Upload
jae-kyeong
View
217
Download
2
Embed Size (px)
Citation preview
This article was downloaded by: [University of Connecticut]On: 08 October 2014, At: 06:40Publisher: Taylor & FrancisInforma Ltd Registered in England and Wales Registered Number:1072954 Registered office: Mortimer House, 37-41 Mortimer Street,London W1T 3JH, UK
Applied ArtificialIntelligence: AnInternational JournalPublication details, including instructions forauthors and subscription information:http://www.tandfonline.com/loi/uaai20
Knowledge acquisition forknowledge based decisionsystemsJae-Kyeong KimPublished online: 26 Nov 2010.
To cite this article: Jae-Kyeong Kim (1997) Knowledge acquisition forknowledge based decision systems, Applied Artificial Intelligence: AnInternational Journal, 11:2, 131-150, DOI: 10.1080/088395197118280
To link to this article: http://dx.doi.org/10.1080/088395197118280
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of allthe information (the “Content”) contained in the publications on ourplatform. However, Taylor & Francis, our agents, and our licensorsmake no representations or warranties whatsoever as to the accuracy,completeness, or suitability for any purpose of the Content. Anyopinions and views expressed in this publication are the opinions andviews of the authors, and are not the views of or endorsed by Taylor& Francis. The accuracy of the Content should not be relied upon andshould be independently verified with primary sources of information.Taylor and Francis shall not be liable for any losses, actions, claims,proceedings, demands, costs, expenses, damages, and other liabilitieswhatsoever or howsoever caused arising directly or indirectly inconnection with, in relation to or arising out of the use of the Content.
This article may be used for research, teaching, and private studypurposes. Any substantial or systematic reproduction, redistribution,reselling, loan, sub-licensing, systematic supply, or distribution in anyform to anyone is expressly forbidden. Terms & Conditions of access
and use can be found at http://www.tandfonline.com/page/terms-and-conditions
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
KNOWLEDGE ACQUISITION FOR KNOWLEDGE-BASEDDECISION SYSTEMS
JAE-KYEONG KIMDepartment of Management Information Systems,Kyonggi University, Yiui-dong, Paldal-ku, Suwon,Kyonggi, Korea
Knowledge-based decision systems support the decision analysis process with decision
analytic knowledge as well as domain-specific knowledge. The knowledge in the knowledge-
based decision system is usually represented by influence diagrams. In the development of
knowledge-based decision systems, knowledge acquisition is one of the most difficult prob-
lems. This article suggests a knowledge acquisition process using verbal protocol analysis and
multi-attribute decision-making methodology. An environmental decision-making problem is
used as an illustrative example to explain the developed knowledge acquisition process. The
suggested process is so flexible that it can be expanded or applied to other problems with
ease. Some guidelines for the knowledge-based decision system are also suggested.
Everyone faces the challenge of making good decisions. Decision making is
complicated by competing alternatives, conflicting objectives, and uncertain conse-
quences. Uncertainty, in particular, makes decision making difficult, whether it
concerns the profitability of a new product or the health risk of toxic chemical
testing. Decision theory is an axiom-based framework that explicitly considers
uncertainty and trade-off. Decision analysis (DA) is an engineering discipline that
tells how to apply decision theoretic principles to real problems in a tractable manner
(Howard, 1984, 1988; Howard & Matheson, 1984). Since the early 1980s when
several knowledge-based systems proved to be successful, there have been some
approaches to apply knowledge-based systems to decision analysis (Chung et al.,
1992; Holtzman, 1985; Howard, 1988; Kim, 1991; Kim et al., 1990, 1992; Kim &
Kim, 1991; Reed, 1989; Russel et al., 1988). A knowledge-based system that
implements the decision analysis is referred to as the decision analysis expert system
(Howard, 1988), the intelligent decision system (Holtzman, 1985, 1989; Kim, 1995;
McGovern et al., 1991) or the knowledge-based decision system (Chung et al., 1992;
Kim, 1991; Kim et al., 1992; Park et al., 1993). The common properties of these
systems are that they provide their users with a substantial amount of domain-
specific knowledge and normative power of decision analysis. In this article, the
Applied Artificial Intelligence, 11:131± 149, 1997Copyright � 1997 Taylor & Francis0883-9514/97 $12.00 + .00 131
This work was done when the author was in the IDS department, University of Minnesota as a visiting faculty
member. The author wishes to thank Professor Carl R. Adams for his careful review and comments on content and
style, and Melissa L. Lange for her help on the interview and analysis of the verbal protocol.
