algo-rithms with average polynomial time complexity[1], when average is taken over all graphs and evenwhen the average is taken over all 3-colourablegraphs with a given number of vertices[5].

If P^NP, then for every algorithm there hasto be a class of 'counterexamples', i.e. graphs onwhich the algorithm either has superpolynomialtime complexity or it fails to produce a correctanswer.

For example, Petford and Welsh noticed thatone of the situations in which the 3-colourablegraphs were not efficiently coloured by their ran-domised algorithm is when graphs are approxi-mately regular of a low vertex degree [8]. Sim-ilarly, approximately regular graphs of a rel-ative low vertex degree are 'difficult' also forthe fc-colouring generalisation of their algorithm[10]. Petford and Welsh conjectured that 'dense'graphs are easy. Indeed, Edwards showed that,when restricted to class of graphs with lovvest ver-tex degree 6 > an for arbitrary a > 0, the deci-sion problem of 3-colouring is polynomial [4].

This may be understood that the 'difficult'graphs are likely to be found among 'sparse'graphs. It is known that the problem of 3-

colouring is NP-complete (even) when restrictedto graphs of maximal vertex degree 4 [6]. Here weshow that the problem can be further 'simplified',proving that the decision problem of 3-colouring isNP-complete when restricted to A-regular graphs(for A > 4). We also show that the decision prob-lem of 4-colouring is NP-complete when restrictedto A-regular graphs (for A > 5).

We assume that the reader is familiar with somestandard definitions of graph theory and of com-putational complexity theory (given, for example,in [2] and [7]).

2 3-colourability of 4-regulargraphs is NP-complete

Let us define the problem Tl(k, A) as follows:Input: A-regular graph GQuestion: Is G &-colourable?

Lemma 1 For any graph G there is a graph G'with no vertex of degree 1 or 2 such that:

G is 3-colourable iff G' is 3-colourable

Remark: G' in the Lemma is either a graph withminimal vertex degree > 3 or the empty graph,which is the case when G is, for example, a cycle.If G' is empty, it is trivially 3-colourable, and fromthe proof of the Lemma 1 it follows that also G is3-colourable.Proof (of Lemma 1): It is easy to see that thefollowing assertions are true:

(a) If there is a vertex v G V(G) with degree1, then G is 3-colourable if and only if the

60 Informatica 17 (1993) 59-63 J. Žerovnik

G

(o) (*)

Figure 1: Vertices of degree 1 or 2 may be omitted

The construction, given in Fig. 2, can be doneas follows. Take two sets, say M and N, of threevertices each. Connect every pair x,y; x € Mand y € N. Add two vertices, say u and v andconnect u to all the vertices of N and v to allthe vertices of M. Now choose arbitrary pair ofdistinct vertices of G, say w and z, and connectu with w and v with z to get the graph G'.Proof: Since the graph H is bipartite, it is easyto see that 3-colouring of arbitrary graph (G onFig. 2) can be extended to 3-colouring of graphG'. On the other hand, since G is subgraph of G',G is 3-colourable if G' is. Q.E.D.

Now we shall prove

Figure 2: G is 3-colourable ifF G' is 3-colourable

induced graph on V \ {v} is 3-colourable (seeFig. l(a)).

(b) If there is a vertex v € V(G) with degree2, then G is 3-colourable if and only if theinduced graph on V \ {v} is 3-colourable (seeFig.

In this way we can reduce any graph G to agraph G' with minimal vertex degree at least 3which is 3-colourable exactly when G is. Q.E.D.

Remark: Extracting G' successively using (a)and (b) can clearly be done efficiently.Remark: It is obvious that maximal vertex de-gree of the graph G" is not greater than maximalvertex degree of the original graph G.

Lemma 2 G is 3-colourable iff G' is 3-colourable

where G' is a graph, obtained from G by theconstruction given in Fig. 2.

Lemma 3 The problem 11(3,4) is NP-complete.

Proof: We will reduce the problem of 3-colourability of graphs with vertex degree at most4 (which is known to be NP-complete [6]) to theproblem 11(3,4).

Let G be arbitrary graph with maximal degreeA < 4. By Lemma 1 there is a graph G\ (whichhas at most as many vertices as G) and G\ is 3-colourable exactly when G is 3-colourable. If G\is empty, then we know that G is 3-colourable.

Now consider the case when G\ is nonempty.By construction, G\ is a graph with vertex degrees3 and 4. Since the sum of all the vertex degrees istwice the number of edges (^vev^" = 2|JE7|), thenumber of vertices with degree 3 must be even.

Now couple vertices of degree 3 in G\ arbitrar-ily. Connect a copy of the graph H to each cou-ple of vertices of degree 3, as defined in Fig. 2.By Lemma 2, this construction gives a graphGi which is 3-colourable exactly when G\ is 3-colourable. Q.E.D.

Remark: Graph H has 8 vertices. Since weadded at most ^ 8 new vertices, the resultinggraph G2 has at most a constant factor more ver-tices than G\.

Remark: The construction can clearly be doneefficiently.

Thus 3-colourability of 4-regular graphs is NP-complete. Now we reduce the problem of 3-colourability of A-regular graphs to the sameproblem on A + 1-regular graphs.

REGULAR GRAPHS ARE 'DTFFICULT' FOR COLOURING Informatica 17 (1993) 59-63 61

g(u)g(v)9(G)

c f(G)f(u)f(v)

Figure 3: Joining two copies of a A-regular graphwe get a A + 1-regular graph

3 3-colourability of A-regulargraphs is NP-complete

Lemma 4 11(3, A) oc 11(3, A + 1)

Proof: Let G be arbitrary A-regular graph. Nowwe give a construction of a graph G".

Take two copies of G, Gx = {VuEi) and G2 =(V2, E2). Denote with / : G -> G\ and g : G ->G2 the corresponding isomorphisms.

Graph G' = (V',E') is defined with: V = V^ UV2 and E' = E^ U E2 U {{/(«), g(v)} \ v e V}(see Fig. 3).

If G' is 3-colourable, then also G is 3-colourable,since it is isomorphic to a subgraph in G'.

On the other hand, if we have a 3-colouring 6 ofG, it is easy to construct a 3-colouring b' of G", forexample with a 'shift' of colours: b'(f(v)) =-b(v)and b'(g(v)) = (b(v) mod 3) + 1.

Since for any A size blow up is a constant fac-tor, the assertion of the Lemma follows. Q.E.D.

By induction, from Lemma 3 and Lemma 4 wehave:

4 4-colourability of RegularGraphs

Here we discuss an attempt to generalise theproposition 1 on the problem of fc-colouring. Withanalogous proof as for the case of 3-colouringswe prove a proposition for 4-colouring, while fork > 4 the time complexity of the decision problemof fc-colouring of A-regular graphs remains openfor some A.

Two of the previous lemmas are easily gener-alised:

Lemma 5 Let G' be any subgraph of G obtainedby the folloiving process: if there is a vertex of de-gree less than k, delete it. Graph G is k-colourableif and only if graph G' is k-colourable.

Proof: Assume we coloured the graph G' withA;-colours. It is easy to see that there is algorithm,which properly extends the proper colouring ofG' to a proper colouring of G. (Take, for exam-ple, vertices of G in opposite order as they weredeleted from G. When a vertex was deleted, ithad less than k neighbours, therefore there is atleast one free colour for it.) Q.E.D.

Lemma 6 For any graph G voith vertex degreesk and k + 1 there is a k + 1-regular graph G', suchthat:

G is k-colourable iff G' is k-colourable

H

Proposition 1 The decision problem of 3-colouring restricted to A-regular graphs 11(3, A)is NP-complete for A > 4.

It is known that for graphs of maximal vertexdegree 3 the problem is polynomial [6]. Hence weknow for all problems 11(3, A) whether they arepolynomial or NP-complete.

Figure 4: G is 4-colourable iff G' is 4-colourable

Proof: If there are at least two vertices of de-gree k in G, then we add a copy of graph H. Forgiven k the graph H is defined as follows. Take acomplete bipartite graph Kk,k- Add two vertices

62 Informatica 17 (1993) 59-63 J. Žerpvnik

and connect one vertex with all the vertices of oneindependent set of the Kk,k and the other vertexwith the second independent set of the Kk,k (forthe case k = 4 see Fig. 4). In this way we reducethe number of vertices of degree k by two.

If there is only one vertex of degree k in G,then we coristruct a new graph as follows: Taketwo copies of G, connect the two vertices of de-gree k with an edge. The resulting graph is ob-viously k + 1-regular and it is easy to see thatit is fc-colourable exactly when G is fc-colourable.Q.E.D.

For a proof of a generalization of the proposi-tion 1 we need a lemma of the following type: de-cision problem of fc-colouring on arbitrary graphcan be reduced to the same problem on a graphof maximal vertex degree k + 1.

In the proof of the proposition for 3-colouringwe used tlie result of Garey, Johnson and Stock-mayer. Here we give the idea of a proof for k = 4.We were not able to generalise the idea for k > 4.

Lemma 7 The decision problem of 4-colouringof graphs of vertez degree < 5 is NP-complete.

Figure 5: Graph for substituting vertices of de-gree 6

Proof (outline): The key of the proof is theidea of how to substitute vertices of large degreewith a graph of small enough maximal vertex de-gree and with property that any 4-colouring ofthe resulting graph Gf defines a 4-colouring of theoriginal graph G. Such graphs are given in Fig-ures 5,6 and 7. The graphs in Fig. 5 and Fig. 6are used for substituting vertices of degrees 6 and7, respectively. For vertices of larger degrees, alonger chain is used, as indicated on Fig. 7. Thegraphs given have the property, that in any proper4-colouring all the vertices with 'free edges' haveto be coloured with the same colour. (This colour

Figure 6: Graph for substituting vertices of de-gree 7

can be then assigned to the substituted vertex inthe original graph. The other vertices of G canthen be assigned the same colours as they hadin the 4-colouring of Gt.) We omit the details.Q.E.D.

With a straightforward generalization of theproof of Lemma 4 we have also:

Lemma 8 U(k, A) oc U(k, A + 1)

Therefore:

Proposition 2 The decision problem of 4-colouring of A-regular graphs 11(4, A) is NP-complete for any A > 5.

Again, because of the theorem of Brooks [3],the problem 11(4, A) has polynomial time com-plexity for A < 4. Thus for all the problems11(4, A) we know whether they are polynomial orNP-complete. Let us conclude with a couple ofconjectures. Since we were unable to generalisethe Lemma 7 we state

Conjecture 1 The decision problem ofk-colouring of graphs with vertez degree < k + 1is NP-complete.

If the first conjecjure was true, then we wouldhave a nice classification of time complexity forall the problems E(k, A).

Conjecture 2 For any k > 2, A > 2 the deci-sion problem of k-colourability of A-regular graphsII(fc, A) is NP-complete if A > k and is polyno-mial othertvise.

Let us conclude with a simple consequence ofthe proposition. Assume we have an algorithm

REGULAR GRAPHS ARE 'DIFFICULT' FOR COLOURING Informatica 17 (1993) 59-63 63

Figure 7: Graph for substituting vertices of degree > 5

A for 3-colouring and we want to characterisegraphs, for which the algorithm does not providethe correct solution in polynomial time. If P^NPthen for any algorithm A for each A > 4 thereexists an iniinite family F(A,A) of A-regulargraphs such that the algorithm A has superpoly-nomial complexity on each family F(A, A). If thiswere not the case for some A then A would be apolynomial algorithm for 3-colouring of A-regulargraphs, which would imply P=NP!

Acknowledgement. The contents of the pa-per is based onapar t of author'sdissertation sub-mitted to University of Ljubljana under supervi-sion of Tomaž Pisanski. Thanks are also due tothe anonymous referees, who helped to improvethe quality of presentation considerably.

References

[1] E.A. Bender, H.S. Wilf, A Theoretical Anal-ysis of Backtracking in the Graph ColoringProblem, Journal of Algorithms 6 (1985)175-282.

[2] J.A. Bondy, U.S.R. Murty, Graph Theorytvith Applications, Amer. Elsevier, New Vork,1976.

[3] R.L. Brooks, On Coloring the Nodes of aNetwork, Proc. Camb. Phil. Soc 37 (1941)194-197.

[4] K. Edwards, The Complexity of Colour-ing Problems on Dense Graphs, TheoreticalComputer Science 43(1986) 123-147.

[5] J.A. Ellis, P.M. Lepolesa, A Las VegasGraph Colouring Algorithm, The ComputerJournal 32 (1989) 474-476.

[6] M.R. Garey, D.S. Johnson, L. Stockmeyer,Some Simplified NP-complete Graph Prob-lems, Theoretical Computer Science 1(1976) 237-267.

[7] M.R. Garey, D.S. Johnson, Computers andIntractability, W.H. Freeman and Co., SanFrancisco, 1979.

[8] A.D. Petford, D.J.A. Welsh, A Randomised3-Colouring Algorithm, Discrete Math., 74(1989) 253-261.

[9] J.S. Turner, Almost All /b-Colorable GraphsAre Easy to Color, Journal of Algorithms 9(1988) 63-82.

[10] J. Žerovnik, A Randomized Algorithm forfc-colorability, Discrete Mathematics (to ap-pear).

Informatica 17 (1993) 65-80 65

METAPHYSICALISM OF INFORMING

Anton P. ŽeleznikarVolaričeva ul. 8, 61111 Ljubljana, [email protected]

Keywords: counterinforming, cyclicity, informational embedding, informing, intelligence, metaphys-icalism, parallelism

Edited by: Jifi SlechtaReceived: January 26, 1993 Revised: February 18, 1993 Accepted: March 1, 1993

This paper is an introduction to the phenomena of metaphysical informing occurringwithin an informational entity [Železnikar 92a, Železnikar 92b]. The basic question ishow to structure and how to organize the processes of informing vvithin the metaphysi-cal triplet of informing, counterinforming, and informational embedding, which performcycHcally, in parallel, and spontaneously in a complex entity-metaphysical cycle. Theproblem of the so-called metaphysical informing (called metaphysicalism) has to besolved conceptually and, then, constructively, that is, in a language-formalized (ma-chined) way. The open cyclic-parallel and spontaneous informing hides the potentialityof intelligence (intelligent information) v/hich through informing, counterinforming, andinformational embedding of an Informational entity in question comes to the informa-tional surface (of an observer). The aim of this paper is to expose certain possibil-ities of metaphysical informing within machines, programs, and tools performing inan informationally arising environment. Metaphysicalism, that is, cydic, spontaneous,entity-intentiona,l and informationally open informing, seems to be the most fundamen-tal problem which has to be solved formally (constructively) on the way to informationalmachine. To some extent, the possibilities of such proceedings are already visible.

Ich wurde hier sagen, dafi uns ein Vor- spontaneous informing. The metaphysical per-stellungsakt als solcher direkt anschaulich tains to an entity's circular and parallel inform-wird, wo wir gerade diesen Unterschied ing, which is spontaneous, that is, speaking gen-zivischen Vorstellung und Vorstellung erally, unforeseeable, unpredictable to certain ex-dieser Vorstellung phanomenologisch tent (sense) or informationally arising, however,konstantieren. entity-intentional, structurally and organization-

ally oriented or, simply, informationally persever-—Edmund Husserl [Husserl 00] 11/1 508 ing.

Informing as a phenomenon of an informa-tional entity is the entity process, by which the

1 IntroductlOn entity arises, that is, develops, maintains bothits structure and its own and from the environ-

Metaph^sicalism1 (called also metaphysically cy- m e n t impacted phenomenality, acts in an infor-clic informing [Železnikar 92b]) is a term denot- mational way. Informing as an entity active com-ing the interior phenomenalism of an informa- ponent performs, as we say, through informing pertional entity. Metaphysicalisin of an informa- Se, counterinforming, and informational embed-tional entity means its own circular-parallel and ding. In the triplet informing-counterinforming-—r=— : embedding, the informing is a regular informa-

lhis paper is a pnvate authors work and no part of .. , , , . . .^, .,it may be used, reproduced or translated in any manner t i O n a i phenomenon being in accord Wlth the en-whatsoever without written permission except in the case tity normal intention, its phenomenalizing in-of brief quotations embodied in critical articles. formational stream, keeping the entity identity,

66 Informatica 17 (1993) 65-80 A.P. Železnikar

structure, and organization. In contrary to in-forming and in regard to it, the counterinformingis a disturbing informing component which arisesduring the process of informing, as a consequenceof interior and exterior informational impacts. Atthe first glance, as a result within an entity in-forming, counterinforming is not well-structuredand well-organized yet, it is not informationallywell-connected in respect to the ruling informing,which determines the character of informationalentity.

Informational embedding as the next compo-nent in the informing of an entity has to connectproperly the arising and from the environmentarriving informational items to the informationalbody of the entity. Embedding as a form of in-forming is arising according to the phenomena ofthe arising information within counterinformingas well as the phenomena of the arriving informa-tion from the entity environment.

In the pointed sense, metaphysicalism is noth-ing else than a common term for the informa-tional phenomenality within an informational en-tity, within which the phenomena of informing,counterinforming, and informational embeddingoccur. This interior phenomenon of an inform-ing entity' is not necessarily evident for tlie en-tity exterior observer and, as one may say, re-mains concealed to a certain informational extent.The aim of this paper is to analyze and to de-termine these phenomena formally by means ofthe informational language [Zeleznikar 92a] and,through this formalization, to capture the con-ceptuality of informing of an informational ma-chine implementation [Železnikar 92c]. Similarneeds can arise within the so-called knowledgearchives projects [Knowledge 92], where knowl-edge and components of knowledge can emergeand have to be determined informationally.

2 Formalizing Some BasicAxioms of Informing

How does an informational entity, marked by a,inform and in which way is it informed? We dis-tinguish four basic types of a's informing calledetternalism, internalism2, metaphysicalism, and

phenomenalism [Železnikar 92b].The externalism3 of the informational operand

a means the possibility of a to inform other enti-ties and itself, called also a 's informing(ness) forothers and itself. The informingness of a is its ba-sic (potential) property (predicate, physical phe-nomenon) marked by the general informationaloperator |= on its right side. The externalism ofentity a is represented by informational forrnulaa (= and reads a inform(s). Thus, a \= is anopen formula (with the open right side of opera-tOT | = ) .

Tlie internalism of the informational operanda means the possibility of a to be informed byother entities and by itself; it is called also a 's in-formedness by others and by itself. The informed-ness of a is its basic (potential) property (pred-icate, physical phenomenon) marked by the gen-eral informational operator (= on its left side. Theinternalism of entity a is represented by informa-tional formula [= a and reads a is/are informedor a is/are being informed. Thus, |= a is an openformula (with the open left side of operator |=).

The metaphysicalism of the informationaloperand a means the possibility of a to informand to be informed by itself; it is called also a 'sinformingness and informedness in itself. The in-terior cyclic and spontaneous informingness andinformedness of a, called a's metaphysicalism, isits basic (potential) property (predicate, physi-cal phenomenon) marked by the general infor-mational formula (expression) a \= a. This for-mula reads a informs and is being informed meta-physically or, in an informationally general way,a informs and is being informed cyclically andspontaneously in itself. However, metaphysical-ism a \= a is in no way a closed formula depend-ing solely on a's own (internal) informingness andinformedness.

The last of basic informational axioms is calledphenomenalism4. The phenomenalism of entity ais a consequence of its externalism, internalism,and metaphysicalism, is an informational system

The informational internalism may be comprehendedas a subjective informational phenomenalism, phenomenal-izing the world and the entity into the entity in question.

3Informational externalism is called also informatioprima because of the basic informational hypothesis thateverything, which is, informs.

4Informational phenomenalism is the most general prin-ciple, by which things inform and are informed in variousways, e.g. physically, biologically, socially, etc. Phenome-nalism may not be replaced by phenomenology, vvhich is aphilosophical discipline (for instance, [Husseil 00]).

METAPHYSICALISM OF INFORMING Informatica 17 (1993) 65-80 67

of those phenomena and a systemic expressionof two basic formulas (connected in parallel by asemicolon) a |=; (= a. Thus, the previous formu-las are subformulas of a's phenomenalism, thatIS

(« K .1= «» a \= a) C (of:p; j= a)

Informational operator C marks the subinformingentities (externalism, internalism, metaphysical-ism), separated by commas, within the informingentity (phenomenalism).

The four basic axioms, which pertain to exter-nalism, internalism, metaphysicalism, and phe-nomenalism, are forms of the so-called informa-tionalism concerning basic modes of an informa-tional entity informing (in Latin, modi informa-tionis). We have formalized them by senseful for-mulas being derived from basic axioms pertain-ing to entity a [Železnikar 92b]. These formu-las are expressions. Externalism means alwaysan expression, that is, exteriorization [Derrida 67](output phenomenalism). On contrary, internal-ism means an impression, that is, interiorization(input phenomenalism). Both externalism and in-ternalism carry meaning (in German, Bedeutung[Husserl 00]). This interwoven meaning causesthe expression and impression of an entity meta-physicalism to an exterior observer together withthe intention of circular and spontaneous meta-physical phenomenality. Similar could be said fora's phenomenalism.

Within this expressional and meaningful scopethe following can be concluded: metaphysicalnessof a possesses its own externalism, internalism,metaphysicalism, and phenomenalism. And, alsoall externalism, internalism, and phenomenalismof a possess each own metaphysicalism (meta-physical recursiveness). But by cyclic and par-allel decomposition of metaphysicalism a |= a,the marker (informational operand) a developsand is being developed through the arising of itsmeaning (contents, significance, structure, orga-nization, informational broadening). In parallelto the cyclic decomposition, as a consequence ofa straightforward metaphysical analysis, parallelmeanings can emerge and in this way altogethercan be composed into a complexly developingscheme of the initial (symbolic) metaphysicalismo (= a. A metaphysical informational systera con-cerning entity o is coming into existence throughinforming of a and it concerning environment.

3 Problems of Informing of anInformational Entity

The question is how to determine, conceptualize,design, construct, organize and, lastly, implementthe process of informing as a regular activity ofan informational entity. What to say in accordto informing which represents the active inform-ing component of an informational entity? Withthe last question, the duality of informingness (ac-tive component) and informedness (passive com-ponent) is implicitly introduced into the meaningof informational entity. The preceding questionsare on the way to possibilities of an informationalmachine implementation which, in its particularcases, reduces in, for instance, electronic dictio-nary, knowledge archives [Knowledge 92], experttools, or intelligent machine.

The general informational operator |= is, fromthe view of informational operand a, an im-plicit (to a belonging) expression of informing of(within) operand a. Informing of operand a canalso be explicated in the form of an operand en-tity, marked by Ia, or by the functional (predica-tive) form I(a). A more directly correspondingnotation for this kind of informing which pertainsto a vvould be simply A. The correspondencebetween entities a,/3,j,- • -u and their informingwould be A, B,C, • • • Z, respectively.

3.1 Basic MetaphysicalDecomposition

Within an informational entity a, the outmostmetaphysical cycle will be a \= a. By definition,through the most primitive metaphysical decom-position, there is

a "Def (£iAs informing Xa is introduced, it represents ana-inner metaphysical cycle in the form

"Def

where ^Def is read as informs or means by defini-tion. In the first step of analysis of informing Ia,we put the question how could Ja be positionedand attitudinized within a and vice versa or, howdo a and Xa impact each other dynamically in an


informational way. The basic informational cyclespertaining to informingness and informedness ofboth entity a and its informing Xa are, in fact,manifold, e.g.,

(a |= Xa) \= a<*\=(Ia\= a)

a-metaphysicalism 1a-metaphysicalism 2

(Xa |= a) \= Xa IQ-metaphysicalism 1Xa \= (a |= Xa) JQ-metaphysicalism 2

(3)

The question is, in which way a particular meta-physical cycle could imply the alternative ones.Thus, considering all possibilities of formula sys-tem 3, hypothetically,

U \= i) N 0t\=(v\= 0;

(4)

where informational operator =$> represents theinformational implication and reads as informs inan implicative way or, simply, implies.

In formula system 3, all formulas, that is, (a \=Xa) |= Q; a \= (Ia |= a); (Ia \= a) \= Xa;and Ia (= (a |= I o ) , are metaphysical and de-duced (decomposed) from the basic metaphysical,that is, informationally cyclic form a (= a andits inner consequence Ta \= Ta. Sometimes, weunderstand these basic formulas as the shortcutsfor a's and Xa's metaphysicalism which complex-ities are hidden in the general informational op-erator (=. Thus, the decomposition of a \= aand Xa (= JQ concerns, in fact, the cyclicallyconnecting informational operator |=. We shalldeal with more complex and parallel basic infor-mational and composed (also perplexed) informa-tional cycles within an entity metaphysicalism.

3.2 Spontaneity of Informing

In the second step of our investigation we rise thequestion of possibilities of (inner, metapliysical)informational spontaneity of informing Xa, thatis, of the phenomenon of informational arising ofentity a.

Informational spontaneity does not mean an ar-bitrary development of an entity contents, struc-ture and organization, that is, of an arbitrary

entity-informational broadening, meaningful fil-tering or notional purification. In an informa-tional situation and attitude, entity a alreadyhas its informational structure and organization(course, orientation, intention, behavior) and inits own way filtered informational impactednessof its exterior (informedness) on disposal, both inthe form of informing Xa and entity a a s a whole(including also the temporarily passive compo-nents of its arising informational structure). In-formational spontaneity of a is caused intention-ally by itself and it impacting exterior and is inaccord with its instantaneously changing (arising)orientation (worldliness).

We have to answer the question of spontane-ity pertaining to the entity-informational in-tention in a constructive way. Which mecha-nisms for informational spontaneity simulation,modeling, and organization are realistic, ma-chine realizable, and conceptually possible? An-svvers to the last question are various and de-pend on particular situation, for instance, mul-timedial, pictorial, acoustic, and linguistic, per-taining to signals, data, written text, dictionar-ies [Dictionary 90a, Dictionary 90b], knowledgearchives [Knovvledge 92], etc.

Spontaneity means, for instance, a free mov-ing along an existing informational net and tak-ing with informational items which correspond,coincide, fit, match, etc. an informational situ-ation and connect, interweave, embed, interpret,etc. them into, in, and within a concrete infor-mational entity. This model of spontaneity couldbe called a model of free informational associa-tion. In fact, the decision, which items in the netto take with, is spontaneous, depending on somedistinguished informational attitudes concerningthe entity in question and its informational envi-ronment.

It is to stress that spontaneity as such is an in-formational entity by itself, which is a constitutivepart of the informational entity and of to the en-tity pertaining informational environment. Spon-taneity is intentlonal, concerns a goal-oriented in-formation, is in no way something information-ally surprising and quite unexpected. It arisesfrom an informational impulse, tendency, but un-planned as an internal force or cause. Informa-tionally spontaneous strategies, procedures, ran-dom associative processes, and like that have to


be conceptualized and used as system supportingmechanisms for spontaneous informing of entities.

3.3 Some Inner States of CircularInforming

How does an informational entity a begin to in-form and how does it inform from one situationinto another? The beginning of an entity inform-ing XQ is caused by the appearance of markero, which is the most simple expression and car-ries the most primitive meaning, for instance,an identifier, a headword, symbol, token, sign,etc. In this very beginning situation, inform-ing la informs the basic meaning being differentfrom nothing5. In an informational environment,the informing of initial entity a is being sup-ported, for instance, through its physical environ-ment, informational machine, knowledge archive,informational dictionary, living actor, linguisticsystem, multimedia exterior, associative mecha-nisms, dispersive algorithmic procedures, artifi-cial surroundings, etc. Several entities can be in-formed of the occurrence of an initial entity a andcan support its informational arising in differentinformational ways.

How does this inner development of inform-ing proceed and which are the proposed (defined)mechanisms, structure, organization, in short, theentity metaphysicalism? We have to develop asystematic approach to the problem of inform-ing, which could be applied in cases of an in-formational environment implementation, for in-stance, in an informational machine or even in acomputer, which can model an adequate informa-tional environment.

Informing Xa is an active part of entity a.Sometimes, by Ja, the whole informational ac-tivity of entity a is meant. But the emphasis ofinforming as a distinguishable entity within anentity as a unit is in its includedness or participa-tion, that is,

In the metaphysical sense, there is

(6)

a C Ot) (5)

4The nothing means the nonappearance of somethingin an informational context. Hovvever, as soon as we speakabout the nothing of something, the something appears inthe informational context and becomes an informing entity.Through this, the informational existence of something (asnothing) cannot be denied anymore.

irrespective of the possible informational struc-ture of cycle Xa (= la. It is understood that in-forming Ia is an informationally subordinated ac-tive component of informational entity (unity) a.According to definition 5, as a consequence of thebasic metaphysical includedness of Xa in a, wecan introduce the includedness of correspondingJa-cycles,

(IQ C a)

((Ia N I«) C (a \= o)) =((iQh«)N^)c\((af=Ja)h«);(la \={a^= Xa)) C(a \= (Xa \= a))

(7)

for the corresponding short-form metaphysicaldecompositions. The decomposition procedurecan extend further to longer and longer forms ofcycles considering the components of informing,counterinforming, and informational embedding,that is, considering the generalized idea of meta-physical informing as discussed in the previoussections. In concrete cases, these general termswill be particularized and, certainly, decomposedin specific (particular) ways.

Within the general context, we can speak aboutshort, medium-sized, and long metaphysical cy-cles of informing. If we introduce the measurefor metaphysical length ^meta of an informationalformula <p, that is, metaC ), then imeta.{Xa) = 0,

\= Ia) = 1,& N Ca) \= la) |= £a) J= ea) \=.

a) |= la) = 6. In the general case of informingwith counterinforming and embedding within en-tity o, the maximal length has the value 6 (a longmetaphysical cycle). In concrete, particularizedcases, the metaphysical length can be extendedby proceeding into a greater detail of the infor-mational problem.

Which are the medium-sized metaphysical cy-cles of informing? We determined an entity in-forming in a basic (formula 2) and in a com-plex way (formula 8), considering counterinform-ing Ca with counterinformation fa and informa-tional embedding £a with embedding informa-tion eQ. Counterinformational and embedding-informational metaphysical subcycles can inform


within the long metaphysical cycle of informing(formula 17).

The long metaphysical cycles concerning in-forming Ta are, systematically,

ya \= {Sa \= (e« N (« N

N ) (M

a |= ca) N 7«) N £«) N^(ea |= (a (= IQ));

((((!„ h Ca) |= 7«) h £a) \= £«)K (« f= J«);

o |= Ca) (= 7«) 1= Sa) |= ea)

(8)

c (iQ |= J«)

where 1=^ marks the so-called point of main (sym-bol /i) cyclic informing.

entia, diversity, associationism, dissociation, frus-trations, etc.) comes to the informational surface.An example of counterinforming in a linguisticway is the coming up of antonyraous meaning toan existing synonymous meaning of a headword,phrase, or sentence and connecting such phenom-ena of meaning reasonably to the already rec-ognized meaning (knowledge). of the original in-formational item. This kind of counterinformingmight mean nothing else than an additional in-terpretation of the original meaning, broadeningthe original meaning in significant and simulta-neously various and varying ways. Such emergingof information concerning a distinct informationalentity is common and obvious vvithin human cul-tures performing discourse, confrontation of be-liefs, democratic dialog, brainstorming, etc.

Counterinforming CQ is a substantial genera-tive part of informing la and, by definition, isthe producer of the counterinformational entityj Q . There is no essential difFerence in regard togeneral formulas 1 and 2 and thus

4 Problems ofCounterinforming of anInformational Entity

To open the possibilities of counterinformingmechanisms, that is, their informational imple-mentation, we have to make a short overview con-cerning concepts of counterinforming which havetheir roots in, for instance, linguistic meaning ofboth words counter and informing. Within thequestion of how to implement the process of coun-terinforming, by which in regard to the informingentity a new, difFerent information is coming intoexistence, we can use various concepts concerningthe meaning of the word counter (adverb, verbtransitive, and word prefix) while the meaningof the word informing is consequently connectedwith the word information, understood in an ex-tended sense.

To counterinform can mean to inform counterto the instantaneous course, intention, ruling, per-severance, phenomenalism of an informing en-tity. This can happen in a pure informationalway, where in parallel to the straightforvvard-ness of an informational entity always to it con-trary phenomenalism (e.g., doubtingness, differ-

"Def

fa ^Def (

((Ca \= 7«) N Ca); \

«7« N C ) t= 7a); A(7a (= (C. (= 7«)) j

(9)

The essential difference between informing Xa andcounterinforming CQ in comparison to formula 5is the following:

la C Ca) Cla)Ca (10)

We see how counterinforming Ca is information-ally subordinated (included) in informing TQ irre-spective of other possible inclusions pertaining todifFerent informational entities.

Metaphysicalism of counterinforming is a phe-nomenon with various faces, also in respect to thepure formalism, that is, to different possibilities offormal expressing. Which are the possible forms(all of them in a given situation) of counterin-formational metaphysicalism vvithin an informingentity a? We can develop (decompose, compose)metaphysical concepts proceeding from the short-


est to the longest cycle, for instance,

C \= C •

(Ca |= 7a) N CQ;iCa\=la)\=EQ)\=Ca\(((Cat=la)\=€a)\=eQ)\=Ca;((((£» N 7«) 1= £») t= £a) h «) 1=(((((C« 1= 7«) M « ) M « ) h «)*=

(11)

(12)

Formulas of system 11 are not in a final shapesince they can be metaphysically decomposed inregard to each in them occurring operand entity.For the last formula of system 11 we can introducemetaphysical markers for ~fa and a, for instance,

((({(Ca N (7« f= 7«)) 1= €a) |= ea) \=(a \= a)) |= Ia) \= Ca

Under such circumstances, in the process of meta-physical decomposition, also a direct informa-tional connection between entity a and its coun-terinformational entity j a can come into existenceas a parallel formula to the informationally aris-ing metaphysicalism. This process is in no wayarbitrarily spontaneous, it keeps the route of in-formational intentionality of a and it information-ally influencing environment.

5 Problems of InformationalEmbedding of anInformational Entity

Informational embedding is a process by whichthe arisen counterinformation and from the exte-rior of an informing entity arriving information isinformationally embedded (connected, meaninglyassociated, interpreted) into the informing entity.At this attempt of investigation, we shall not re-search into particular details, by which in a cer-tain case the embedding process could be deter-mined according to some concrete informationaldemands.

Informational embedding £a is a part of a's in-forming and, by definition, the producer of theembedding information sa. There is no notionaldifference in regard to formulas 1 and 2, thus

"Def

"Def

((£a\=eQ)\=£a);

£a) |= ea);

The essential difFerence between informing Xa andinformational embedding £a in comparison to for-mula 5 is the following:

((((£« C £a) Cla)cCa)C Ia) Ca (14)

We see how informational embedding £a is in-directly, via counterinformational entity fa andcounterinforming Ca, informationally included ininforming la.

To interpret concretely informational embed-ding £a and by embedding produced embeddingentity (information) ea, let us introduce under-standing Ua instead or as a part of £a and mean-ing fia of an interior or exterior entity /?, that is,Ha(/3). In accord to formula system 13 and for-mula 14, there is ,

(15)c ua) c 7a) c cQ) da)ca

(13)

Meaning /J.a(P) is, for instance, an interpretationof meaning pertaining to entity /3, produced byunderstanding Ua. Additionally, entity Ua canconcern other meanings of other informationalsubjects, e.g., fia(£),iia(r}), etc. To mean meansto apply different informational modi or rules(modes) of inference, known (in Latin) as modusponens, modus tollens (affirmative and negativemode in traditional logic, respectively), modusobliquus (e.g., logic of absurdity, for instance,devious inversion, adjustment, or directionality,which appears together with the direct), modusrectus, modus procedendi (logic of intentionalityand processing to reach a certain goal, respec-tively), modus vivendi (logic of tolerant coexis-tence, for example), modus possibilitatis, modusnecessitatis (modal logic), etc.

6 A Constitution of theMetaphysical Cycle asInforming,Counterinforming, andInformational Embedding

So far, the structure of metaphysical cycles wasdetermined with basic informational componentsand their length was measured by the number ofoccurring informational operators (= within the

72 Informatica 17 (1993) 65-80 A.P. Žcleznikar

cycle. Let us remind and systemize the previouslydisctissed metaphysical cycles of various lengths.We distinguish several types of metaphysical cy-cles belonging to an informing entity. Thesecycles can be classified as, for instance, primi-tive, basic, medium-sized (counterinformational-embedding), and long (the longest, holistic) meta-physical cycles.

6.1 Primitive Metaphysical Cycles ofan Informing Entity

Although primitive metaphysical cycles of an in-forming entity are trivial, they are the startingpoints, from which the cyclical decomposition be-gins. The longest metaphysical cycle belongs tothe informational entity a in question. Its inform-ing component Xa is for one step shorter, etc,thus, at the end, embedding informational entitye remains, at the present state of decomposition,in its trivial form. At the beginning, the primitiveinformational cycles are the following:

a\= a informational entityIa |= 2"a informingCQ \= Ca counterinforming7a 1= la counterinformational entity - (16)£a (= £a informational embeddingea |= ea embedding-informationa!

entity

In the beginning conceptual state of decomposi-tion, the basically extended metaphysical cyclesare as follows:

h ca) (= 7«) 1= ea) h((o \= iQ)

£ h c«) h 7«) h 4.) 1= *.)(((Ca h 7») h £a) h e«) h CQ;((7a 1= €a) \= £c) \= 7«!

£a \=ea

(17)

However, we shall see, how these initial cycles willbecome as long as the longest cycle of entity a(formula 25) because of the strict circular natureof a's informing.

6.2 Basic Metaphysical Cycles of anInforming Entity

We distinguished six informational entities as ba-sic components of an metaphysical cycle belong-

ing to an informing entity, namely: a as the in-forming entity itself in its wholeness and to itbelonging informing Ia; a's counterinforming Ca

and by it informed (produced) counterinforma-tional entity fa; and, lastly, a's informational em-bedding £a and by it informed (produced) embed-ding information (entity) ea.

All short metaphysical cycles of the involvedinformational entities tt,Io,Ca,7a,£a> and ea arethe following:

(18)

Informational operator € is read as informs in thecontext of a set of entities.

6.3 Metaphysical Cycles Pertaining toInforming Xa

Informing Ta of informational entity a is the innermechanism of the activity of informational entitya. Within this cycle of informing several othercycles inform, for instance, both the counterinfor-mational and the embedding-informational one.

The primitive metaphysicalism of a's inform-ing (the starting point of metaphysical decompo-sition) is expressed by formula Ja \= Ja. Thenext step of metaphysical decomposition are twopossible short cycles, that is,

{la \=a)\= I a ; Ia \= (a |= Ia) (19)

Counterinforming Ca and informational embed-ding €a are the informing components (parts) ofinforming Xa. Thus,

(Ca t= Ca) C Ia; {£Q \= SQ) c (20)

with (7a |= 7O) c Ca and (ea (= ea) C Sa. Theremay not exist an informational includedness be-tween entities Ca and £a.

Several medium-sized metaphysical cycles con-


cerning la can be observed, for instance:

( C h ) | )(lQ\=Ca)\=(*fa\=Ia);Ta \= (Ca \= (7« N I«));

la\=(SQ\=(eQ\=la));(((2a N C«) [= 7«) 1= £») h â;P« N c«) N 7«) N (£» h -2-«);(Ia \= Ca) f= (7a |= (£a (= Ja));2« 1= (C« (= (7. N (£« N 2-«)));

(21)

«) |= (7a N (*« 1= (ea2a h (Ca \= (7a 1= (£« 1= 2a))))

The next metaphysical cycle of informing, la |=Xa, belongs already to the long metaphysical cycleof Xa and will be treated in subsection 6.5.

6.4 Counterinformational-embeddingMetaphysical Cycles

It seems reasonable to discuss counterinformingand informational embedding within common in-formational cycles. Counterinformational entity7O, produced by counterinforming Ca, has to beinformationaUy embedded before it could becomelost in the informational realm of the informingentity a. The basic question within this contextis what does the arisen counterinformational en-tity j a mean at all. The answer to the meaningof 7 a is its embedding, that is informational con-nection into the informational realm of informingentity a.

We introduce the following informational in-cludedness hierarchy concerning the basic meta-physical counterinforming-embedding subcycle:

iCa h Ca) C Ia) C a;

\â r a) •" Def

(22)) )

ea) |= Ca) \= 7«;Ma) h 7a) M«;

6.5 Long Metaphysical Cycles

The long metaphysical cycles6 are deduced fromthe most primitive forms, marking the involvedinformational entities within an informing entitya. The most simple surveying definitional schemeis, for instance,

a "Def

/ 2 a ; \

(a\=a)

\

(23)

\ £* JJ

We see how the metaphysical component a \= aremains informationally open because of the pres-ence of the system formula a |=; f= ot in whicha's metaphysicalism is recursively open, that is,can inform and can be informed both interiorlyand exteriorly in concern to entity a. The meta-physical openness of entity a means that meta-physicalism of a informs and is informed, that is,(a |= a) f= and (= (a \= a), which is a recursiveproperty of system a (=; a. This definitionalscheme can be expressed in the primitive meta-physical form, which is

a |=; |= a;

(a\=a)£a\=£a

(24)

Formula 23 and formula 24 are merely the initialschemes for the construction of the so-called longmetaphysical cycles of informing within informa-tional entity a. The implicational part of the long

6The term long tnetaphysical cycle concerns the longestcyclically structured operand-operator informational for-mula, which considers all the identified components occur-ring in a metaphysical cycle, constructed by a constructinginformational entity. Within characteristic systemic com-ponents (informing, counterinforming, counterinformation,embedding, and embedding information) of an entity, con-crete infoimational entities appear (look at, for example,formula 30), which can concretely lengthen the conceptu-ally basic metaphysical cycle.


metaphysical cycle of entity a's informing is

( a (= a)iong Def

<X \= la) \= Ca) |= 7«) 1=(= e a) |= a ) = •

•a \= Ca) \= la) |= £a) (= ea).1=^)1=2«;

» N 7«) N ) t= e«) h «) (25)|= 2«) |= Ca;

NCQ)h7*;

h -v l b £ •

i F"JF xaj P flj p 7a

In long metaphysical cycles, entity a observes itsconstituting parts Ia,Ca,la,£a, and ea, but theseparts can observe the entity as a whole too. Inprinciple, all parts can observe each other andthe entity as a unit. This situation is in no wayprincipally contradictory. Of course, the ques-tion arises, how do different parts "know" thewholeness of entity a, also this knowledge can beequally exhaustive for all constitutive entities, ifit is supported (delivered) from a central informa-tional place, where all information concerning in-formational entity a is collected (e.g. in an infor-mational machine, in the informational operanddictionary [Železnikar 92c]). In this way, in anymetaphysical cycle, within which a occurs, entitya is one and the same entity.