Address correspondence to Jae-Kyeong Kim, Department of Management Information Systems, Kyonggi
University, San 94-6, Yiui-dong, Paldal-ku, Suwon, Kyonggi, Korea.
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
knowledge-based decision system (KBDS) is used as a representative tool for
decision making. A KBDS constitutes a means by which decision makers can exploit
the normative power of decision analysis in a relatively simple and inexpensive way.
In the development of KBDS, there is no general agreement on the use and
appropriateness of the representation scheme and various measures of uncertainty.
It is, however, widely recognized that influence diagrams are used to represent
domain knowledge, and probability inference using Bayesian theory is used to
handle uncertainty (Chung et al., 1992; Holtzman, 1985, 1989; Howard, 1988; Kim,
1991, 1995; Kim et al., 1990, 1992; Park et al., 1993). The concept underlying
influence diagrams is the representation of variables critical to the problem under
consideration, and their interrelationships, in a graph-theoretic manner (Howard &
Matheson, 1984; Olmsted, 1983). Much of the power of decision analysis lies in its
ability to effectively integrate the many factors that commonly affect a decision, so
influence diagrams are used as a knowledge representation scheme of KBDS in this
research.
Domain-specific knowledge of KBDS is represented as influence diagrams,
which consist of decision variables, uncertainty variables, criteria values, and their
relationships (Holtzman, 1985, 1989; Kim, 1991; Kim et al., 1990, 1992). However,
it is not an easy problem for the decision makers and decision analysts to elicit (and
define) many variables and integrate them effectively to structure a decision prob-
lem. These difficulties caused by the complex and evolutionary nature of the
structuring process are considered to be a bottleneck to the widespread use of
decision analysis as well as to the development of KBDS (Holtzman, 1985, 1989;
Howard, 1988; Kim, 1991). Another problem is that there has been little research
on how to modify an existing knowledge base or model base of KBDS to solve
similar problems (Kim et al., 1990; Owen, 1984). The amount of effort, money, and
time spent formulating a decision problem into an influence diagram is very
burdensome. It is also difficult to apply the knowledge or information obtained while
formulating one decision problem into other similar problems. This is another
difficulty to the development of KBDS (Holtzman, 1989; Howard, 1988).
It is not easy to find a research resource on the implementation of KBDS,
especially on the knowledge acquisition of that system. Research on the influence
diagram structuring process surveyed in this research is related to the knowledge
acquisition process of KBDS. RACHEL is the first KBDS to diagnose infertility
(Holtzman, 1985, 1989). It defines a class of decisions to build a domain-specific
system for that class. However, RACHEL is a character-based system, having no
graphic user interface. KIDS is for strategic decision problems by HyperCard
(Chung et al., 1992; Kim, 1991; Kim et al., 1992). It has a graphic user interface and
knowledge base to structure an influence diagram. The main limitation of this system
is its limited knowledge base capacity. The interview process (Holtzman, 1989;
Howard, 1988) and the goal-directed approach (Holtzman, 1989; Kim et al., 1990;
Newell & Simon, 1977) are frequently used to structure an influence diagram in the
132 J.-K. Kim
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
decision-making process. The interview process is a kind of conversation process
interchanging domain-specific knowledge and decision-analytic knowledge be-
tween decision makers (or domain experts) and decision analysts. One of the main
purposes of a knowledge-based system (including KBDS) is to mimic or replace an
expert (Kim, 1991). Thus the interview process alone is not suitable for the KBDS
because it needs the decision analyst(s) at all times. INFORM by Moore and Agogino
(1987; Russel et al., 1988) focuses on editing the knowledge base using a domain
independent tool. The induction of influence diagrams from a set of examples is
conducted in the domain of strategic management constructing a knowledge base.