The other approach would be to understand en-tity a, at any place or situation it appears, as asample of a. For instance, in a formula system,

(a 1= Ia) \= a; (Itt \= a) \= la

entity a in both formulas could mean one andthe same entity. In a different situation, entitya in the first formula could represent one sampleand, in the second formula, the other sample ofone and the same thing. This situation is not sosurprising as it might be understood by the tra-ditional philosophy (readiness-to-hand). Similaraffairs take place in a living mind, where difFerentsamples of one and the same thing occur vvithindifferent mental situations and attitudes.

7 Metaphysicalism of anInformational Entity

Metaphysicalism is a common notion, by whichthe internal informing of an informational entityis determined. This informing is entity-cyclic andsensitive in respect to the entity environment ina semantic-pragmatical (ontic, ontological) way.The entity-cyclic pertains to the informing, coun-terinforming, and embedding-informational na-ture of an entity. The semantic-pragmatical ofan entity concerns the internalization of externalentities, so that they inform within an entity'scyclicity. This cyclicity is structured in paralleland has its short, medium-sized, and long cy-cles, which pertain to internalized external enti-ties. In this way, metaphysicalism is not only asimple, direct, and straightforwardly shaped in-formational cyclicity determined once and for aH.It is an entity-ontological process, which developsby entity informing in its own way and by meansof entity-sensitive impacting of exterhal entities.We shall show the semantic-pragmatical part ofmetaphysicalism by an example of intelligent in-formational entity i in section 8.

Metaphysicalism is a composed and parallelstructured cyclicity of an informing entity, whichhas its internal intention, but remains open forexternal informational impacts. In this respect,it is a particular view of possible informing of anentity, of its individual and externally impactedprocessing, in which the identity of an informa-tional entity emerges in an internal way and, foran external observer, also in an externally influ-enced way.

In this section we have to show the metaphys-icalism of an informational entity in its entirety.We have to join the results obtained in the pre-vious sections and interpret them in a compactform. So, let us make a verbal compilation of dis-cussed metaphysical possibilities.

Decomposition of an entity a metaphysicalismstarts by the trivial formula a |= a. This formulaacts as a title (idea as the linguistic-informationalmeaning) in the top-down design of metaphysical-ism. Within this initial situation, metaphysical-ism can be tackled by two basic ideas of itiform-ing. The first one is in the cyclically based tripletinforming, counterinforming, and informationalembedding, which is the fundamental way of an


entity informing. The second idea is semantic-pragmatical, which particularizes the first conceptand makes it more concrete. In fact, the secondidea (iti section 8) is an informational projectionto the first one.

The next step further from the trivial situationa \= a is the introduction of informing {Ta), coun-terinforming (Ca) with counterinformational en-tity (7o), and informational embedding (£a) withembedding informational entity (ea)- In this waya long basic metaphysical cycle is obtained and allshorter cycles can be derived as parallel informingcycles. At this step of development more concreteentities, particularizing the previous ones, can bebrought into the game. For mstance, intelligentinformational entity t in section 8 is a good exam-ple of a general concept of an intelligent system,which concerns intelligently an entity a.

The question "What is the informing Xa ofan informational entity a?" particularizes entityla with the next question, which is "What arecounterinforming Ca with counterinformationalentity 7 a and embedding £a with embedding-informational entity £a?" In this point of view,both 7 a and ea are certain results of entities Ca

and Sa, respectively.

For instance, we can introduce some strate-gic functions and their results as counterinforma-tional entities (Ca and ja) in the form of intention,significance, sense, etc. Embedding componentscan embrace certain sensing, observing, perceiv-ing, etc. situation in concern to an observed in-terior or exterior entity. Then, informing IQ pro-duces cyclically a result (e.g. meaning to the un-derstood situation), etc. In this manner, a prag-matical way of a's informational structure andorganization remains open for possibilities of fur-ther development, improvement, intention, e tc ,that is, an entity's metaphysicalism.

8 Intelligence as anInformational Entity'sMetaphysicalism

The question of intelligence can be tackled by theinformational theory of metaphysicalism in an in-novative, that is, informationally arising, circular,and creatively spontaneous way. Informationalschemes of intelligence become highly parallel and

circularly perplexed; this yields together with in-formational formula systems, an informationallyarising formula system, describing parallel, circu-lar, and intervroven intelligent informational phe-nomena. On this basis, intelligent entities canbe treated as emerging systems, which can bemodeled (machined) by the proposed metaphysi-cal conceptualism.

Intelligent informational entities concern sev-eral other mutually perplexed informational en-tities, for instance, understanding, meaning as aresult of understanding, consciousness as a specifi-cally circularly structured informational phenom-enality, phenomena of observing, perceiving, con-ceiving, concluding, comprehending, etc. produc-ing observation, perception, conception, conclu-sion, comprehension, meaning, etc. as intelligentinformational items, respectively. Further, intel-ligence concerns knowledge, truth, belief, faith,significance, e tc , which meaningly overlap eachother and form a redundant informational over-lapping, that is interweavement, parallelism, com-munity. Only an informational entity possess-ing some of these characteristics as commonlyrecognized properties occurs as intelligent, thatis, informationally satisfactory in an intelligentway. Thus, the intelligent means to have a suf-ficiently dynamically phenomenalizing meaning,contents, externalism, internalism, and, certainly,metaphysicalism.

Let us structure metaphysically intelligent in-formational entity to some extent, proceedingfrom basic informational cycles to more cyclicallyand parallel complex ones. As we said in the pre-vious paragraph, intelligent information must besuch and such when conceptualizing it from a ba-.sic point of view.

We can build up a hierarchy of intelligent func-tions as they appear in a circular and spontaneousact of understanding. At the bottom is the capa-bility of sensing, which is followed by functions ofobserving and perceiving of sensed informationalentity. Then, on this basis, the conceiving gen-erates concepts, which concern observation andperception of something. A mutual game betweenperception and conception entity delivers conclu-sions in the framework of consciousness, which isa kind of comprehension. On the top of intel-ligent informing metaphysicalism are cycles andparallels of understanding, which produce topi-


cal and detailed meaning of something and be-have in their individual and common intelligentways. Products of such informing are senses, ob-servations, perceptions, conceptions, conclusions,consciousnesses, comprehensions, and meanings,which may embrace other essentially conscious orunconscious entities as there might be intention,significance, sense, identification, etc. within theinformational garae. AU those components canbe informationally determined, structured, orga-nized, and formalized.

Now we can sketch some initial attempts of astrategy, by which concepts of intelligence couldbe informationally implemented. An intelligentinformational entity i informs and is informed inan intelligent metaphysical way and intelligenceis always demonstrated in concern to a concretesubject, that is, concrete informational entity, saya. An intelligent entity phenomenalism is markedsimply by

<•(«) ^ D e f (<-(«) Nintelligent; Nntelligent (26)

Intelligent entity t is an informational function ofentity a and notation t(a) expresses this func-tionality. The informational arising of intelligententity i depends on informing of entity a. Fur-ther, i(o) is a regular informational entity in-clusive with its components, which are inform-ing IL(a), counterinforming C t(a), counterinfor-mational entity 7 ((a), informational embedding£ t(a), and embedding informational entity e t(o).We can take a more concretely componential andcircular formula, where the so-called intelligencepertains to a certain entity a, by

*(«) "Def

(«) F\ \

ôbserverperceiveV xl I— "

^conceive(a) Fperceive

be_conscious b Cl • (nYI be_consciousv / '

Qomprehend(Q ') F ^comprehend(a) '

(27)

Ci(a)

In this definition of informational inclusion, whichpertains to intelligent entity i(a), its componentshave a superscript i while a subscript expresses a

^conclude

more concise pragmatical property. In this defi-nition, the so-called operands of informing, that

")) Cobserve(a)' ^perceiveC")) ^conceiveC")'( a ) ' ^be.conscious(a)' Qomprehend(a)' an<*

ûnderstand(a) produce (adequate) results in theform of informational entities, as ^gensationC01)'"observationV01/' ^perceptionV0/' TconceptionV0-/'

'conclusionv /» 'consciousnessv / ' l comprehension v. / 'and MmeaningC0)' respectively. Arbitrary pragmat-ical components of informing can be introducedat the informational formalization of philosophi-cal texts7.

Within a long metaphysical cycle concerningintelligent entity t(a), its informing-active prag-matical components in formula 27 can be struc-tured metaphysically as follows:

(a) p

Qonceive(«)) 1= Qonc.ude(«)) (28)

(a)) NZCderstandC")) N^bejconscious Qomprehend(a))

In this cycle we kept informing I t ( a ) of intelligententity i(a) to be involved cyclically in informing ofchosen pragmatical components. The basic prag-matical metaphysical cycles in formula 28 are

1= ^ ssensation l«)) 1=

(°observe(a) N °observation(a)) 1=

(^perceive(«) F ^perceptionC«)) 1=7?» . (/vV' perceivev / '

v^conceivev / I 'conceptionV » I

*-'conceive\C1'/>

(Qonclude(«) N ^conclusionC«)) 1=

(29)

(«);consciousness.(«))

V comprehendv / I 'comprehensionv // '

^comprehendV0))

(ûnderstandl«) F /'meaning(«)) t=iinderstand (a)

and a marks something, for instance, a word, sen-tence, text paragraph, picture, etc. (in short, aninformational entity), which will be in the processof understanding within intelligent entity t(a).

TSuch an attempt was made at informational formaliza-tion [Železnikar 92d] of f 31 (Being-there as Understand-ing) in [Heidegger 62].


As in the discussion pertaining to general in-forming of an entity, various short, medium-sized,and long metaphysical cycles for intelligent en-tity L and its components can occur, making theintelligent structure as complex, perplexed, in-terwoven, cycled, parallel, spontaneous, etc. aspossible. Additional components of intelligentinforming can be considered in an intelligentinformational game, for instance, intention^01)'^significanceC")' CTsense(»), etc. E.g., while inten-tion may appear already on lower levels of anunderstanding process, significance can become astrategic role on a higher level of a recognizingprocess, etc. Also, the object of understanding abecomes meaningly more and more information-ally identified. Different parallel metaphysical cy-cles inform the intelligent entity i as a whole aswell as its components. The reader can imaginehow extremely complex schemes and scenarios,that is, informational formulas of understandingwill come into existence.

An essential question is how one can choose in-formational components, which interact in an un-derstanding process. That what is known andcomes into the consciousness about understand-ing of something, concerns certainly the sensing ofsomething. But, sensing of something, by which asensation of something comes into existence, is alower (or the lowest) function in an observing sys-tem. This system generates the observing infor-mation as a consequence of the observing analysisand synthesis, which certainly have some percep-tional and conceptional characteristics. We seehow the main metaphysical cycle of understand-ing begins to appear via sensing, producing sen-sation into observing, producing an observationalinformation with elements of perceiving and con-ceiving of something. But, this is only the be-ginning of a cycle, which becomes raore and morestructured and complex in a componential way.Information of perception and conception is theresult so far. After this initial situation, higherinformational functions of understanding can en-ter into the informational game of understanding.Concluding is a highly informationally integrativeentity, which takes the arisen and cyclically struc-tured components of sensing, observing, perceiv-ing and conceiving, and produces a sort of the firstapproximation of that, which we can call a conclu-sion about something. But, conclusion pertaining

to the concluded of something is in no way the fi-nal result. First, it can be cycled informationallywith the intention to obtain a more sophisticatedinformation about something and, second, it canmediate the concluding results to hierarchicallyhigher positioned entities as, for example, com-prehending and understanding are. The processof comprehending has, for example, the functionof an informational comprehension of an integralform of conclusion within the being-conscious.

Informational comprehending is an action ofinformational grasping, seizing, comprising, andincluding as a consequence of the previous in-formational consciousness. Everything, which inthe cyclic process of understanding was producedin informational ways till this situation, has tobe grasped anew and included into the consid-eration of comprehending. In this way, compre-hending functions as an overtaking of somethingand coming up with the overtaken. This processof grasping reminds on sensing on a higher leveland, certainly, conceiving. The result of compre-hending is a comprehensional information, whichnow waits to be understood in a new way, whenthe action proceeds into new metaphysical cycle(middle-sized or long one) for the informationalrefinement of that, which vvas sensed, observed,perceived, conceived, concluded, conscious, com-prehended, and understood up to this situationby an intelligent informational entity and its in-forming.

Understanding something is a function of ap-prehending the meaning of something, that is,grasping the idea, information, concept of some-thing. On this level of informing, understandingis thoroughly acquainted or familiar with some-thing, so it can deal with something properlywhen producing the meaning pertaining to some-thing. When the result of this acquaintance isnot satisfactory or not final (informationally stillrelevant), a further cycling of the understand-ing process is going on, producing a refined ormore sophisticated meaning of something. Usu-ally, understanding of something can never reachthe point of being final or satisfactory, becausethe arisen meaning is a structure of an informa-tional (linguistic, semantic, tautological) net withvarious unexplored possibilities. On this way ofinforming, especially through its cycling, under-standing as an informationally acting entity can


proceed to higher informational levels of knowl-edge, which concerns something. As the highestand all-embracing component within the meta-physical cycle, understanding is in the posses-sion of faculties of lower metaphysical componentsthat concern something. In this sense, togetherwith participating metaphysical components, un-derstanding masters the informing (communica-tion, language, information), which expresses andimpresses the meaning of something.

As an additional matter, metaphysical cyclespertaining to understanding are, in their nature,cyclic-parallel.

After this discussion, we can introduce amedium-sized and pragmatically conceptualizedparallel-serial metaphysical cycle, which consid-ers all the mentioned components and delivers,when analyzed, a set of evident serial cycles inparallel. At this occasion still modes of inform-i n g 2"Žrtend(o)» «5rignify(a)' a n d ^makeênseC«) a r e

introduced, which are self-explanatory. Thus, in-stead of formula 28 we have a parallel-serial meta-physical scheme for intelligent entity v.

((((««) |= Xt(a)) |=

V1 • (perceiveV

Cl • (^conceivevconcludeV01/'

C- (a\-bejconsciousV / '

^comprehend(a)'

\ ûnderstand(a) .

(30)

\ \°observationVa/'

""perceptionv^Ji-v4 • (cx\-'conceptionv / '^condusion V. / '

I consciousness V " ) >

' comprehension v / '^meaning(.a)

\=t(a)

The last informational formula is a shortcut fora parallel system of serial formulas of all possibleforms, that is,

i-întend^signifjr^makelsense/'

{^sense? '-'observe'

^conceive

V1

' perceive'(31)

condude' be_conscious'I i A r*

comprehend' understandJ'

\ sensation' observation' perception'

'conception' 'condusion' 'conscionsness'

'comprehension' r*m

Formula 30 shows how a serial-parallel expressioncan be formally presented by a system of cycles,in which all possible informational permutationscome into the foreground. We see how specific in-formational entities, belonging to specific entitiesof informing, can become informationally influ-enced not only by informing entities, but also byentities themselves and vice versa. In this way,each entity can informationally impact and canbe informationally impacted by an occurring en-tity.

With formula 30, we can suggest a long meta-physical cycle, considering all components of in-telligent eritity i{ot), which appear in the formula.We can "properly" permute the positions of com-ponents, e.g., vvithin four parallel blocks in for-mula 30. The length of the long metaphysical cy-cles pertaining to i(a) is the number of successiveinformational operators in a long cycle, that is,^metaM«)) = 22. However, 22 is not a final value,because some of the appearing components canbe additionally decomposed (for example, Tv(a)).

Let us show only one of the long metaphysicalcycles of intelligent entity i(a), within which weconsider the informational pairs, as follows: in-tending and intention; signifying and significance;making sense and sense; sensing and sensation;observing and observation; perceiving and per-ception; conceiving and conception; concludingand conclusion; being-conscious and conscious-ness; comprehending and comprehension; and un-derstanding and meaning concerning something(that is, entity a). Thus, one of the examples(possibilities) of long metaphysical cycles is the


following one:

U°% \=a)) \= mtentionv // i

^signifjrv^// F ^significancev // i

make_sensev // i senseV // I

ênse(«)) N ^sensationt")) Nro*. r«A) b r>*. . f ^ 1=— oDservev /./ ' ~ ODScrv&iionA - ' / / r~

êrceive(«)) N ^ e r c e p t i o » ) \= (32)

^conceiveV^J/ r1 7ConceptioiAQ:.'/ PT' aT b= V • fai^ 1=

condudeV )) I 'conclusionv // ICl (n~S\ \= t1 • C«Yl b

be-consciousv // I 'consciousnessv // I

comprehendV // I 'comprehensionv " I

W.t_J....._Jfa^h/4«uunB(«))l=The reader can imagine how many other longmetaphysical cycles concerning intelligent entityi(a) are possible and how each of them representsan alternative case of metaphysical informing. Inthis variety of syntactic possibilities of long meta-physical formulas, which can inform in parallel (si-multaneously, cooperatively), participating com-ponents can arise in various informational ways,constituting the spontaneity and cyclicity of anintelligent informational entity.

9 Conclusion

Metaphysicalism is a concept of inner informing ofentities. By metaphysicalism, the informationalarising of entities in scopes of their informationalcontents is performed in an informational con-structive way. Metaphysicalism is a basic prin-ciple (called informatio tertia) and is an essentialcircular-spontaneous property of an informing en-tity. Entities within an informational system (e.g.informational machine) can obtain metaphysicalsupport by the system, but can also have theirown metaphysical "mechanisms". In this sense,metaphysicalism is a constructive approach in aconceptual and a machine-oriented way.

Metaphysicalism may not be paralleled bymetaphysics as philosophia prima or supernaturalpower. The term expresses that, which concernsthe informational emerging of an entity's possi-bilities and lies in different presentations (in Ger-man, Vorstelhmgen) of one and the same thing atdifferent observing places with difFerent informa-tional possibilities. Metaphysicalism is an innercreative power of an informing entity. Under such

circumstances, it is never completely foreseeablein advance8 for it can be impacted interiorly andexteriorly in respect to the informing entity.

Within the conclusion pertaining to metaphys-icalism, we have to say which are the possibil-ities of its technological implementation. Infor-mationally supported computing system is thefirst step to such implementation. Such a sys-tern must deliver a basic informational supportto informing entities and, within its operatingsystem, there must be informational dictionar-ies [Dictionary 90a] and, for instance, knowledgearchives [Knowledge 92], in which informationalentities needed for informational arising of an in-forming entity can be searched.

The speed of an informational machine im-plementation and machine's functional (informa-tional) performance will dramatically depend onthe sophistication of machine metaphysicalism,that is, on basic informing, countermforming andinformational embedding mechanisms by whichinformational machine as an informing entity byitself will systemically support the informing ofoccurring informational entities (operands andformulas, informational programs, informationalbases) [Železnikar 92c].

In some former essays [Železnikar 92d] it wasshown how semantically reasonably structuredwritten texts, that is, words, word groups, idioms,sentences, paragraphs, sections, etc. inform andare informed in various circular and parallel ways,which are mutually interwoven. It became evidentthat an informational interpretation (understand-ing) of a text surpasses the conventional, humanstyle of linguistic comprehension, which is on aglobal level serial, atomistically structured (con-sciously particularized), also non-parallel, andnot dynamically structured in the way of a systemof text and its parts interpreting informational

To foresee in advance may not be equaled with to pre-dict. By informational terms, we can put, for example,•* foreseeJn-advance ^ ((^see fînâdvance) f^Ln-advance) whue,for the other case, there is, Tpredict — (<5say (=before). Afurther difference is in the semantical nature of both casesand concerns, in the first case to see before in advance or,also, to see in advance, in advance and, in the second case,to say in advance. As we understand, the seeing and sayingmight be completely different informational phenomena.In the common speech, it may be inappropriate to say toforesee in advance, in advance. Within the informationaldiscourse, this case can become a matter of informationalexternalism and internalism.


formulas. Of course, besides unforeseeable prag-matical approaches of a text recognition, an infor-mational system (machine) of informing entities(operands and formulas) can consider various tra-ditional and scientifically organized methods andstructures of text interpretation, cognition, infor-mational processing, etc. But, all that may notsuffice for a dynamically understood written text,which informs highly parallel in an openly struc-tured informational realm, that is, in the world,where information and informational understand-ing arise at every time.

Thus, let us close the discourse on metaphys-icalism of informing with the following rumina-tion. Perceiving of the physical is metaphys-ical. Components as sensing, observing, per-ceiving, conceiving, concluding, being-conscious,comprehending and, at the end, understanding9

are characteristic metaphysical informational en-tities. But, metaphysicalism does not mean thatthese components do not possess their own phys-ical, biological, chemical, genetic, neuronal, so-cial, etc. backgrounds of matter, energy, infor-mation (structure, organization), which enablethe appearance of "metaphysical" phenomenal-ism. Or, said in another way ([Husserl 00] II/2,p. 244): Und sie existieren dabei keineswegs blofiphdnomenal und intentional (als erscheinendeund blofl vermeinte Inhalte), sondern voirklich.

References

[Derrida67] J.Derrida: La Voix et le Phenomene, PressesUniversitaires de France, Paris, 1967.

[Heidegger 62] M. Heidegger: Being and Time,Harper k Row, New York, 1962.

[Husserl 00] E. Husserl: Logische Untersuchun-gen, Max Niemeyer Verlag, Tiibingen, 1900-1901.

[Železnikar 90a] A.P. Železnikar: Understandingas Information, Informatica, Vol. 14 (1990),No.3, pp. 8-30.

[Železnikar 90b] A.P. Železnikar: Understandingas Information II, Informatica, Vol. 14 (1990),No.4, pp. 5-30.

[Železnikar 92a] A.P. Železnikar: Tovrards an In-formational Language, Cybernetica, Vol. 35(1992), No. 2, pp. 139-158.

[Železnikar 92b] A.P. Železnikar: Basic Informa-tional Axioms, Informatica, Vol. 16 (1992),No. 3, pp. 1-16.

[Železnikar 92c] A.P. Železnikar: An Introduc-tion to Informational Machine, Informatica,Vol. 16 (1992), No. 4, pp. 8-29.

[Železnikar 92d] A.P. Železnikar: An Informa-tional Approach of Being-there as Under-standing (in three parts), Informatica, Vol.16 (1992), No.l, pp. 9-26; No.2, pp. 29-58;and No. 3, pp. 64-75.

[Dictionary 90a] An Overview of the EDR Elec-tronic Dictionaries, TR-024, Japan Elec-tronic Dictionary Research Institute, Tokyo,April, 1990.

[Dictionary 90b] English Word Dictionary, TR-026, Japan Electronic Dictionary ResearchInstitute, Tokyo, April, 1990.

[Knowledge 92] A Plan for the KnowledgeArchives Project, The Economic ResearchInstitute, Japan Society for the Promotionof Machine Industry, and Systems Research& Development Institute of Japan, Tokyo,March, 1992.

An initial philosophy of understanding as an informa-tional entity and a first attempt to its infoimational for-malization is presented in [Železnikar 90a] (Part One: AGeneral Philosophy and Theory of Understanding as In-formation) and in [Železnikar 90b] (Part Two: A FormalTheory of Understanding as Information).

Informatica 17 (1993) 81

MISSION AND RESEARCH REPORTS

1 Introduction

The two project frameworks—one the Japaneseconcerning a plan for the Knoivledge ArchivesProject and the other pertaining to the researchprogram of the Center for the Study of Lan-guage and Information at Stanford University—have much in common. Both concern knowledgeof language and information in an extended viewof understanding and both tend towards new,the so-called informational theories, methodolo-gies, programs and architectures. The carefulreader will observe the crucial perplexity of bothundertakings—the first one in the form of a globalproject of knowledge understanding and technol-ogy, and the second one as a set of theoreti-cally, methodologically, and experimentally ori-ented projects concerning knowledge in linguisticand informational sense.

In columns entitled Mission and Research Re-ports we shall present the most significant re-search and technology projects running in theworld at present time. In this issue of Informat-ica the first parts of both project frameworks arepresented. The second parts will be published assequels in the next issue of Informatica. Thus, thereader can compare the new research and technol-ogy trends within both frameworks leading intothe realm of informational, where the informa-tional becomes an acting, meaning (semantic),and intelligent environment.

2 A Plan for the KnowledgeArchives Project I

2.1 Introduction

Again, as in the case of the Fifth Generation Com-puter Systems Project in 1980's, the Japanese re-search and development initiative has surprisedthe professional research and technology commu-nities over the globe by its grandiose plan tounite (accumulate, standardize, collect, define,aggregate, transform, arise, etc.) various kindsof knowledge (languages, sciences, technologiesor, lastly, cultures of the world, in general) ina powerful, eflfective, dynamic, emerging, signif-icant, intentional, e tc , that is, intelligently oper-

ating knowledge informational system. Three em-inent Japanese institutes have launched this planthrough a special publication1 and, as it seems,by an advanced organizational arrangement.

The plan sets some new standards in the ter-minology (definition) of knowledge related in-formational entities. Knowledge archives is avery large-scale knowledge base considering anautomated acquisition and collection of knowl-edge, stored systematically; supporting the cre-ation of new knowledge; and translating andtransmitting knowledge. These technologies willimpact knowledge itself, that is, its determina-tion, dynamic structure, and intelligent organi-zation. The knowledge archives shifts the per-spective of knowledge processing, which was con-cerned with classic development of computers,programming languages, and high level technolo-gies. The Knowledge Archives is not only amechanism for massive information and electroniclibrary. Some relevant domains of knovvledgearchives are natural language processing, knowl-edge engineering technology, multimedia, nextgeneration databases (e.g. deductive and ob-ject oriented), and software engineering, by whichknovvledge will be developed by itself. The lastview—developing knowledge by knowledge—is acharacteristic shift to the new informational wavethat can be marked as informational. This shift-ing into the informational will require new lin-guistic verbal and formal styles of expression andnew ways of dynamic semantics, meaning, inten-tion, and significance—to follow new paradigmsof informing, counterinforming, and embedding ofknowledge.

Computers as tools have to be re-evaluatedfrom the standpoint of the user through the newapproaches to technology, humanities, and socialsciences in understanding of inforraation, knovvl-edge, and their mechanisms. In this way, knowl-edge archives should become the most universalof all application systems, e.g. the most universalexpert system. The other view of the project isto be a common basis of knowledge for interna-

1A Pfon tor the Knovrledge Archives Project, The Eco-nomic Research Institute; Japan Society for the Promotionof Machine Industry; Systems Research & Development In-stitute of Japan, Tokyo, March, 1992, 1-78.

82 Informatica 17 (1993)

tional and interdisciplinary exchange in researchand technology communities, where extremely ef-fective and cooperative international relationshipscan be built up. This view concerns, for in-stance, subcultural techniques of each country'slanguage, which have to come together and bringmutual study and learning of different languagetechniques.

2.2 Technical Background

Within the project plan, knowledge is looked onas information carrying meaning, that is, consid-ering semantics of informational expression. To-day information processing technology focuses itsattention on the views of form and syntax (e.g."formulas" in natural and artificial or formal lan-guages). This platform is generally accepted indifferent information processing systems. But,it becomes also evident that conventional tech-niques of form and syntax can not produce usefuland efficient results. The task is to challenge thecontents and semantics of information, that is, toproceed into higher, parallel, interwoven levels ofform, syntax, and meaning pertaining to some-thing as information. In this respect, the wordknotvledge remains novel and concerns the infor-mational realm of that what is commonly (cultur-ally) recognized as the factical, true, believable,faithful, etc. Knowledge as a term is novel be-cause of possibilities of its informational emergingin every specific case, where meaning of somethingcan be extended in difFerent informational ways,in a straightforward and circular direction.

"How to handle the contents and semantics ofinformation as knowledge?" is the basic questionand, to this one, the second is "How the generatedknovvledge of something is information, which car-ries its own contents and semantics?" This pro-cess of knowledge generation on the linguistic levelcan proceed further and further and is in the lastconsequence always tautological. The AI boomlacked this point of view: it limited the range ofinformational objects and tried to search meaningin depth under certain limitations. The projectpreviews to deal with informational entities in abroader range and with the meaning in a shallov/range. Said literally: "The Knowledge Archivesis the technology which can grasp the meaning ina shallow but as wide as possible range and whichcan broaden and generalize the application area

as much as possible."By all means, two of the essential questions

are "Did AI research arrived to a dead end be-cause of having dealt with just a small amount ofknowledge or, self-critically, with toy problems?"and "Was AI only a phase of infantilism, whichended in proverbial dead lock on the way to anew informational research and technology per-spective?" The new informational research andtechnology perspective accentuates large-scale in-formation processing, vvhere both the amount ofknowledge and the capacity to process knowledgeare enlarged. E.g. the Fifth Generation Com-puter Systems Project is a representative of thelatter view. In this respect, new informationalconceptualism and technologies are required toacquire and store massive knowledge automati-cally and as efficiently as possible. New atten-tion has to be spend to, for instance, massiveparallel computing technology known as memorybased reasoning; neural network computing forlarge-scale symbol manipulation; and, last butnot least, theories and methods handling knowl-edge as a dynamic informational entity, whicharises in every case and in every moment in thesocial and physical environment of a living being.

The diverse knowledge representation media isthe next inevitable demand. Knowledge has to beunderstood by humans using computers, but inthis function, "a part" of knowledge must be un-derstood by computers, which support the func-tions of human understanding of knowledge. Inhuman culture, media representing knowledge arenatural, artificial and graphical languages, diver-sified images, sounds, and other, yet not identifiedhuman sensory and mental information. Multi-media technology is on the way to the highly andflexibly fused demands for informationally mixedand complex informational presentation.

The next question is, in which form knowledgecould be presented to human intelligence and, inthis respect, the necessity for research based onecology of knovvledge arises. Types of media andtypes of knowledge will determine the presenta-tion to humans. Knowledge is a regular infor-mational entity2, which arises, varies, diversifies,and comes in many forms and it is being gener-ated, edited, transformed, stored, retrieved, and

2A.P. Železnikar, Metaphysica]ism of Informing, Infor-matica, Vol. 17 (1993), No. 1, 65-80.


transmitted. Simple theories of the informationaland insignificant experiences will not suffice forthe phenomenalism of highly diversified knowl-edge ais informational entity. Knowledge as infor-mational phenomenon has to be clarified togetherwith proper research environment and develop-tnent of technologies, considering the ecology anddynamics of knowledge.

One of the next task is to standardize knowl-edge representation media. New types of logicand logic programming (e.g., informational logicand informational language3) have to be consid-ered. The next generation data base is necessaryfor storing the large amounts of knowledge andretrieve it on request.

In this context some Japanese national projectshave to be mentioned. The Fifth GenerationComputer Systems Project will proceed into theresearch and development of logic programming,parallel computers for knowledge processing, andapplication systems for verification. The JapanElectronic Research Institute (EDR) electronicdictionary project implements natural languageprocessing technology. At this project, a natu-ral language is the kernel language of knowledgerepresentation media and a part of the dictionaryis a large-scale knovvledge base of lexical knowl-edge. The multilanguage electronic dictionary isprovided for translation of sentences. Further,machine translation systems are under develop-ment at software houses and computer industryof Japan.

2.3 Social Background

Knowledge Archives Project is the first projectever, which tackles knowledge itself completelyand, in this way, meets a wide range of the so-cial needs. A knowledge (culture) base usableamong nations (civilizations) is certainly a desire.Human intelligence and knovvledge is a sociallyinterwoven phenomenon, which should be tack-led together with broadened informational theory,humanities, social sciences, and computer tech-nology. That could restructure the joined infor-mation industry.

Knowledge is a turbulent information and theignoring of its dynamism can lead to the collapseof social systems (e.g., communism, socialism).

3A.P. Zeleznikar, Towards an Infortnationa.1 Language,Cybernetica, Vol. 35 (1992), No. 2, 139-158.

Today, knowledge must be exchanged on a globallevel. Unraveling the mystery (revealing) of in-telligence and actualizing it as informational phe-nomenalism belongs to the most hopeful, dream-ful and exciting challenges in progressing towarda new human civilization.

Overproduction of information in the informa-tion age is becoming a tremendous burden for in-dividuals, who must decide, by help of knovvledge,what to accept informationally and what to leavewithout perception, that is, what to ignore. Newterms must be coined to characterize entities likeredundant, misused, insignificant, damaging, mis-understood, polluted, unworthy information, etc.

Current research and technology projects mustconsider the demands of an international society,long-term industrial needs, and the mode of tech-nology. Knowledge Archives Project is a projectfor projects, by which other projects (e.g. the fifthgeneration, electronic dictionary, language trans-lation, international projects, etc.) will be cov-ered and coordinated. The technology, which willmake computers more intelligent is going beyondsmall-scale projects conducted by business ven-tures. The whole industry is to be made more in-telligent and the high-quality information in noth-ing else than a higher intelligence. Knowledgemay be mass of textual material (static intelli-gence) or it may be informational entities floatingaround in minds of experts.

2.4 Functions and System Structureof the Project

Let us describe in short the knowledge internalstructure. Knowledge representation media arenatural languages, formal languages, picture lan-guages, images and sounds. Natural languages areJapanese and various foreign languages. Formallanguages are algebraic formulas, logical formulas,and programming languages. Picture languagesenable the representation of diagrams, tables, ar-chitectural design dravvings, electronic circuits di-agrams, music scores, etc. Images include staticimages, dynamic images and animations. Soundsare speech, music, and sounds in general. Theseknowledge representation media are appropriatelycombined and all media are treated equally.

"Knowledge documents" is appropriately repre-sented, observed, and objectively analyzed knowl-edge. Knowledge documents act as a processing


and understanding system for humans and com-puters. Some documents are merely data, somecan be understood syntactically, and some canbe understood semantically. By the advance oftechnology, the understanding of knowledge doc-uments will shift more toward computers. Knowl-edge documents will be normalized especially inthe domain of natural languages.

The massive amount of knowledge documentswill be provided by learning and self-organization.For instance, a knowledge object is an expressionof one aspect of a knowledge document. Objectsare mutually related, attributed, inferential, etc.Functions and structure will clarify the final formof the knowledge base by an arising improving.Basic units of knovvledge documents are stories (atype of text), paragfaphs, sentences, words. Ba-sic types of knowledge objects are surface objects(knowledge documents) and meaning of them aresemantic objects. Semantic objects are of a kind,which is a superior (semantic) object, etc. Thereare relations and rules for composition of surfaceobjects and semantic objects (relations betvveensemantic objects, between surface objects and se-mantic objects, and betvveen knowledge objects).

Mechanisms of inference include deductionand inferring of semantic objects, semantic rela-tions (equivalence, super and subinclusion, e tc) .Knovvledge base will learn by using its own in-ference mechanism and knovvledge will be self-organized. Hovvever, the knowledge base will notbe completed within the duration of the project(until year 2000).

The knowledge in the base will be determinedin the following ways:

(i) Characteristics of knowledge for various fieldswill be determined carefully.

(ii) "Summary knovvledge documents" will be at-tached to knowledge documents, defining re-lationships (e.g. synonymous, antonymous).

(iii) At creating of word knowledge documents,the expanded EDR electronic dictionarieswill be used.

(iv) For every field, the narrative (story) knovvl-edge will be included in the knowledge base.This type of knowledge is the most generalfor semantic recognition of knowledge docu-ments.

"How to chose the diverse fields of knowledge?"is another problem. Documents can be divided inthree types: general texts; knowledge representa-tion and programs; and data.

Texts in natural la.ngua.ges are narratives, news-paper articles, scientific and technical papers,patent documents, legal documents and prece-dents, and manuals.

Knowledge and programs /n formai languagesare knowledge documents in system knowl-edge representation languages, knowledge docu-ments in constraint logic programming languages,knowledge documents using expert systems shells,and knowledge documents in general-purpose pro-gramming langnages.

Data will have its own database language forextracting knowledge from massive amount ofdata, e.g. from MITI database.

The functions of Knowledge Archives are thefollovving: knowledge storage and retrieval func-tion; knowledge collection and acquisition func-tion; knowledge creation and utilization function;and knowledge translation and communicationfunction.

The basic functions of the knowledge base are:knowledge representation language and executingenvironment. A function defines knowledge ob-jects, their attributes, relations between objectsand attributes, inference rules, and rules forlearn-ing. This function stores the defined relations,rules, etc. and responds efRciently to demands.An expanded mechanism of deductive and object-oriented database seems appropriate.

Knowledge being common to various fields willbe stored and retrieved systematically. Knowl-edge is organized hierarchically in respect tothe document structure: words, sentences, para-graphs, texts, and stories (narratives). The hier-archy is described by natural languages. Knowl-edge documents can be structured only by naturallanguages, knowledge representation languages,and programming languages. Graphics, images,and sounds will be handled as paragraphs and or-ganized as sentences and words.

Knowledge extraction is a function, which ex-tracts and stores knovvledge automatically andeffectively. The extraction means the formingof knovvledge documents in a condensed contentsform (content knowledge documents, which gen-erally correspond to index words, key words, e tc) .


Knowledge extraction on the level of words isa kind of automatic index extraction. Sentencecontent knovoledge documents correspond to sum-mary sentences and Paragraph content knovoledgedocuments correspond to summary texts or ex-cerpts or abstracts. Thus, knowledge extractionis an automatic abstraction.

Knowledge documents will be created effec-tively from undocumented knowledge and storedin knowledge bases. Only knowledge required willbe selected from the enormous amount of knowl-edge.

A function for the translation of knowledgedocuments for translation into various languageswill be available. Existing translation methodswill be improved and expanded. Knowledge willbe communicated among knovvledge archives lo-cated in different countries. Standardization andother forms of agreement will be important, forinstance, standardization of protocols, unificationof specifications, etc. The protocol language willbe one of the representation media, so that spec-ification of various ontologies will be unified.

Knowledge archives is a system structure, is alinked system or network of four types of serversand clients. Knowledge workbench is a highly ef-ficient work station possessing the necessary func-tions for operations related to knowledge extrac-tion, creation, transformation, etc. Knowledgebase center server stores commonly usable knowl-edge in large amounts and answers retrieval in-quiries coming from different places of the globe.Knowledge base site server is located at each re-search site. It emphatically accumulates knowl-edge characteristics to each site and responds toretrieval inquiries from other sites. Knowledgecommunication server is an intelligent gateway todifferent countries, other projects, and researchinstitutes. Small-scaled servers of this type areset up at each site. The described system struc-ture of knowledge archives is a knowledge archivesnetwork.

3 Center for the Study ofLanguage and Information

Based on the CSLI 1991 Annual Report

3.1 Overview and Background

CSLI is an independent laboratory at StanfordUniversity devoted to research in the emergingscience of inforcnation, computing, and cogni-tion. This new science had its origins in the late1970s as computer scientists, linguists, logicians,philosophers, psychologists, and artificial intelli-gence researchers, seeking solutions to problemsin their own disciplines, turned to one another forhelp.

The problems that brought them together wererooted in issues that crossed the traditionalboundaries among the disciplines. A shared in-terest in how agents, vvhether biological or arti-ficial, acquire, process, and convey informationforced researchers in these different fields to con-front many of the same issues concerning commu-nication, perception, action, reasoning, and rep-resentation. Many researchers saw that problemstargeted by their own discipline were linked tosolutions in the others, and, as interaction devel-oped, began to view the common issues as defin-ing a science in its own right.

In 1985, the National Science Foundation spon-sored a meeting of representatives from the allieddisciplines to investigate the nature and potentialof the new science.4 They found that in the inter-est of progress, researchers in the individual disci-plines had idealized along dimensions that madesense in light of the questions central to their re-spective fields, but that these idealizations werecomplementary when viewed across disciplines.For example, mathematical logic had traditionallyignored the resource limitations of information-processing agents, while computer science had re-stricted itself to data bases describable withinsimple fragments of first-order logic. The confer-ence participants concluded that the new sciencehad reached a critical point. For future progress,it was now necessary to replace traditional ideal-izations with the more realistic assumptions that

[To be continued]—Summarized by A.P. Železnikar

4See Report of Workshop on Information and Represen-tation, ed. Barbara H. Partee, Stanley Peters, and Rich-mond Thomason. Stanford, Calif.: CSLI, 1985.


would emerge from serious interdisciplinary col-laboration.

CSLI's foremost goal is to provide an interdis-ciplinary setting for research in this new science.AU of the research projects currently under waybenefit from significant interaction across disci-plines. Many are collaborative projects involvingsenior researchers from different fields. The inter-action has born fruit in both expected and unex-pected ways. For example, research on speech-act theory, originally the province of philoso-phers and linguists studying human communica-tion in natural language, has influenced severalCSLI projects whose goal is the design of bettercomputer languages, architectures, and protocols.Conversely, attention to computational efficiencyand tractability has afFected the basic structure ofsyntactic and semantic theories pursued at CSLI,as well as the models of human rationality, com-munication, and learning under development invarious of the projects.

Besides being interdisciplinary, CSLI is an in-terinstitutional laboratory. Founded in 1983 byresearchers from Stanford, SRI International, andXerox PARC, CSLI has since its inception pro-moted collaboration between industrial laborato-ries and academic departments. In recent years,this collaboration has expanded to include re-searchers from additional universities, laborato-ries, and companies, both within the immedi-ate geographical vicinity and around the vvorld.CSLFs Industrial Affiliates Program currentlyincludes fourteen corporate members, many ofwhich send researchers to participate in researchprojects on site.

This interinstitutional collaboration has hadan equally important efFect on the nature andprogress of many CSLI projects. It has informedthe more theoretical projects with an awarenessof current technology and potentjal applications,while providing the more applied projects withaccess to the latest theoretical advances.