McGovern et al. (1991) suggest the elicitation of decision structures through text
analysis. They emphasize that this approach is primarily exploratory with less well
structured problems. They argue that when combined with natural language process-
ing software, it will be possible to acquire knowledge of KBDS. Most real decision
problems are very diverse and slightly different even within similar problem
domains. Thus the knowledge from text alone is not suitable for the KBDS.
In this research, we develop a knowledge acquisition process based on verbal
protocol analysis (VPA) and multiple-attribute decision making (MADM) for the
development of KBDS. The VPA has been extensively applied to understanding
human problem solving to identify the problems and current decision-making
processes (Ericsson & Simon, 1984; Newell & Simon, 1977). VPA is used in this
research because it is a powerful tool for extracting knowledge from human experts
when they solve real managerial decision problems with their heuristic knowledge
combined with decision-analytic knowledge and domain-specific knowledge. VPA
makes it especially easy to transfer the extracted knowledge to a knowledge base
using a well-defined coding scheme. Merkhofer (1990; Merkhofer & Keeney, 1987)
suggests the use of an influence diagram to solve MADM problems. His idea of
combining an influence diagram and MADM is used in this research to model a
complex decision problem and to apply the developed model to other similar
problems. The main ideas proposed in this article can be summarized as follows.
1. VPA helps to determine the scope or boundary of the decision problems. The
analysis makes it possible to capture abstract knowledge of the influence diagram
and related information with ease, even if decision makers have their own heuristic
knowledge, rather than decision-analytic knowledge.
2. A hierarchical-type MADM makes it easy to classify the knowledge of
components or objectives of decision problems. Influence diagrams combined with
MADM methodology make it possible to enlarge or modify an existing model to be
adjusted to similar problems. MADM can also handle well the trade-off between
conflicting objectives.
3. This study provides the expansion procedure from a high-level influence
diagram (i.e., general, rough, or abstract-level influence diagram) to a low-level
influence diagram (i.e., more detailed or concrete influence diagram). This proce-
Knowledge Acquisition for Knowledge-Based Decision Systems 133
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
dure is very helpful for modeling of complex decision problems. It also makes it
easy for decision makers (including domain experts) to make their abstract
knowledge more concrete.
The process is explained by applying it to an environmental decision problem.
Some guidelines are suggested in regard to how to transfer or modify the knowledge
base for another similar problem. In the next section, the basic concept of influence
diagrams with inference problems is provided, and the knowledge acquisition
process, and VPA are described. Then, building an influence diagram, expansion
procedure, and transferring knowledge between decision problems are explained for
an environmental problem. Difficulties of developing KBDS and guidelines to
overcome them are also discussed. Finally, conclusions and possible future research
areas are provided.
KNOWLEDGE REPRESENTATION
There have been many approaches to implement the decision analysis process
with the help of computers. A KBDS applies the descriptive ideas of knowledge-
based system theory at a higher level, so it can manage the process of formulating,
evaluating, and appraising a decision problem. The knowledge of this system would
be knowledge for the creation of decision models, knowledge for their evaluation,
and knowledge for explaining their implications (Holtzman, 1985, 1989). Although,
the human decision analyst is too general-purpose to be captured everywhere, the
KBDS can be built only within a particular domain of application. An influence
diagram is used as a knowledge representation tool and inference tool of KBDS in
this research.
Influence Diagrams
An influence diagram is a major knowledge representation tool of KBDS. It was
developed as a computer-aided modeling tool by Miller et al. (1976) and Howard
and Matheson (1984). Olmsted (1983) proposed a computational architecture for
automating influence diagrams along with rules specifying the topological transfor-
mations needed for his solution procedure. Shachter (1986) developed a goal-
directed algorithm based on these rules. In contrast to other models, influence
diagrams focus decision makers’ attention on only those problem features that are
most relevant to the focus at hand. Excluding irrelevant information from an
influence diagram can save decision makers time and effort, since there are fewer
variables to be interpreted. If decision makers want to pay attention to a certain part
more specifically, they can expand part of an influence diagram to a lower level.
The flexibility of influence diagrams makes it easy for decision makers to communi-
cate with each other and to store influence diagrams for future use.