CSLI researchers are pursuing a wide vari-ety of topics, including robotics design, plan-ning and reasoning, speech recognition, machine-aided translation, language acquisition, text un-derstanding, computer languages, and softwaredesign strategies, among others. Each project fo-cuses on one or more aspects of the use of infor-mation by natural and artificial agents. Roughly

half deal with languages, vehicles by which infor-mation is communicated between agents. Thesein turn divide into those concerned with nat-ural (human) languages, and those concernedwith computer languages. The other half dealwith a variety of questions involving the ac-quisition and manipulation of information: howagents acquire and use information to guide ac-tion; what information-processing architecturesare best suited to various tasks; how representa-tional format affects information processing andhuman comprehension; and so forth.

The Annual Report is intended as a record ofthe researchers and projects a$sociated with CSLIduring 1991, as well as the seminars and collo-quia held during the year. In presenting reportsfrom the individual projects, we divide them intothe following rough groupings, according to theprojecfs principal concerns:

- Intelligent Agents

- Human/Computer Interaction

- Computer Languages and Architectures

- Natural Language

- Foundations

Readers who would like more detailed informationabout specific research projects should consult thelist of publications and references for that project,or contact the publications department at CSLI.

3.2 Intelligent Agents

Information gets its value as a guide to intelligentaction. Whether we are deciding when and whereto market a new product, or when and where toinvest our money, or simply when and where tocross the street, information is the crucial ingre-dient in our deliberations. It is what allows us tonavigate through a dynamic and changing envi-ronment.

Intelligent agents, biological or artificial, mustbe equipped to gather information about their en-vironment and to bring that information to bearon their behavior. Understanding how humans dothis, and how to build artificial agents capable ofdoing it as well, is no simple task. The projectsdescribed in this section are devoted to variousaspects of this problem. Their goals range from


the highly theoretical to the practical, from de-veloping general models of rational agency to in-vestigating techniques for improving the real-timeperformance of autonomous robots.

3.2.1 Autonomous Agents5

Goals. The goal of the Autonomous Agentsproject is to build agents, such as robots, capa-ble of learned and planned behavior in dynamicenvironments.

Significance. The control of real-time, em-bedded systems has presented severe challengesto computer scientists, yet control theorists havelong been able to use continuous feedback as atechnique for reactive control of processes. Themajor significance of our work to date is thatit imports some powerfiil control-theoretic ideasinto the core of computer science—permittingprograms whose control structure is radically dif-ferent from that of conventional programs.

3.2.2 Integrating Perception andReasoning6

Goals. We are interested in the role that reason-ing plays in perception. More specifically, we areconcerned with the boundary between low-levelinterpretation of sensory input and higher-levelcognitive processes such as planning and beliefderivation. It is obvious that sensory interpre-tation is useful as input to high-level cognition;it is less clear how and to what extent informa-tion flows in the other direction. This is the areawe are exploring, using computational models ofperception and cognition.

Signiflcance. Artificial intelligence (AI) hasdeveloped powerful computational models of bothsensory interpretation and cognitive skills such as

4Project participants: Nils J. Nilsson, Stanford Com-puter Science (project leader); Andrew Kosoresovv, Stan-ford Computer Science student; plus occasional undergrad-uates. The project is funded primarily by NASA undergrant NCC2-494. The project monitor is Monte Zvveben ofNASA-Ames.

5Project participants: Kurt Konolige, SRI International(project leader); Karen Myers, SRI International and Stan-ford Computer Science postdoctoral fellow; Daniela Musto,Italian National Research Council (CNR); Chad Walters,Stanford Symbolic Systems student. This project wasfunded by the Office oi Naval Research under contractN00014-89-C-0095 and by CSLI internal research funds.

planning. There has been relatively little inter-action between these areas, although recently theimportance of inferential and planning capabili-ties in visual interpretation has been recognizedand given the name "active vision." Starting fromthe point of view that perception is a form of infer-ence, and further that explicit symbolic reasoningis an integral part of perception, we are tryingto integrate representation and theorem-provingtechniques into the perceptual process. The ap-plication areas include abstract reasoning aboutperception, map-making for mobile robots, andperception-based text editors.

3.2.3 Rational Agency (RATAG)7

Goals. The aim of the Rational Agency(RATAG) project is to provide models and theo-ries of situated, resource-bounded, intelligent ac-tion. The scope of the research covers rational ac-tivities of human beings, including the use of lan-guage, and also the activities of artificial agentswith a variety of rational and deliberative abili-ties.

Significance. The models developed byRATAG should facilitate the development of ar-tificial agents capable of using information toguide their action in real-life environments, as wellas enhance our understanding of how biologicalagcnts bring information to bear on action.

6Project participants: John Perry, Stanford Philosophy(project leader); Michael Bratman, Stanford Philosophy;Philip Cohen, SRI International and Stanford Linguisticsconsulting ptofessor; Guven Guzeldere, Stanford SymbolicSystems student; Felix Ingrand, SRI International; DavidIsrael, SRI International and Stanford Philosophy consult-ing professor; Kurt Konolige, SRI International; HectorLevesque, Toronto Computer Science; Betsy Macken, CSLI;Karen Myers, SRI International and Stanford ComputerScience postdoctoral fellovv; Robert C. Moore, SRI Inter-national and Stanford Computer Science consulting pro-fessor; Eunok Paek, Stanford Computer Science student;John Perry, Stanford Philosophy; Martha Pollack, SRI In-ternational; Yoav Shoham, Stanford Computer Science;Syun Tutiya, Chiba Philosophy; Len Wesley, SRI Interna-tional. The project was funded by a grant from the SystemDevelopment Foundation.

Project participants: Grigori Mints, Stanford Phi-losophy, Stanford Computer Science, and Institute forCybernetics, Estonian Academy of Sciences (projectleader); Tanel Tammet, Institute for Cybernetics, Esto-nian Academy of Sciences; Vadim Basylev, Kazan Univer-sity, and Institute for Mathematical Studies in the SocialSciences (IMSSS). Funding for this project was providedby the Institute for Cybernetics, by IMSSS, and by CSLI


3.2.4 Resolution Procedures forReasoning8

Goals. Many projects in artificial intelligence(AI) and computer science presuppose a logicengine capable of doing more or less sophisti-cated reasoning. This project involves research inproblem-oriented logical systems of the resolutiontype. In particular, the following research areasare to be addressed: (1) improvement and appli-cation of the existing systems for classical and in-tensional logics; (2) construction of new systerasfor linear and belief logics.

Significance. The system we design shouldbe useful in developing specialized reasoning toolsand effective tools for teaching nonclassical logics.

3.2.5 Robotic Machine Learning ofNatural Language9

Goals. The goal of the Robotic Machine Learn-ing of Natural Language project is to develop anatural-language learning interface for a roboticsystem that can be taught to execute, in an ap-propriate environment, commands like "Put thescrew left of the nut into the hole between thewasher and the black nut!"

Signiflcance. The system we are developingwill contribute to the machine-learning theory ofnatural language. Our approach contrasts withothers in tliat it is semantically rather than syn-tactically based.

3.2.6 Situated Automata (SA)10

Goals. The Situated Automata (SA) project

internal research funds.8Project participants: Patrick Suppes, Stanford Philos-

ophy and Institute for Mathematical Studies in the SocialSciences (IMSSS) (project leader); Michael Bottner, Max-Planck-Institute for Psycholinguistics; Lin Liang, StanfordMechanical Engineering student; Weile Zhu, Universityof Electronic Science and Technology, China, ElectricalEngineering; Michael Donio, Paris student. Funding forthe project was provided by IMSSS and by the DeutscheForschungsgemeinschaft (DFG).

9Project participants: Stanley J. Rosenschein, TeleosResearch and Stanford Computer Science consulting pro-fessor (project leader); David Chapman, Teleos Research;Neil Hunt, Teleos Research; Leslie Pack Kaelbling, Stan-ford Computer Science student and Teleos Research; H.Keith Nishihara, Teleos Research; Lambert Wixson, Uni-versity of Rochester student; Nathan Wilson, Teleos Re-search. Funding for this project was provided by NASAand DARPA.

is engaged in a long-term program of researchaimed at developing theoretical foundations anddesign methods for sophisticated embedded com-puter systems. A significant portion of the workis devoted to real-time machine perception andto the integration of results in working systemsto control physical systems such as robots.

Significance. New theories being developed inthe SA project promise dramatic improvements inreal-time robotics applications, since much of thecomputational cost of current approaches can beshown to be eliminable in principle. For exam-ple, all costs involved in representing and deriv-ing consequences of invariant facts can be elimi-nated by exploiting the embedding circumstances.The project has also developed practical symboliclanguages and other tools to aid in the develop-ment of high-performance, parallel artificial in-telligence (AI) systems in the situated-automataframework.

3.3 Human/Computer Interaction

The computer is the tool of the information age.It allows us to gather, store, and transform mas-sive amounts of information in ways unimaginedthirty years ago. But the development of thetool has outpaced our ability to put it to pro-ductive use. One bottleneck is the communica-tion channel that links the computer and its hu-man users. Often, the form in which informationis most easily managed by computers is not theform most natural, or most readily understood,by their users. To us, a picture can be worth athousand words, but to tlie computer, a line ofPascal is worth a thousand pictures. To bridgesuch gaps, we need techniques for converting in-formation into a variety of forms and modalities,and a clear understanding of both the computa-tional and psychological efFects of the conversion.

This is just one dimension of the problem of hu-man/computer interaction. Large computerizedsystems, such as those used in banking and air-traffic control, must interact with a heterogeneousgroup of organizations, workers, and other com-puters. The difficulty of tailoring such a systemto a complex work setting is an order of magni-tude greater than the design of single-user soft-ware. New tools and methods are needed to aidin the construction ofsuch systems.

The projects described in this section are con-


cerned with these and related topics. They rangein scope from projects that address conceptual is-sues in software design to a project whose goal isto introduce computerized tools into the teachingof cognitive psychology.

3.3.1 Experiments in CognitivePsychologyn

Goals. Ideally, courses in cognitive psychologyshould have a laboratory to introduce students toexperimental techniques and results, to demon-strate the classic findings instead of telling aboutthem. In practice, implementing a laboratory hasbeen expensive in terms of equipment and per-sonnel. Recent advances in microcomputers haverendered developing such a laboratory possible.The goals of this project were to develop a labora-tory to accompany a course in cognitive psychol-ogy that would be: (a) comprehensive, includinga broad range of basic and classic phenomena; (b)self-running and self-explanatory, obviating theneed for personnel; (c) provide feedback to thestudent-subjects on their own performance, thegroup's performance, and typical performance.

Signiflcance. As a whole, our package demon-strates most of the classic techniques and phe-nomena used in experimental psychology. Theprograms can be used as a demonstration labo-ratory for beginning and intermediate level stu-dents, and can be used as an experiment generatorby more advanced students. That is, the standardexperiments can be revised to create new experi-ments, and then collect data on them. The pack-age was awarded a Distinguished Software Avvardby NCRIPTAL/EDUCOM in 1990.

10Project participants: Barbara Tversky, Stanford Psy-chology (project leader); Approximately fifty undergrad-uates in Symbolics Systems and Psychology participatedin creating the first versions of the experimental mod-ules. Two Symbolic Systems interns vvorked summers, JimWhite and Paul Glanthier. Glauthier continued the work,funded by the Dean's Innovation Fund.

nProject patticipants: John Etchemendy, Stanford Phi-losophy (project leader); Gerrard Allvvein, Indiana Com-puter Science student; Jon Barwise, Indiana Philoso-phy, Mathematics, and Computer Science; Douglas Felt,Stanford Academic Information Resources; Mark Greaves,Stanford Philosophy student; Michael Lenz, Stanford Sym-bolic Systems and Computer Science student; KennethNorman, Stanford Symbolic Systems student; Sun-JooShin, Notre Dame Philosophy. This project has receivedfunding from the System Development Foundation, the Of-

3.3.2 Hyperproof12

Goals. The Hyperproof project has two goals.The first is to develop a mathematical theory ofheterogeneous reasoning—reasoning in which in-formation is presented in more than one modalityor representational form. The second is to con-struct a computer application that allows the userto reason using both propositions and diagrams,one common form of heterogeneous reasoning.

Signiflcance. We believe that the Hyperproofproject could have far-reaching significance in avariety of fields.

(1) We anticipate that the first version of theHyperproof program will be an effective toolfor teaching analytical reasoning skills. Thisis an important pedagogical goal that is notachieved by standard techniques of teachinglogic.

(2) Hyperproof II, the successor program, couldprovide the prototype for a powerful general-purpose reasoning tool, useful in a wide va-riety of problem-solving contexts such asscheduling and planning.

(3) The Hyperproof architecture should be use-ful in developing special-purpose reasoningtools, such as intelligent CAD/CAM systems,in which both graphical representations andlinguistically stated constraints must be sat-isfied.

(4) Techniques used in the Hyperproof systemshould be transferable to more general prob-lems of information management in multi-modal data bases (e.g., data bases containingpictures, graphs, and text).

(5) From a psychological perspective, the Hyper-proof inference system is suggestive of an al-ternative view of human reasoning, distinctfrom both the purely deductive view associ-ated, for example, with Jerry Fodor (RutgersPsychology), and the purely model-buildingview associated with Philip Johnson-Laird(Princeton Psychology).

fice of Academic Information Resources at Stanford, theCenter for Innovative Computer Applications at IndianaUniversity, and NATO.

12Project participants: Betsy Macken, CSLI (projectleader); Cathy Haas, Stanford Special Language Program.This project was funded by CSLI internal research funds.


3.3.3 Hyperproof and American SignLanguage13

Goals. We are investigating the potential of Hy-perproof for teaching reasoning skills to deaf col-lege students, especially those whose preferredlanguage is American Sign Language (ASL). Inparallel, we seek to provide an information-basedcharacterization of ASL inspired by research re-lated to multi-representational computing andheterogeneous reasoning.

Significance. Our characterization of ASL be-gan as a way of developing two related conjec-tures, which, if correct, would have clear implica-tions for the deaf community and for researchersinterested in heterogeneous reasoning.

The first is that Hyperproof, which teaches for-mal reasoning based on both visual and senten-tial forms of information, holds special potentialfor teaching reasoning skills to deaf students andfor allowing us to make explicit reasoning skillsthey already possess but are not currently rec-ognized in traditional logic courses. The secondis the complement that theories of heterogeneousreasoning will benefit from an understanding ofthe crudal features of ASL.

In addition, our characterization vrould haveimplications for the teaching of ASL to Englishspeakers and the teaching of English to thosewhose first language is ASL. It also may be ofinterest to psychologists interested in spatial per-ception and spatial mental models.

3.3.4 Language Shapes14

Goals. Traditionally a very strong distinctionhas been drawn between texts and pictures. Ap-plied to computer representations, however, thisdistinction is less clear. We are exploring thisdistinction on a number of levels, from ways oftextually describing shape, to the developmentof languages with shaped terms. The long-rangeaim is to develop a theory of the semantics ofshapes—how shapes can carry meaning, both inthemselves and by virtue of their relationshipsto others—and to understand how to apply it toknowledge representation in artificial intelligence

(AI).Signiflcance. Various aspects of this work

have practical significance. The shape-descriptionwork with Leyton has direct application to medi-cal image perception, and is part of a developingfield of automatic visual perception mechanismswith many obvious applications.

More generally, however, the analysis of thesemantic foundations of shape has implicationsfor our understanding of cognition. For example,there has been an ongoing debate in psychologybetvveen those who believe that mental images aresomehow pictorial in nature and those who wishto map them into something essentially propo-sitional. This work suggests that both may beright, or at any rate the differences between themmay be less significant than has been supposed.

3.3.5 People, Computers, and Design(PCD)15

Goals. The goal of our research is to develop thetheories of communication and interaction thatunderlie the design of computer systems for co-operative work. Our focus is on developing thetheoretical and practical background needed toincorporate human contextual elements into thedesign and analysis of computer systems. Thetheoretical emphasis is on the development of a"language/action perspective" in which currentand potential software and hardware devices areanalyzed and designed in the context of their em-bedding in work and communicative structures.The language/action perspective grew out of ear-lier work in artificial intelligence, but it shifts thefocus of attention away from the mental and theindividual, to the social activity by which peo-ple generate the space of cooperative actions inwhich they work—and to the technology that isthe medium for those actions.

13Project participants: Patrick J. Hayes, Xerox PARCand Stanford Computer Science consulting professor(project leader); Michael Leyton, Rutgers Psychology.This project was funded by CSLI internal research funds.

14Project participants: Terry Winograd, Stanford Com-puter Science (project leader); Eric Babinet, StanfordComputet Science student; Clarisse Sieckenius de Souza,Pontificia Universidade Catolica, Brazil, Infoimation Sci-ences; Rafael Furst, Stanford Computer Science student;Brad Hartfield, Stanford Computer Science; Annie Kreyen-berg, Stanford Computer Science student; Rafael Pardo,University of Madrid; Alice Wu, Stanford Computer Sci-ence student; Mountaz Zizi, Paris VI Computer Sciencestudent. Funding Sources: NSF Grant CDA9018898, Di-rectorate of Computer and Infoimation Science and Engi-neering, and CSLI internal research funds.


Significance. The project is developing the-oretical models and methodologies that will becentral to the design of computer systems thatinclude interactions with people of any kind. ThePeople, Computers, and Design (PCD) project isalso funded by an NSF grant to develop a seriesof innovative courses on human/computer inter-action.

3.3.6 Spatial Mental Models16

Goals. The goals of the Spatial Mental Modelsproject are to characterize people's mental repre-sentations of space for a variety of spatial layouts,and to study how they are acquired and how in-formation is accessed from them. We are inter-ested in situations acquired solely from descrip-tion as well as those acquired from visual interac-tion. A secondary goal is to characterize linguisticdescriptions of space.

Significance. From this work, we will learnhow people forin mental representations of spacefrom language, what spatial properties are re-flected in those mental representations, and howpeople use language to describe space. Thepresent results indicate that the spatial mentalrepresentations formed are not like the internal-ized perceptions proposed by most theories ofmental imagery. Nevertheless, they are spatialin the sense that they contain information aboutspace and have biases that reflect general concep-tions of the spatial world. Spatial knowledge isparticularly accessible to investigation because allhumans have it and communicate about it. Men-tal representations of, and communication about,space can serve as a model for representation andcommunication about more abstract domains ofknowledge.

"Project participants: Baibara Tversky, Stanford Psy-chology (project leader); David Bryant, Northeastern Psy-chology; Nancy Franklin, New York Stony BrookPsjPchology; Caren Jones, Stanford Psychology student;Joe Kim, Stanford Psychology student; Jay Leahy, Stan-ford Psychology student; Scott Mainwaring, Stanford Psy-chology student; Deborah Tatar, Stanford Psychology stu-dent; Holly Taylor, Stanford Psychology student. The re-search was funded by the Air Force Office of Scientific Re-search, Air Force Systems Command, USAF under giantAFOSR 89-0076, and by CSLI internal research funds.

16Project participants: Peter Sells, Stanford Linguistics(project leadei); Michael Calcagno, Stanford Symbolic Sys-tems student; David Whelan, Stanford Symbolic Systemsstudent. This project has received funding from the Dean's

3.3.7 Syntax WorkBench17

Goals. Teaching introductory syntax is primarilya matter of teaching the basic ideas of hypothe-sis construction and testing, rather than the de-tails of any particular theory that has been de-veloped within the field. Syntax WorkBench is aproject intended to put in the students' hands atool that allows them to focus on analyzing dataand testing different ideas about it, thereby im-proving this aspect of the class. Our goal is toreplace the standard "textbook" method of teach-ing with an interactive environment provided bythis program, preconstructed exercises within it,and an accompanying text.

Significance. The program represents a pre-cise implementation of one particular set of ana-lytic assumptions (transformational grammar) ina way that makes exploration of classic problemsin linguistics accessible to the student in a veryshort time. We are unaware of any comparableprogram that is publicly available.

3.3.8 Tarski's World Study (TWS)18

Goals. The Tarski's World Study (TWS) projecthas three goals: to study human interaction withthe Tarski's World first-order logic microworld, todesign a cognitive model that accounts for humanreasoning within Tarski's World, and to model themeans by which learning takes place in this envi-ronment.

Signiflcance. The project has both practi-cal and theoretical significance. Practically, it isimportant to understand processes of learning inopen-ended exploratory environments. Theoreti-cally, we need a better understanding of situatedreasoning processes and mental representationsinvolved in human/computer interactions.

3.4 Computer Languages andArchitectures

The first computers were used to compute—tocarry out numerical calculations. Because of this,most theories of computation, as well as most

Innovation Fund, matched by CSLI internal research funds.17Project participants: James G. Greeno, Stanford Ed-

ucation (project leader); Eric Cooper, Stanford Educationstudent; John Etchemendy, Stanford Philosophy; HiroshiKato, NBC Corporation. This project was supported byCSLI intemal research funds.


programming languages, were based on an im-age of the computer as a programmable calcula-tor. But since then, computers have become full-fledged information-processing devices. We findthem controlling the brakes in our cars, regulat-ing load distribution on the power grid, trackingand guiding missiles, running elevators. They stillcompute, but computation in the old sense formsonly a small kernel of what they do, an interven-ing step in the process of acquiring, manipulating,and putting to use information about the worldin which they are embedded.

Computers have become engines of informa-tion, and this change requires a new understand-ing of computation, new programming languagesand paradigms, new techniques for ensuring theadequacy of a program. The projects describedin this section are devoted to this cluster of top-ics. They are exploring a wide variety of ideas—from linear logic to speech-act theory to situationsemantics—and applying them to issues afFectingthe current and future state of computation.

3.4.1 Agent-oriented Programming(AOP)19

Goals. The Agent-oriented Programming (AOP)project aims to develop a new programmingparadigm that exploits the computational advan-tages of attributing mental state to machines,which employs ideas from speech act theory, andwhich relies on computational versions of sociallaws.

Significance. Agent-oriented programmingprovides a new approach to programming dis-tributed systems, emphasizing explicit repre-sentation of time, beliefs, and commitments,which includes speech-act-like communicativecommands. Our work included the theoretical in-vestigation of mental state, interpreter design andimplementation, and applications such as traffic

control and robotics.

18Project paiticipants: Yoav Shoham, Stanford Com-puter Science (project leader); Alvaro del Val, StanfordPhilosophy student; Nita Goyal, Stanford Computer Sci-ence student; Ron Kohavi, Stanford Computer Science stu-dent; Fangzhen Lin, Stanford Computer Science researchassociate; Eyal Mozes, Stanford Computer Science stu-dent; Anton Schvvartz, Stanford Computer Science student;Moshe Tennenholtz, Stanford Computer Science postdoc-toral fellow; Becky Thomas, Stanford Computer Sciencestudent.

3.4.2 AMALA20

Goals. The goal of the AMALA project is tofind better ways to design, understand, and rea-son about computer programs that must interactwith the world external to the program. This willlead to improved programming languages, bet-ter frameworks for specifying their semantics, andmore effective techniques for reasoning about par-ticular interactive programs.

Significance. Iri addition to its direct rele-vance to the design and theory of programrninglanguages, the AMALA project has also led to abetter appreciation of just what kinds of informa-tion facilitate understanding and practical reason-ing about programs that interact with the exter-nal world. This is expected to lead to improvedprogramming methodologies helpful to program-mers using existing imperative programming lan-guages.

Furthermore, our new view of programminglanguage semantics brings it more into line withthe notion of semantics in natural-language un-derstanding, mathematical logic, and knowledgerepresentation. This approach promises to pro-vide a more comprehensive and more realisticview of complete computational agents interact-ing with the world in accordance with the seman-tic interpretation of their components.

3.4.3 Connectionist Models of LinguisticInformationProcessing (CMLIP)21

Goals. The goal of the Connectionist Mod-els of Linguistic Information Processing (CMLIP)project is to develop connectionist models of vari-ous aspects of linguistic information processing—including various modalities of language produc-tion and comprehension/recognition of such ut-terances. These models are being built around the

19Project participants: Michael Dixon, Stanford Com-puter Science student and Xerox PARC (project leader);Michael Ashley, Indiana Computer Science student; Jimdes Rivieres, Xerox PARC. This project was supported byXerox Corporation and by a grant from the System Devel-opment Foundation.

20Project participants: David E. Rumelhart, StanfordPsychology (project leader); Ben Martin, Stanford Psy-chology student. The project was partly funded by DARPAand Ricoh Corporation.


two strengths of connectionist systems—namely,their ability to learn and their ability to solvelarge constraint-satisfaction problems with many"soft" constraints. We have been developing mod-els in the areas of morphophonology, lexical rep-resentation, syntax, acquisition of semantic rep-resentations, and some of the pragmatic aspectsof conversation and communication. This projectbegan in 1989.

Significance. The work on connectionist com-putation has a large potential range of payoffs.On the one hand, since brains are inherently par-allel computational devices, it may be possible todevelop parallel computational systems modeledon the algorithms and methods found effective inreal biological systems. Moreover, some of the al-gorithms have proven to be as good as, or betterthan, existing methods in such areas as automaticspeech recognition, recognition of cursive hand-writing and other image processing tasks. A fur-ther understanding of how to apply connectionistsystems to the general language-processing sys-tem will allow for the development of a more totalparallel language-processing system. At anotherlevel, the development of these systems shouldfeed back to biology and allow the developmentof models of actual brain systems, which will im-prove our understanding of how relatively large-scale brain systems function to produce the kindsof intelligent behavior we observe in biological sys-tems.

3.4.4 Declarative Programming22

Goals. The main goal of the Declarative Pro-gramming project is to extend declarative pro-gramming beyond its present static realizations tocover many dynamic and concurrent applicationsthat at present are considered beyond the scopeof declarative programming. In particular, theproject aims to bring concurrent programmingand object-oriented programming within the foldof declarative programming, and also to unify dif-ferent programming paradigms in a declarative

way.Significance. Extending the scope of declara-

tive programming will make a wider range of taskseasier by supporting high-level reasoning aboutthe problem to be solved, and by allowing thedirect formal expression of such reasoning in adeclarative program. An important added bene-fit is the greater ease with which parallelism canbe exploited once a program has been expressedat a high level of abstraction.

3.4.5 Embedded Computation (EC)23

Goals. The long-term goal of the EmbeddedComputation (EC) project is to develop a foun-dational theory of computation as a physical pro-cess that both interacts with and represents itsenvironment. During the course of development,intermediate results and insights are applied toa variety of projects—in computer science, to thedesign of new computer languages and flexible ar-chitectures; in linguistics, to the development ofa situated language for interaction between peo-ple and machines; and in cognitive science, to theunderstanding and development of systems of per-ception and cognition.

Significance. The theoretical significance ofthis project is twofold. In terms of computer sci-ence, it will provide a rigorous framevvork for un-derstanding interactive programs—an analysis asdeep as current theories of purely functional andnumerical computation. In terms of cognitive sci-ence, it rests on the view that computational pro-cesses constitute one variety of representationalor semantic phenomenon. Asking about the na-ture of computation, on this view, involves issuesin semantics, representation, and intentionality—

21Project participants: Jose Meseguer, SRI Interna-tional (project leader); Patrick Lincoln, Stanford Com-puter Science student and SRI International Computer Sci-ence Laboratory; Narciso Marti-Oliet, Madrid ComputerScience student and SRI International; Timothy Winkler,SRI International. Funding Sources: OfRce of Naval Re-search Contracts N00014-88-C-0618 and N00014-90-C-0210, and NSF grant CCR-8707155.

22Project participants: Brian Cantwell Smith, Xe-rox PARC and Stanford Philosophy consulting profes-sor (project leader); Kathleen Akins, Xerox PARC andIllinois Philosophy and Neuroscience Cognitive ScienceGroup; John Batali, San Diego Cognitive Science; Jim desRivieres, Xerox PARC; Michael Dixon, Xerox PARC andComputer Science student; Vinod Goel, Berkeley SpecialProgram in Cognitive Sdence student; Patrick J. Hayes,Xerox PARC and Stanford Computer Science consultingprofessor; Gregor Kiczales, Xerox PARC; John Lamping,Xerox PARC; Geoffrey Nunberg, Xerox PARC and StanfordLinguistics consulting professor; Susan Stucky, Institute forResearch on Learning; John VVoodfill, Stanford ComputerScience student; Ramin Zabih, Stanford Computer Sciencestudent. This research was funded by Xerox Corporationand by a grant from the System Development Foundation.


all problems of critical concern to the cognitivesciences as a whole.

The practical significance of this project (andof a group of collaborative projects being carriedon at Xerox PARC) stems from the developmentof a series of specific architectures. Three that arepresently under way are:

(a) AMALA, a new programming language forinteractive systems based on a reworked con-ception of semantics, reference, and state (seebelow, tenet (1), and also its independentproject report).

(b) Intrigue, a metalevel compiler for Scheme (inthe tradition of 3-lisp, CLOS, and other re-flective systems) that, by exploiting interna!semantic relationships, provides an elegantway in which a program can be tuned forhigh performance without compromising thesimplicity or modularity of its original func-tional conception. The basic insight is to usemetalevel techniques to provide orthogonalcontrol of the different aspects of a compu-tation. As a first example, Intrigue is focus-ing on performance and implementation, onestep towards the long-term goal of makingintended physical realization an integral partof program design.

(c) Pidgin, a designed language for human/com-puter interaction, based not only on thetenets of the EC project, but also on manyof the ideas of efiiciency, context-dependence,indexicality, and discourse structure that arebeing studied throughout CSLI.

3.4.6 Logic and Language forComputation24

Goals. This project is concerned with logics forreasoning about computation and with languages

23Project participants: John Mitchell, Stanford Com-puter Science (project leadei); Patrick Lincoln, StanfordComputer Science student and SRÎnternational ComputerScience Laboratory; Brian Howard, Stanford ComputerScience student; Dinesh Katiyai, Stanford Computer Sci-ence student. Funding Sources: NSF Presidential YoungInvestigator Award to Mitchell, with matching funds fromDigital Equipment Corporation, Powell Foundation, Xe-rox Corporation, a gift from the Mitsubishi Corpoiation,VVallace F. and Lucille M. Davis Faculty Scholaiship. Lin-coln was supported by an AT&T Bell Laboiatories DoctoialScholarship and SRI International intemal funding.

for describing computational processes. There aretwo complementary long-term goals. One is todevise logics for stating and proving propertiesof programs. In this area, we are primarily in-terested in simple logics with limited expressive-ness. Such logics may not provide a basis for full"program verification," but may lead to tractablemethods for preventing some- common kinds ofprogrammer error. The second goal is to developprogramming languages that more accurately ex-press distinctions of interest while hiding irrele-vant concerns. We aim to develop foundationsfor systematic analysis of current programming-language features and provide a logical basis forcurrent and future programming tools.

Signiflcance. Type theory, the general studyof type systems, provides a basis for reasoningabout programs and suggests extensions to cur-rent programming languages. Type systems incurrent use prevent simple but common forms ofprogrammer error. In addition, type systems canguarantee the correctness of certain implementa-tion decisions. In a field that has lacked a generalparadigm and standard terminology, type theoryhas had signiiicant influence.

The primary shortcomings of current type the-ory are that there is little application to such im-portant topics as computational efficiency, imper-ative programming, and concurrency. The recentdevelopment of linear logic raises the possibilitythat current type-theoretic methods could be ex-tended to algorithms for automatic complexityanalysis and new perspectives on concurrency.

A number of researchers are currently investi-gating connections between linear logic and Petrinets (a basic model of concurrent program execu-tion), logic programming and nonmonotonic logic.Our current work aims to generalize type systemsto include simple logics for reasoning about the re-source requirements of a computation, and otherimportant program charaeteristics not presentlycovered by type systems.

24Project participants: Stanley Peters, Stanford Linguis-tics (project leader); Per-Kristian Halvorsen, Xerox PARCand Stanford Linguistics consulting professor; HideyukiNakashima, Electrotechnical Laboratories (ETL) CognitiveScience Section; Hinrich Schutze, Stanford Linguistics stu-dent; Hiroyuki Suznki, Matsushita Electiic Industrial Co.,Ltd. This project was supported by funds from the SystemDevelopment Foundation.


3.4.7 PROSIT25

Goals. The PROSIT project is designingand implementing a programming/knovvledge-representation language based on situation the-ory. The language is intended to improvefacilities for computing with context-depend-

;• ent information—as is essential in natural-" language processing, problems involving cooper-

ation among multiple agents, and interpretationof sensory data.

Significance. Putting situated inference un-der the control of programmers could signifi-cantly increase the power of logic-programminglanguages, whose primitive notion is inference.The PROSIT project moves in this direction byreplacing Horn clauses with situation-theoreticconstraints and allowing concrete reference to sit-uations that respect these constraints, therebyimplementing a useful notion of situated infer-ence.

Hypothetical reasoning can then be pro-grammed using inference in a hypothesized situ-ation. Reasoning mixing private knowledge withshared or common knowledge is carried out us-ing PROSIT's facilities for treating situation hi-erarchies; efRciency is gained by virtue of inheri-tance within the hierarchy, often replacing infer-ence vvithin a single situation. An emerging styleof situated programming provides new tools forsolving the computational problems that are thegoals motivating the PROSIT project.

3.5 Natural Language

Language is the most distinctively human vehicleof communication. Using it, we exchange infor-mation in remarkably efficient bursts of sound.But while sound technology—techniques for re-producing, storing, and transmitting auditorysignals—is highly advanced, language technologyis in its infancy. A human's ability to recognizean acoustic signal as a piece of language, to parseit into its component words and phrases, and thento access its information content, is a remark-able feat. Engineering artificial systems that canaccomplish almost any step in this process hasproven a formidable task.

Language technology encompasses a broadrange of research eflforts, from speech recogni-tion and synthesis to text understanding and ma-

chine translation. Many early forays into theseareas followed a similar pattern. Limited successwas achieved by applying seat-of-the-pants tech-niques, but then a sudden barrier was encoun-tered. More often than not, such barriers arisefrom an impoverished theoretical understandingof the phenomenon in question.

Just as mechanized flight eluded inventors untilaeronautical theory was sufficiently developed, sotoo must efforts in the area of language technologybe supported by a strong theoretical base. Theprojects described in this section address some ofthe immediate problems in language technology,as well as the broader theoretical issues that un-derlie them.

3.5.1 Computational Models ofRepresentationalSignals (CMORS)26

Goals. The goal of the Computational Models ofRepresentational Signals (CMORS) project is to provide a scientific base for themachine perception of natural, spoken communi-cation and visual symbols.

Significance. The theories and computationaltechnologies developed will result in improved un-derstanding of language and its communicativefunction.

3.5.2 Dialogue as a Joint Activity27

Goals. The goal of the Dialogue as a Joint Ac-tivity project is to develop a formal theory of di-alogue that takes seriously the jointly interactivenature of the dialogue process, predicts discoursebehavior, and can serve as the basis for computa-tional dialogue partners.

Significance. Understanding the structure ofdialogue will help increase the naturalness of lin-

2sProject participants: Meg VVithgott, Xerox PARC(project leader); Francine R. Chen, Xerox PARC; RonaldM. Kaplan, Xerox PARC and Stanford Linguistics consult-ing professor; Julian Kupiec, Xerox PARC; Ramin Zabih,Stanford Computer Science student. This project wasfunded by Xerox PARC and by a grant from the SystemDevelopment Foundation.

26Project participants: Philip Cohen, SRI Inteinationaland Stanford Linguistics consulting professor (piojectleader); Hector Levesque, Toronto Computer Science.Funding for this project was piovided by ATR Intetpret-ing Telephony Reseaich Laboiatories and a giant fiom theSystem Development Corporation.


guistic interaction between humans and comput-ers, as well as improve the performance of ma-chine translation and interpretation systems.

3.5.3 Grammatical Processing28

Goals. The goal of the Grammatical Process-ing project is to discover computational principlesfor the efficient and psycholinguistically plausi-ble interpretation of grammatical formalisms thatare expressive enough to characterize natural lan-guages.

Signiflcance. Rich grammatical formalismshave been developed to permit the informationaldependencies in natural-language syntactic andsemantic relations to be expressed in an accu-rate and insightful way. The descriptive powerof these formalisms is purchased at a computa-tional cost, however, in that there are no knownalgorithms for parsing and generating grammat-ical sentences in less than worst-case exponen-tial time. This worst-case behavior is not justa theoretical possibility, since exponential explo-sion is frequently observed with the programsthat are typically written to process these for-malisms. This undermines the hypothesis thatsuch grammatical systems can play a direct role inpractical applications or plausible psycholinguis-tic models. The Grammatical Processing projectis investigating alternative methods of computingwith complex constraint-based formalisms thatmay show average-case polynomial behavior whenapplied to real sentences and real grammars. Suchalgorithms vvould resolve the apparent incompat-ibility of linguistically motivated and computa-tionally tractable grammatical descriptions.

3.5.4 Knowledge-based TextUnderstanding (TACITUS)29

Goals. The long-range goal of the Knowledge-

27Project participants: Ronald M. Kaplan, Xerox PARCand Stanford Linguistics consulting professor (projectleader); John Maxwell, Xerox PARC. This project wasfunded by Xerox PARC.

28Project participants: Jerry Hobbs, SRI International(project leader); Douglas Appelt, SRI International; JohnBear, SRI International; Megumi Kameyama, StanfordComputer Science student; Mark Stickel, SRI Interna-tional; Mabry Tyson, SRI International. This researchwas funded by the Defense Advanced Research ProjectsAgency under Office of Naval Research contract N00014-85-C-0013.

based Text Understanding (TACITUS) projectis to investigate the problem of how knowledgeis used in the interpretation of discourse. Thisvery large problem decomposes into two morevery large problems: (1) how should our common-sense knowledge of the world be encoded? and (2)how should this knowledge be manipulated by in-ference processes to interpret • discourse? In thepresent phase of the project, we are building com-putational models of text understanding, basedon abductive inference. This is done primarily byimplementing a system to analyze and extract theappropriate information from naturally generatedtexts of some complexity.

Significance. Text-understanding systemshave a substantial economic potential for helpingorganizations collect information and route thatinformation to the people who need it.

3.5.5 Language Use in InteractionalSettings30

Goals. This project is concerned with the studyof discourse as a joint activity among coordinatingparticipants.

Significance. This work contributes to threeresearch efForts. One is to develop theories ofplanning, intentions, and acting. A second is tobring in the notion of collective action: how twoor more people accomplish goals by acting in en-semble. And a third is to expand the notions ofinformation beyond the traditional concern withwords and other symbols.

3.5.6 Lexical Acquisition in LanguageDevelopment (LALD)31

Goals. The goal of this project is to account for

29Project participants: Herbert H. Clark, Stanford Psy-chology (project leader); Bridget Bly, Stanford Psychol-ogy student; Jean Fox Ttee, Stanfoid Psychology student;Elizabeth Wade, Stanford Psychology student. Fundingfor this project, which was provided by the Max-Planck-Institute for Psycholinguistics, Nijmegen, The Nether-lands, supported both Clark and Fox Tree for the academicyear.

30Project participants: Eve V. Clark, Stanford Lin-guistics (project leader); Trisha Svaib, Stanford Lin-guistics student; Ruth A. Berman, Tel Aviv Linguis-tics. This research was suppotted in part by the Max-Planck-Gesellschaft during a sabbatical year spent at theMax-Planck-Institute for Psycholinguistics, Nijmegen, TheNetherlands.


how children build up their lexicon during acqui-sition.

Significance. This issue is central to any the-ory of language acquisition since it requires anaccount of what children know about the lexi-con and word-formation, what constraints theymay or must observe as they learn new words andword-forms, and how they put their knowledge touse in talking and understanding.

3.5.7 Machine-aided Translation (XL)32

Goals. The primary goal of the Machine-aidedTranslation (XL) project is to develop a theo-retical and computational foundation for researchin computational linguistics in general, but withparticular emphasis on machine translation. Itsscope thus includes any representation and pro-cessing issues that might be relevant to systemsthat perform parsing, generation, morphologicalanalysis, and so forth. One of the projecfsspecific goals is to provide computational toolsfor investigating the Resolution Problem (see re-port of the Resolution project). Another specificgoal is to build a prototype machine-translation(MT) system that explores techniques for context-dependent negotiation.

Significance. The computational tools we aredeveloping are intended to play an important sup-

31Project participants: Martin Kay, Xerox PARC andStanford Linguistics (project leader); Michael Calcagno,Stanford Symbolic Systems student; Suk-Jin Chang, SeoulNational Linguistics; Dan Fish, Stanford Symbolic Sys-tems student; Jerry Hobbs, SRI International; MasayoIida, Stanford Linguistics student and Hevvlett-PackardLaboratories; Michio Isoda, WACOM Co., Ltd.; MegumiKameyama, Stanford Computer Science student; MakotoKanazawa, Stanford Linguistics student; Charles Lee,Stanford Linguistics student; Yo Matsumoto, Stanford Lin-guistics student; Hideo Miyoshi, Sharp Corporation; Hi-roshi Nakagawa, Yokohama National Linguistics; NaohikoNoguchi, Matsushita Electric Industrial Co., Ltd.; RyoOchitani, Fujitsu Laboratories Ltd.; Stanley Peters, Stan-ford Linguistics; Livia Polanyi, Rice Linguistics; Alan Ra-maley, Stanford Symbolic Systems student; Amnon Ribak,Tel Aviv Linguistics student; Ivan A. Sag, Stanford Lin-guistics; Hinrich Schutze, Stanford Linguistics student;Peter Sells, Stanford Linguistics; Hadar Shem-Tov, Stan-ford Linguistics student; Hidetosi Sirai, Chukyo Computerand Cognitive Sciences; Yoshihiro Ueda, ATR InterpretingTelephony Reseatch Laboratories; Chris Weyand, StanfordSymbolic Systems student; Shuichi Yatabe, Stanford Lin-guistics student. Funding for this project was providedby CSLI internal research funds and Center for East AsianStudies (CEAS) funds.

porting role for future computational linguistic re-search at CSLI, including providing new tools forinvestigating the Resolution Problem. The theo-retical and conceptual research conducted by thisproject, for example, the development of notionslike "negotiation," is likely to make an importantcontribution both to the theory of translation andto MT applications.

3.5.8 Parallel Constraint Grammar(PCG)33

Goals . The goal of the Parallel ConstraintGrammar (PCG) project is to explore the organi-zation of linguistic structure into parallel systemsof constraints—structural, functional, semantic,and prosodic.

Signiflcance. The parallel constraint gram-mar architecture has computational and theo-retical advantages over current alternative de-signs. These include advantages of order-freecomputation through the removal of informationdependencies that require ordering of computa-tional processes, extensibility through the ease ofadding structures and correspondence principlesrepresenting new domains, and conceptual claritythrough increasing understanding of interactionsof different domains that have been cleanly fac-tored.

32Project participants: Joan Bresnan, Stanford Linguis-tics and Xerox PARC (project leader); Alex Alsina, Stan-fotd Linguistics student; Jan Borota, Charles Linguistics;Miriam Butt, Stanford Linguistics student; Mary Dalrym-ple, Xerox PARC; Ki-Sun Hong, Stanford Linguistics stu-dent; Chu-Ren Huang, Institute of History and Philol-ogy, Academia Sinica, and Tsing Hua University; MichioIsoda, WACOM Co., Ltd.; Smita Joshi, Stanford Linguis-tics student; Jonni Kanerva, Indiana Linguistics; RonaldM. Kaplan, Xerox PARC and Stanford Linguistics consult-ing professor; Paul Kroeger, Stanford Linguistics student;Tibor Laczko, Kossuth English; Malillo Machobane, Na-tional University oi Lesotho Linguistics; Christopher Man-ning, Stanford Linguistics student; Yo Matsumoto, Stan-ford Linguistics student; John Maxwell, Xerox PARC; SamMchombo, Berkeley Linguistics; Lioba Moshi, Georgia Lin-guistics; Tetsuro Nishino, Tokyo Denki Information Sd-ences; Gillian Ramchand, Stanford Linguistics student;Peter Sells, Stanford Linguistics; Whitney Tabor, Stan-ford Linguistics student; Annie Zaenen, Xerox PARC andStanford Linguistics consulting professor. The project wasfunded by a grant from the System Development Founda-tion and by NSF grant BNS-8919880.

33Project participants: Paul Kiparsky, Stanford Linguis-tics (project leader); Young-Mee Yu Cho, Stanford AsianLanguages student; Jennifer Cole, Stanford Linguistics stu-


3.5.9 Phonetics and Phonology (P&P)34 natural-language processing technology.

Goals. The Phonetics and Phonology (P&P)project focuses on the representation and in-terpretation of spoken language, with the goalof characterizing more explicitly the informationtransmitted in the acoustic signal.

Significance. Since utterances are physicallylocated acoustic signals, parsed by listeners intodiscrete symbolic units that serve as the basis forhigher-level linguistic processing, a theory of themapping between signal and symbol is an essen-tial part of any theory of the relation betweenutterances and the information they convey.

3.5.10 The Resolution Project35

Goals. The goal of this research project is to de-velop a basic scientific theory of the "resolutionproblem," which may be defined as the problemof how diverse kinds of information—linguistic,contextual, and encyclopedic—are integrated inreal-time language use; the problem of how com-munication can proceed rapidly and efRciently (orat all) in light of the fact that interpretation isradically underdetermined by language.

Signiflcance. The resolution problem is per-haps the most important issue in the theory oflanguage processing and arguably the most sig-nificant obstacle facing the development of useful

dent; Eunjoo Han, Stanford Linguistics student; MichaelInman, StanfoidLinûistics student; Goh Kawai, Stanford Linguistics stu-dent and SRI International; Andias Kornai, StanfordLinguistics student; William Leben, Stanford Linguis-tics; William Poser, Stanford Linguistics; Linda Uyechi,Stanford Linguistics student. Funding for the projectwas provided by a grant from the System DevelopmentFoundation.

34Project participants: Ivan A. Sag, Stanford Linguis-tics (project leader); Herbert H. Clark, Stanford Psy-chology; Lewis Creary, Hewlett-Packard Laboratories; An-thony Davis, Stanford Linguistics student; Jane Edvvards,Berkeley Psychology; Aaron Halpern, Stanford Linguis-tics student; Koiti Hasida, Institute for New GenerationComputer Technology (ICOT); Jerry Hobbs, SRI Interna-tional; Martin Kay, Xerox PARC and Stanford Linguis-tics; David Magerman, Stanford Computer Science stu-dent; Daniel Morrow, Stanford Psychology and Schoolof Medicine; David E. Rumelhart, Stanford Psychology;Hinrich Schutze, Stanford Linguistics student; WhitneyTabor, Stanford Linguistics student; Michael Tanenhaus,Rochester Psychology. The project was funded by CSLIinternal research funds and a contract from the BoeingCorporation.

3.5.11 Spoken Language in InterpretedTelephoneDialogues36

Goals. The goal of the Spoken Language in In-terpreted Telephone Dialogues project is to an-alyze dialogue and performance characteristicsof Japanese/English telephone conversations con-ducted through an interpreter.

Significance. This study allowed us to pro-vide preliminary target requirements for auto-matic telephony interpretation systems, that is,systems in which speech input in one languageis automatically translated into speech output inanother.

3.5.12 Theory of Aitiational Frames(TAF)37

Goals. The Theory of Aitiational Frames (TAF)project is developing a word semantics, or lexi-cal semantics that cuts across syntactic theories,and endeavors to formulate a theory with tiesto formaj semantics as practiced by philosophersand generative grammar as practiced by linguists.Our aim is to provide a conceptual framework foranalyzing our understanding of nouns, verbs, andother parts of speech that can be used by work inpsycholinguistics, such as that of S. Pinker (MIT),and cognitive psychologists such as R. Case (Stan-ford CERAS).

Significance. This theory ofFers a uniformtreatment of the semantic representation of wordsthat can be used in predicate expressions. It aimsto unify insights from philosophy of language,linguistics, and psychology. It can represent a

35Project participants: Philip Cohen, SRI Internationaland Stanford Linguistics consulting professor (piojectleader); Sharon L. Oviatt, SRI International; Ann Pod-lozny, SRI International. Funding for this projectwas provided by ATR Interpreting Telephony ResearchLaboratories.

36Project paiticipants: Julius Moravcsik, Stanford Phi-losophy (project leader); Jane Aronson, Stanford Philoso-phy student; Michael Calcagno, Stanford Symbolic Sys-tems student; Paul Glauthier, Stanford Symbolic Sys-tems student; Stephen Irish, Stanford Philosophy student;Martin McDermott, Stanford Philosophy student; RayRavaglia, Stanford Philosophy student; Katherine Uum,Stanford Philosophy student. This project was funded bya grant from the System Development Foundation.


variety of semantic relations including not onlyhomonymy, synonymy, and polysemy, but alsopartial overlaps of a variety of sorts. Practicalapplications include facilitating translation fromlanguage to language, and dealing with recalci-trant semantic facts such as metaphor.

3.5.13 Verbmobil: A Translation Systemfor Face-to-Face Dialog38

Goals. This project arose as the result of a re-quest from the Bundesministerium fiir Forschungund Technologie (BMFT), the German FederalMinistry of Research and Technology. They wishto embark on a program of research to developan experimental prototype for "Verbmobil," aportable simultaneous interpreter. They askedsome of the researchers at CSLI to examine thefields of science and engineering bearing on Verb-mobil and to assess what they could contributenow and what advances of importance to the pro-gram could be expected soon. In addition, theyasked us to draw up a plan for the Verbmobil pro-gram for the period beginning now and ending inthe year 2000, saying how close it might be pos-sible to come to the goal in that time and whatsteps should be taken to assure the best outcome.

Significance. Verbmobil is an extremely am-bitious program. It depends on finding solutionsto many difficult problems. For example, Verb-mobil must be able to pick sentences out of thestream of sounds that impinge on its microphone.But the sounds are run together and confused bybackground noise. Nothing in the stream corre-sponds to the spaces that separate written lettersand words, or the punctuation marks that set offmajor phrases. There is also much that is notunderstood about the way language is used ineveryday conversation, such as interpretation ofintonation and rhythmic variation, how mistakesare corrected, and how one person yields the floor

37Project participants: Martin Kay, Stanford Linguisticsand Xerox PARC (project leader); Jared Bernstein, SRI In-ternational; John Etchemendy, Stanford Philosophy; JeanMark Gavvron, SRI International; Jerry Hobbs, SRI Inter-national; Megumi Kameyama, Stanford Computer Sciencestudent; Betsy Macken, CSLI; Peter Norvig, Sun Microsys-tems; Ryo Ochitani, Fujitsu Laboratories Ltd.; Stanley Pe-ters, Stanford Linguistics; VVilliam Poser, Stanford Lin-guistics; Ivan A. Sag, Stanford Linguistics. This projectwas funded by the BMFT contract Tnternational Trendsin Machine Translation and Speech Understanding."

to another. Above all, Verbmobil must be able totranslate what it hears into another language and,while this problem has motivated much researchin recent years, much more must be learned beforeeven the first prototype can be built.

Thus, while Verbmobil is clearly important asan end in itself, for the immediate future its im-portance will lie in the way in which it will focusattention on central scientific issues and channelenergy into key enabling technologies.

3.6 Foundations

The concept of information—like those of energyand mass in physics, species and gene in biol-ogy, inflation and deflation in economics—derivesmuch of its meaning from our everyday interac-tions with the world. But the demands placed onthese concepts by the working scientist require adegree of precision and refinement that goes be-yond that of the ordinary notion.

Many issues that arise in the study of action,communication, and computation require a pre-cise characterization of the information content ofsentences or signals or representations. The spe-cific goal may be to translate sentences from onelanguage to another, to present large amounts ofdata in a humanly comprehensible form, to searchand summarize a body of text, or to compare thecontents of differently structured data bases. Butwhatever the goal, such notions as informationcontent and informational equivalence loom largein the investigation.

The project described in this section is devotedto the abstract study of information. Its goal isto develop a mathematically precise theory of in-formation, and to apply that theory in a varietyof areas within the study of language and compu-tation.

38Project participants: John Etchemendy, Stanford Phi-losophy (project leader); Peter Aczel, Manchester Com-puter Science; Jon Barvvise, Indiana Philosophy, Math-ematics, and Computer Science; Robin Cooper, Edin-burgh Cognitive Science; Keith Devlin, Colby Mathemat-ics; Elizabet Engdahl, Edinburgh Linguistics; TimothyFernando, Stanford Symbolic Systems student and Am-sterdam Institute for Logic, Language, and Information;Jean Mark Gavvron, SRI International; David Israel, SRIIntetnational and Stanford Philosophy consulting profes-sor; Yasuhiro Katagiri, NTT; Paul J. King, CSLI postdoc-toral fellow; Kuniaki Mukai, Institute for New GenerationComputer Technology (ICOT); Hideyuki Nakashima, Elec-trotechnical Laboratories (ETL) Cognitive Science Section;


3.6.1 Situation Theory and SituationSemantics (STASS)39

Goals. The Situation Theory and Situation Se-mantics (STASS) project serves as a clearinghousefor worldwide research on situation theory. Thegoals of situation theory are to develop a unifiedmathematical theory of meaning and informationcontent, and to apply that theory to specific areaswithin the study of language, computation, andcognition.

Significance. Most work in information the-ory adopts a purely quantitative approach tothe subject, dealing only with the information-carrying capacity of a signal or information chan-nel. From this perspective, a meaningless stringof characters or bits is equivalent to a meaningfulstring, so long as the two signals diverge equallyfrom random background noise. This approachto information is valuable for certain purposes,particularly when our concern is the error-freetransmission of information by devices that arenot themselves users of information.

But quantitative information theory ignores is-sues that arjse when signals must be dealt withat the level of information content, for example,in the context of machine translation (when thecontent of sentences from different languages mustbe compared), machine perception (when usableinformation must be extracted from sensory in-put), or, potentially, multimedia (when the sameor related information is represented in differentmodalities or formats). As information technol-ogy has matured, the need for a general mathe-matical framework in which to address these is-sues has become increasingly evident. The theoryof meaning and information content being devel-oped in situation theory is meant to provide sucha framevrork.

John Nerbonne, Saarbrucken Computational Linguisticsand Deutsches Forschungsinstitut fur Kunstliche Intelli-genz (DFKI); John Perry, Stanford Philosophy; StanleyPeters, Stanford Linguistics; Gordon Plotkin, EdinburghComputer Science; Jerry Seligman, Edinburgh CognitiveScience and Indiana Logic Group; Sun-Joo Shin, NotreDame Philosophy; Hidetosi Sirai, Chukyo Computer andCognitive Sciences; Btian Cantwell Smith, Xerox PARCand Stanford Philosophy consulting professor; HiroyukiSuzuki, Matsushita Electric Industrial Co.; Syun Tutiya,Chiba Philosophy; Dag Westerstahl, Stockholm Philoso-phy; Edward N. Zalta, Stanford Philosophy acting assis-tant professor. Funding for this project was provided by agrant from the System Development Foundation.

NEWS


CACM, Vol. 36, No. 1, January 1993; New-strack:According to the General Accounting Office, 10 of the11 key industrial sectors in USA fell to competitors.President (at the time candidate) Clinton announcedhis plans for promoting USA high technology, such asrobotics, biotechnology, fiber optics, netvvorks, digitalimaging, data storage, CAD and AI.Daily News reports that the project to computerizestudent attendance records is millions over budgetwhile less than half of the intended 1,000 city pub-lic schools are wired.Benchmark tests have shown that on three selectedbenchmark problems traditional vector supercomput-ers were faster than massively parallel machines.A new magazine Future Sex is devoted to cybersex.

CACM, Vol. 36, No. 2, February 1993; New-strack:IFIP appeals to everybody worldwide to censure harm-ful games, to raise awareness of the issues involved, andto support only computer games that respect humandignity.Under President Clinton and Vice President Gore, anAl-based system Resumix helped to sort 100,000 appli-ances for 4,000 federal jobs thus saving about a millionpieces of paper.Carnegie Mellon University has created a five-year pro-gramme leading to a bachelor's degree and a master'sdegree in software engineering in coordination withfirms such as Apple, Intel, Microsoft and Motorola.

CACM, Vol. 36, No. 1, January 1993; Con-tents:Devoted to multimedia.In the Presidenfs letter, the ACM President GwenBell introduces bridges between worlds. As claimed,the ACM membership reaches approximately 85,000,however, only 20% of them are outside North Amer-ica. Similar proportions seems valid for several otherimportant special interest groups.In the ACM Forum, Gregory Aharonian notes that in1960 - 1992 over 9,000 software patents were issued,and in 1992 alone about 1,300. Robert Milner, a pro-fessor at the University of Edinburgh and a TuringAward winner, presents his work on concurrency, andhis viewpoints on several aspects of computing in anintervievv.

CACM, Vol. 36, No. 2, February 1993; Con-tents:Devoted to Digital's Alpha Chip Project (see articlein this issue of Informatica).In the Forum, debate concerns the proposed motive

by governmental institutions to broaden the computerscience by encompassing computer engineering and ap-plications. Science is not to be done just because it isscience, but it must either produce results or be atleast very promising in the future. The principal issueraised is who is to set priorities for computing research,and if it is reasonable to advocate linear (mutually sup-portive) transmissions of discoveries into applications.

ACM Membernet:The 5-hour ACM sponsored television series 'TheMachine That Changed The World' was ratedamong the top six PBS prime time productions inUSA, and received generous commends. Regular re-tail prices are about $4000 (contact P.O. Box 2053,Princeton, NJ 08543). Personally, and from commentsof colleagues, we share acclamations by American crit-ics with a small remark that it should be shipped toEurope in European VHS recording standards.

Artificial Intelligence, Vol. 59, No. 1-2,February 1993Special volume: artificial intelligence in perspective.Dedicated to the memory of Allen Newell who died onJune 19, 1992, one of the founders of AI.Based on the SCI's databases, authors of the papersin the magazine which had over 24 citations were in-vited to write a short paper about their work, area andprospectives. There were around 30 of them includ-ing Nevvell, which was able to write his work althoughprobably knovving he won't be able to read it.Indeed, we should pay him greatest respects to severalof his important achievements, especially Soar in thelast decade or so, to his great intellectual power, hisdreams and devotion.

Minds and Machines, Vol. 3, No. 1, Febru-ary 1993; Contents:Larry Hausler and "VVilliam J. Rapaport exchange ar-guments regarding the question 'is a pocket calcula-tor a thinking thing'. Hauser starts argumenting thatthere is no sufficient proof (at least without a reason-able doubt) that a calculator is not a thinking thing.It seems that he is deliberately posing the frontier ofhis arguments at the level even he does not believe in,just to show that there are no real arguments that com-puters can not be (are not?) intelligent. Although Ra-paport demolishes Hauser's claims, he seems to nearlycome into terms with the claim that an integrated pro-gram (system) could be considered as intelligent to acertain degree.


IJCAI-93 WORKSHOP

Machine learning and knowledge acquisition: com-mon issues, contrasting methods, and integrated ap-proaches.

29 August 1993, Chambery, France.

Machine learning and knovvledge acquisition share thecommon goal of acquiring and organizing the knowl-edge of a knowledge-based system. However, each fieldhas a different focus, and most research is still donein isolation from each other. The focus of knowledgeacquisition has been to improve and partially auto-mate the acquisition of knowledge from human ex-perts. In contrast, machine learning focuses on mostlyautonomous algorithms for acquiring or improving theorganization of knowledge, often in simple prototypedomains. Also, in knovvledge acquisition, the acquiredknowledge is directly validated by the expert that ex-presses it, while in machine learning, the acquiredknowledge needs an experimental validation on datasets independent of those on which learning took place.As machine learning moves to more 'real' domains, andknowledge acquisition attempts to automate more ofthe acquisition process, the two fields increasingly findthemselves investigating common issues with comple-mentary methods. Hovvever, lack of common researchmethodologies, terminology, and underlying assump-tions often hinder a close collaboration. The pur-pose of this symposium is to bring together machinelearning and knowledge acquisition researchers in or-der to facilitate cross-fertilization and collaboration,and to promote integrated approaches vvhich couldtake advantage of the complementary nature of ma-chine learning and knowledge acquisition.

Topics of interest include, but are not limited to,the following: Case Studies, Comparative Studies,Hard Problems, Knowledge Representation, Key Is-sues, Overviews, Position Papers.

It is recommended that the papers make explicit theresearch methodology, the underlying assumptions,definitions of technical terms, important future issues,and potential points of collaboration. They should notexceed 15 pages. The organizers intend to publish aselection of the accepted papers as a book or the spe-cial issue of a journal. They encourage the authorsto take this into account while preparing their papers.The format of the workshop will be paper sessions withdiscussion at the end of each session, and a concludingpanel on the integrated approaches, guidelines for suc-cessful collaboration, and concrete action items. Thenumber of the participants to the workshop is limitedto40.

Each workshop attendee must also register for theIJCAI conference and must pay an additional 300FF

(about $60) fee for the workshop. One student attend-ing the vvorkshop and being in charge of taking noteswill be exempted from the additional 300 FF fee. Vol-unteers are invited.

WORKSHOP CO-CHAIRS .

Smadar Kedar, NASA Ames & Inst. for Learning Sci-ences, ([email protected])Yves Kodratoff, CNRS & Universite de Paris-Sud,([email protected])Gheorghe Tecuci, George Mason Univ. & RomanianAcademy, ([email protected]) ,

PROGRAM COMMITTEE

Ray Bareiss, Institute for the Learning SciencesCatherine Baudin, NASA AmesGuy Boy, European Inst. of Cognitive Sc. and Eng.Brian Gaines, University of CalgaryMatjaz Gams, Jozef Stefan InstituteJean-Gabriel Ganascia, Univ. Pierre and Marie CurieNathalie Mathe, European Space Ag. and NASAAmesRyszard Michalski, George Mason UniversityRaymond Mooney, University of Texas at AustinKatharina Morik, Dortmund UniversityMark Musen, Stanford UniversityMichael Pazzani, Univ. of California at IrvineLuc De Raedt, Catholic University of LeuvenAlan Schultz, Naval Research LaboratoryMildred Shaw, University of CalgaryMaarten van Someren, University of AmsterdamWalter Van de Velde, University of Brussels

ADDRESS FOR CORRESPONDENCE

Gheorghe TecuciArtificial Intelligence Center, Computer Science De-partmentGeorge Mason University, 4400 University Dr., Fair-fax, VA 22030email: [email protected], fax: (703)993-3729

Those who would like to attend without a presen-tation should send a one to two-page description ofrelevant research interests and a list of selected publi-cations.

(Informations about events should be sent by e-mailto [email protected], possibly already in LaTeX or atleast in ASCII)


ERK'93Electrotechnical and Computer ConferenceElektrotehniška in računalniška konferenca

27.-29. September 1993

Conference ChairmanBaldomir ZajcUniversity of LjubljanaFaculty of Electr. Eng. and Comp. ScienceTržaška 25, 61000 Ljubljana, SloveniaTel: (061) 265 161, Fax: (061) 264 990E-mail: [email protected]

Conference VicechairmanBogomir HorvatUniversity of MariborTechnical Faculty,Smetanova 17, 62000 Maribor, SloveniaTel: (062) 25 461, Fax: (062) 212 013E-mail: [email protected]

Program Committee ChairmanSaša DivjakUniversity of LjubljanaFaculty of Electr. Eng. and Comp. ScienceTržaška 25, 61000 Ljubljana, SloveniaTel: (061) 265 161, Fax: (061) 264 990E-mail: [email protected]

Programe CommitteeTadej BajdZmago BrezočnikJanko DrnovšekMatjaž GamsFerdo GubinaMarko JagodičJadran LenarčičDrago MatkoMiro MilanovičAndrej NovakNikola PavešičFranjo Pernuš

Publications ChairmanFranc SolinaUniversity of LjubljanaFaculty of Electr. Eng. and Comp. ScienceTržaška 25, 61000 Ljubljana, SloveniaTel: (061) 265 161, Fax: (061) 264 990E-mail: [email protected]

Advisory BoardLudvik GyergyekDali DjonlagičKarel JezernikPeter JerebAlojz KraljMarjan PlaperJernej VirantLojze Vodovnik

Call for Papersfor the second Electrotechnical and Computer ConferenceERK'93, which will be held on 27-29 September 1993 in Portorož,Slovenia. All presentations during the first day of the conference (in-vited lectures and selected sessions) will be held in English.

The following areas will be represented at the conference:

- electronics,- telecommvnicaiions,- measurement,- automatic control and robotics,- computer and information science,- ariificial intelligen.ee and pattern recognition,- biotnedical engineering,- power engineering.

The conference is organized by the Slovenia Section IEEE andother Slovenian professional societies:

- Slovenian Society for Automatic Control,- Slovenian Measurement Society (ISEMEC 93),- SLOKO-CIGRE,- Slovenian Society for Medical and Biological Engineering,- Slovenian Society for Robotics,- Slovenian Artificial Intelligence Society,- Slovenian Pattern Recognition Society.

Authors who wish to present a paper at the conference should sendthree copies of their abstract (500 words) to the chairman of the Pro-grame Committee prof. S. Divjak. The abstract should include also:

1. the title of the paper,2. author's address,3. telephone, telefax and e-mail of the contact author,4. the paper's subject area.

Authors of accepted papers will have to prepare a four page camera-ready copy of their paper for inclusion into the proceedings of theconference.

Time schedule: Abstracts dueNotification of acceptanceCamera-ready paper

1 June 199330 June, 19931 September 1993.

For all additional information please contact the conference chairmen.


JOZEF STEFAN INSTITUTE

Jožef Stefan (1835-1893) was one of the mostprominent physicists of the 19th centurtj. Born toSlovenian parents, he obtained his Ph.D. in Vi-enna University, where he was later Director ofthe Physical Institute, Vice-President of the Vi-enna Academy of Sciences and member of severalscientific institutions in Europe. Stefan ezploredmany areas from hydrodynamics, optics, acous-tics, electricity, magnetism and the kinetic theoryof gases. Among other things, he originated thelaw that the total radiation from a black body isproportional to the J^th potver of its absolute tem-perature, known as the Stefan-Boltzmann law.

The Jožef Stefan Institute (JSI) is a researchorganisation for pure and applied research in thenatural sciences and technology. Both are closelyinterconnected in research departments composedof different task teams. Emphasis in basic re-search is given to the growth and education ofyoung scientists, while applied research and devel-opment serve for the transfer of advanced knowl-edge, contributing to the development of the na-tional economy and society in general.

At present the Institute, totalling about 800,has 500 researchers: about 250 of them arepostgraduates, over 200 have doctorates (Ph.D.),and around 150 have permanent professorships ortemporary teaching assignments at the Universi-ties.

In view of its activities and status, the JSI playsthe role of a national institute, complementingthe role of the universities and bridging the gapbetvveen science and applications.

Research at the JSI includes the following ma-jor fields: physics; chemistry; electronics, infor-matics and computer sciences; biochemistry; ecol-ogy; reactor technology; applied mathematics.Most of the activities are more or less closely con-nected to information sciences, in particular com-puter sciences, artificial intelligence, language andspeech technologies, computer-aided design, com-puter architectures, biocybernetics and robotics,computer automation and control, professionalelectronics, digital communications and networks,and applied mathematics.

The Institute is located in Ljubljana, the cap-ital of independent country Slovenia (or S^nia).The capital today is considered as a crossroad be-

tween the East, West and Mediterranean Europe,ofFering excellent productive capabilities and con-solidate business opportunities with strong in-ternational connections. Ljubljana is connectedto important centers such as Praha, Budapest,Wien, Zagreb, Milano, Roma, Monaco, Nice,Bern, Miinchen all inside the circle of 600 km.

In the last year at the location of the Jožef Ste-fan Institute, the Technology park "Ljubljana" isproposed as a part of the national strategy fortechnological development to foster synergies be-tween research and production industry, to pro-mote joint ventures between university bodies, re-search institutes and innovative industry, to act asan incubator for high-tech initiatives and to accel-erate the developing cycle of innovative products.

At the present time a part of the Institute isbeing reorganized in several high-tech units sup-ported and connected within the Technology parkat "Jožef Stefan" Institute, established as a be-ginning of an regional Technology park "Ljubl-jana". The project is being developed at a partic-ular historical moment characterized by a processof state reorganization, privatisation and privateinitiative. The national Park will take the forraof a shareholding company and will host an inde-pendent financial institution for venture kapital.

Promoters and operative entities of the projectare the Republic of Slovenia, Ministry of Scienceand Technology and the Jožef Stefan Institute.The frame of the operation includes also Univer-sity of Ljubljana, Institute of Chemistry, Insti-tute for Electronics and Vacuum Technique, In-stitute for Materials and Construction Researchand some others. Furthermore the project is sup-ported by Ministry of Small Business, NationalChamber of Commerce and the City of Ljubljana.

Jožef Stefan InstituteJamova 39, 61000 Ljubljana, SloveniaTel.:+38 61 159 199Fax.:+38 61 161 029Tlx.:31 296 JOSTIN SI

E-mail: [email protected]

Contact person for the Park: Iztok Lesjak, M.Sc.

Public relations: Ines Černe


INFORMATION FOR CONTRIBUTORS

1 Submissions and Refereeing

Please submit three copies of the manuscript withgood copies of the figures and photographs to one ofthe editors from the Editorial Board or to the Con-tact Person. At least two referees outside the au-thor's country will examine it and they are invitedto write as many remarks as possible directly on themanuscript, from typing errors to global philosophicaldisagreements. The chosen editor will send the au-thor copies with remarks, and if accepted also to theContact Person. The Executive Board will inform theauthor that the paper is accepted, meaning that it willbe published in less than one year after receiving orig-inal figures on separate sheets and the text on an IBMPC DOS floppy disk or through e-mail - both in ASCIIand the Informatica LaTeX format. Style (attached)and examples of papers can be obtained through e-mail from the Contact Person.

2 News, letters, opinions

Opinions, news, calls for conferences, calls for papers,etc. should be sent directly to the Contact Person.

\documentstyle[tBoside,informat]{article}

\begin{document}\title{TITLE 0F THE ARTICLE}\author{First Author \\ Address \\ E-mail \\

AID \\Second Author \\ Address \\ E-mail}

\titleodd{TITLE II HEAOER}\authoreven{Author's naiae in header}\keysords{article keysords}\edited{editor in charge}\received{date}\revised{date}\accepted{date}

\abstract{abstract — around 200 aords}\maketitle

\8ection{Introduction}Text of Introduction.

\sect ion{Subj ect}Text of the Subject.

\begin{figure}

An example of a figure over 1 column.\caption{Caption of the figure 1}

\end{figure»}

\begin{figure*}An example of a figure over 2 columns.\caption{Caption of the figure 2}

\end{figure*}

\begin{thebibliography}{99}\bibitem{} First Author: {\sl Title},Hagazine, Vol. 1, lo. 1.

\end{thebibliography}

\end{document}

\def\journal{Informatica Vol. 17, lfo. 1, Harch 1993}

\immediate\¥ritel6{'Informatica' VI.1 by B."Z}\newif\iftitle \titlefalse\hoffset=llmm \voffset=8nun\oddsidemargin=-21mm \evensideraargin=-14ir«n\topmargin=-33mm\headheight=17mm \headsep=10mm\footheight=8.4mn \footskip=52.5nun\textheight=242mm \text»idth=170mm\columnsep=5mm \columnseprule=Opt\tnocolumn \sloppy \flushbottom\parindent lem

Heftmargini 2em \leftmargin\leftmarginiMeftnarginv . 5em \leftmarginvi . Sem\def\labelitemi{\bf — } \def\labelitemii{—}\daf\labelitemiii{—}\setcounter{secnumdepth}{3}\def\maketitle{\twocolumn[X

\vbox{\hsize=\textwidth\Large\bf\raggedright\8t itle}\vss\bigskip\bigskip \vbox{\hsize=\textHidth \8author}\bigskip\smallskip\vbox{\hsize=\textwidth {\bf Keywords:}\8keywords}

\bigskip \hbox{{\bf Edited by:} \8edited}\smallskip\hbox{{\bf Received:}\hbox to 10om{ \8received\hss}{\bf Revised:}\hbox to 10em{ \8revised\hss}{\bf Accepted:}\hbox to 10em{ \8accepted\hss}}

\bigskip \vbox{\hsize=\textoidth\parskip=\baselineskip \leftskip=3em\rightskip=3em \sl \8abstract}

\bigskip\bigskip]\titletrue}\def\maketitleauthor{\tBOColumn[y.\vbox{\hsize=\text»idth \Large\bf\raggedright\®title}\vss \bigskip\bigskip\vbox{\hsize=\textwidth \8author}\bigskip\bigskip] \gdef\8title{}\titletrue}

\def\makeonlytitle{\tBocolumn[y,

\vbox{\hs ize=\textnidth \Large\bf\raggedright\8title}\vss\bigskip\bigskip]\gdef\8title{>\titletrue}

\def\8title{} \def\8author{}\def\8titleHO \def\8authorH{}\def\8key¥ords{} \def\8edited{} \def\8abstract{}\def\8rsceivedO \def\8revised{} \def\8accepted{}\def\authoreventl{\gdef\8authorH{tl}}\def\titleoddtl{\gdef\8titleH{*l}}\def\keyBordstl{\gdef\®keyBords{tl}}\def\edited*l{\gdef\8edited{#l}}\def\receivedtl{\gdef\8recoived{tl}}\def\revisedtl{\gdef\8revised{tl}}\def\acceptedtl{\gdef\8accepted{#l}}\long\daf\abstracttl{\gdef\8abstract{#l}}\def\section{\8startsection {section}

{I}{\z8}{-3.5ex plus -lex minus -.2ex}{2.3ex plus .2ex}{\Large\bf\raggedright}}

\def\subsection{\8start8ection{subsection}{2}{\z8H-3.25ex plus -lex minus -.2ex}{1.5ex plus .2ex}{\large\bf\raggedright}}

\def\subsubsection{\8startsection{subgubsection}{3}{\z8K-3.25ex plus -lex minns -.2ex}{1.5ex plus .2ex}{\normalsize\bf\raggadright}>

\def\8evenhead{\hbox to 3em{\bf\thepage\hss}{\small\journal\hfil \iftitle\else

\8authorH \fi}\global\titlefalse}\daf\8oddhead{{\sniall\iftitle\else \8titleH \fi\hfil\journal}\hbox to 3em{\hss\bf\thepage>\global\titlefalse}

\dof\8ovenfoot{\hfil} \def\8oddfoot{\hfil}\endinput

Infoimatica 17

INFORMATICAAN INTERNATIONAL JOURNAL OF COMPUTING AND INFORMATICS

INVITATION, COOPERATION

Since 1977 Informatica has been a major Slovenian sci-entific journal of computing and informatics, includingtelecommunications, automatics and other related ar-eas. In its 17th year it is becoming truly international,although it still remains connected to Central Europe.The basic aim of Informatica is to impose intellectualvalues (science, engineering) in a distributed organisa-tion.

Informatica is a journal primarily covering the Eu-ropean computer science and informatics community;scientific and educational as well as technical, commer-cial and industrial. Its basic aim is to enhance commu-nications between different European structures on thebasis of equal rights and international refereeing. Itpublishes scientific papers accepted by at least two ref-erees outside the author's country. In addition, it con-tains information about conferences, opinions, criticalexaminations of existing publications and news. Fi-nally, major practical achievements and innovations inthe computer and information industry are presentedthrough commercial publications as well as throughindependent evaluations.

Editing and refereeing are distributed. Each editorcan conduct the refereeing process by appointing twonew referees or referees from the Board of Refereesor Editorial Board. Referees should not be from theauthor's country. If new referees are appointed, theirnames will appear in the Refereeing Board.

Informatica is free of charge for major scientific, peda-gogical and governmental institutions. Others shouldsubscribe (institutional rate 50 DM, individual rate25 DM, and for students 10 DM). Send a check in theappropriate amount to Slovene Society Informatica,Vožarski pot 12, Ljubljana, account no. 50101-678-51841.

Please, return this questionnaire.

QUESTIONNAIRE

[ I Send Informatica free of charge

| | Yes, we subscribe

We intend to cooperate (describe):

Propositions for improvements (describe):

Informatica is published in one volume (4 issues) peryear.If possible, send one copy of Informatica to the locallibrary.Please, fulfil the order card and send to dr. RudiMurn, Informatica, Institut Jožef Stefan, Jamova 39,61111 Ljubljana, Slovenia.

ORDER CARD - INFORMATICA

Name: Office Address and Telephone (optionally):

Title and Profession (optionally):

E-mail Address (optionally):Home Address and Telephone (optionally):

Signature and Date:

Inforrnatica 17

INFORMATICAREVIJA ZA RAČUNALNIŠTVO IN INFORMATIKO

VABILO K SODELOVANJU

Od leta 1977 je Informatica najpomembnejša sloven-ska strokovna revija na področju računalnitva in infor-matike pa tudi telekomunikacij, avtomatike in drugihsorodnih področij. V 1993 vstopa Informatica zuglednim mednarodnim uredniškim odborom in novo,panevropsko usmeritvijo. V njej bodo štirikrat lethoobjavljeni strokovni članki, ki jih bosta recenziralavsaj dva recenzenta izven države avtorja članka. Vdrugem delu revije bodo objavljene strokovne debate,ocene, mnenja, pomembne organizacijske in tržnespremembe in produkti na znanstvenem, pedagoškemin gospodarskem področju.

Prosimo Vas, da na ustrezen način predstavite revijosvoji okolici, npr. tako da jo postavite na viden pros-tor, omogočite dostop do nje in jo uvrstite v svojoknjižnico.

Revija Inforrnatica bo brezplačna za nekatere naj-pomembnejše znanstvene, pedagoške in ostale institu-cije v Sloveniji, drugi pa jo bodo lahko dobili z enoletnonaročnino, ki znaša za podjetja 3000 SIT (50 DEM), zazasebnike 1500 SIT (25 DEM) in za študente 500 SIT(10 DEM). Izpolnite prijavnico in nakažite ustrezniznesek na Slovensko društvo Informatika, Vožarski pot12, Ljubljana, žiro račun št. 50101-678-51841.

Če se aktivno udejstvujete na ožjem ali širšem po-dročju računalništva in informatike, Vas vabimo, dapošljete svoje prispevke (npr. članke ali propagandnemateriale) komurkoli izmed "Executive Editors". In-formatika bo skrbela za pošiljanje revije v vse pomem-bnejše raziskovalne, pedagoške, gospodarske in infra-strukturne institucije po Sloveniji in Evropi pa tudiAmeriki, zato je to enkratna priložnost za predstavitevVašega dela.

Prosimo Vas, da vrnete ta list z izpolnjeno anketo.

ANKETA

Pričakujemo brezplačno prejemanje

Naročamo revijo

Želimo si sodelovati

| I Način sodelovanja (opisno):

Predlogi izboljšav (opisno):

Izpolnjeno naročilnico pošljite na naslov: dr. RudiMurn, Informatica, Institut Jožef Stefan, Jamova 39,61111 Ljubljana.

Naročilnica na revijo INFORMATICA

Ime in priimek: .• Elektronski naslov:

Izobrazba: EMŠO (neobvezno):

Domač naslov in telefon:

V , dne

Služben naslov in telefon:

Podpis:

EDITORIAL BOARDS, PUBLISHING COUNCIL

Informatica is a journal primarily covering the Eu-ropean computer science and informatics community;scientific and educational as well as technical, commer-cial and industrial. Its basic aim is to enhance commu-nications between different European structures on thebasis of equal rights and international refereeing. Itpublishes scientific papers accepted by at least two ref-erees outside the author's couhtry. In addition, it con-tains information about conferences, opinions, criticalexaminations of existing publications and news. Fi-nally, major practical achievements and innovations inthe computer and information industry are presentedthrough commercial publications as well as throughindependent evaluations.

Editing and refereeing are distributed. Each editorcan conduct the refereeing process by appointing twonew referees or referees from the Board of Refereesor Editorial Board. Referees should not be from theauthor's country. If new referees are appointed, theirnames will appear in the Refereeing Board. Each pa-per bears the name of the editor who appointed thereferees. Each editor can propose new members fortHe Editorial Board or Board of Referees. Editors andreferees inactive for a longer period can be automat-ically replaced. Changes in the Editorial Board andBoard of Referees are confirmed by the Executive Ed-itors.

The coordination necessary is made through the Ex-ecutive Editors who examine the reviews, sort the ac-cepted articles and maintain appropriate internationaldistribution. The Executive Board is appointed bythe Society Informatika. Informatica is partiallysup-ported by the Slovenian Ministry of Science and Tech-nology.

Each author is guaranteed to receive the reviews ofhis article. When accepted, publication in Informaticais guaranteed in less than one year after the ExecutiveEditors receive the corrected versiori of the article.

Board of AdvisorsIvan Bratko, Marko Jagodič,Tomaž Pisanski, Stanko Strmčnik

Publishing CounčilTomaž Banovec, Ciril Baškovič,Andrej Jerman-Blažič, Dagmar Šuster,Jernej Virant

Board of RefereesC. Anglano, F. Abbattista, S. Bezjak, D. Bojadžiev,G. Božič, P. Giolito, I. Kononenko, F. Lanubile,N. Lavrač, J. P. Merlet, R. Sabatino, M. Sapino,J. Taylor, A. Werbrouk

Executive Editors

Editor in ChiefAnton P. ŽeleznikarVolaričeva 8, Ljubljana, SloveniaE-mail: [email protected]

Associate Editor (Contact Person)Matjaž GamsJožef Stefan InstituteJamova 39, 61000 Ljubljana, SloveniaPhone: +38 61 159 199, Fax: +38 61 161 029E-mail: [email protected]

Associate Editor (Technical Editor)Rudi MurnJožef Stefan InstituteJamova 39, 61000 Ljubljana, SloveniaPhone: +38 61 159 199, Fax: +38 61 161 029

Editorial BoardSuad Alagic (Bosnia and Herzegovina)Vladimir Batagelj (Slovenia)Andrej Bekeš (Japan)Leon Birnbaum (Romania)Marco Botta (Italy)Pavel Brazdil (Portugal)Andrej Brodnik (Canada) •Janusz Brozyna (France)Luca Console (Italy)Hubert L. Dreyfus (USA)Jozo Dujmovič (USA)Johann Eder (Austria)Vladimir A. Fomichov (Rvišsia)Janez Grad (Slovenia)Noel Heather (UK)Francis Heylighen (Belgium)Bogomir Horvat (Slovenia)Jadran Lenarčič (Slovenia)Angelo Montanari (Italy)Peter Mowforth (UK)Igor Mozetič (Austria)Oliver Popov (Macedonia)Sašo Prešern (Slovenia)Luc De Raedt (Belgium)Giacomo Della Riccia (Italy)Claude Sammut (Australia)Jiti Slechta (UK)Branko Souček (Italy)Harald Stadlbauer (Austria)Gheorghe Tecucci (USA)Robert Trappl (Austria)Terry Winograd (USA)Stefan Wrobel (Germany)

Vohime 17 Number 1 March 1993 ISSN 0350-5596

An International Journal of Computing and Informatics

Contents:Towards an Informational Orientation

Profiles: Terry Winograd

Methodoloeies from Machine Learnine: in Data

Editorial

A.P. Zeleznikar

D. Michie

1

2

3Analysis and Softvvare \

Analytical Form Solution of the DirectKinematics of a 4-4 Fully In-parallel ActuatedSix Degree-of-freedorh Mechanism

Towards a Mathematical Theory ofNatural-language Communication :

Alpha AXP Overview ;

Open Secure Model and its Functionality inNetworks with Value-added Services /

Real AI

Regular Graphs are 'Difficult' for Colouring

Metaphysicalism of Informing

C. Innocenti 13V, Parenti-Castelli -

V.A. Fomichov

J. Šilc

2 1

. 35

J. Buh : _.\- :

Borka Jerman-Blažič 41

M. Gams ; 53

J. Zerovnik ;

A.P. Železnikar 65

Mission and Research Reports •

- A Plan for the Knowledge Afchives Project- Center for the Study of Language and Information

81

8185

News and Announcements 101