134 J.-K. Kim
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
An influence diagram is defined as a network consisting of a directed graph with
no directed cycles. The diagram consists of nodes and arcs (influences, or arrows)
that represent a decision basis (Howard & Matheson, 1984; Howard, 1988). Each
node in the graph represents a variable in the model. This variable may be a constant,
an uncertain quantity, a decision to be made, or a criterion value. The nodes can be
viewed from three levels: relational, numerical, and functional (Holtzman, 1985;
Howard & Matheson, 1984; Kim et al., 1992). Figure 1 shows an influence diagram
for the decision problem of a raw material buyer (Chung et al., 1992; Kim, 1991;
Kim et al., 1992). The nodes and arcs represent the relational level, i.e., showing the
major components, their types, and their interdependence.
In Figure 1, ª amount of oil purchaseº is a decision node. Decision nodes
represent decision options that are available to a decision maker (DM). The arcs to
a decision node represent information available at the time the decision is made.
Each decision node denotes variables under the control of the DM. It has branches
to represent the possible options. Each chance node has an underlying probability
distribution to quantify the uncertainty for the variable represented by that node.
Arcs into chance nodes represent information affecting the probability distribution
for that node. Chance nodes are generally represented as a cycle or an oval in the
influence diagram. ª Total oil priceº is a value node, and it summarizes the preferen-
ces of the DM for the decision outcomes. A mathematical function (value function
or utility function) associated with a value node can be used for deriving a numeric
value representing the trade-off among attributes of the problem. Direct predecessor
Figure 1. Influence diagram for the oil purchase problem.
Knowledge Acquisition for Knowledge-Based Decision Systems 135
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
nodes of a value node indicate the attributes included in the evaluation of preferen-
ces. See Olmsted (1983) and Howard and Matheson (1984) for more details about
influence diagrams. The next subsection discusses the probabilistic inference using
influence diagrams containing only chance nodes.
Inference Problem Using Influence Diagram
A probabilistic influence diagram is a special type of influence diagram that
contains only chance nodes. Once a problem is modeled using the influence diagram,
problem analysis and probabilistic relationships can be manipulated at the graph
level of the diagram. In the KBDS, knowledge inference can be made if the decision
maker needs a more detailed analysis. In 1988, Shachter showed how any prob-
abilistic inference problem can be solved through basic operations on an influence
diagram (Schachter, 1988). Also, Rege and Agogino (1988) formally defined the
consistent transformation of influence diagrams, probabilistic inference problems,
and a polynomial-time algorithm to solve the same transformed problem at a
symbolic and topological level. In this section, a brief explanation of probabilistic
inference is given.
Consider n state variables x1, x2, . . . , xn involved in the decision problem being
modeled. Each variable represents a factor or component of the problem, which has
a finite set of possible outcomes and a conditional probability distribution over those
outcomes. The information about the problem can be represented in terms of a joint
probability distribution of all variables denoted by P(x1, x2, . . . , xn). Such a joint
distribution can be expanded into any of n! expansions. The following equation
represents one of the possible expansions: P(x1, x2, . . . , xn) = P(x1)P(x1|x2), . . . ,
P(xn|x1, x2, . . . , xn ± 1).
Shachter’ s (1988) probabilistic influence diagram is an influence diagram contain-
ing only nodes that represent constant or uncertain quantities. A probabilistic influence
diagram represents one expansion of the joint distribution of the state variables by means
of a directed graph. Each variable in a probabilistic influence diagram has a data frame
within its associated node, in which there is a finite set of outcomes, and a conditional
probability distribution over those outcomes. The conditional relationship of each node
is represented by arcs in the diagram. If there is no undirected path between two nodes,
then they must be independent. If arc (i, j) is part of the diagram, then the assessed
distribution for the variable j is conditional on the value of i. When a chance node has
no arcs to it, then the assessed distribution is a marginal distribution. There are two types
of variables in a probabilistic influence diagram. A deterministic variable has a degen-
erate conditional distribution. If it does not, it is called a probabilistic variable. We might
be uncertain about the value of a probabilistic variable even after observing the values
of its conditioning variables. In the case of a deterministic variable, we are certain of its
value, given the value of its conditioning variables, although we might be uncertain
about its value if we could not observe their values.
136 J.-K. Kim
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
The general probabilistic inference problem is to determine P[fo(xj)|xk], where
J and K are arbitrary subsets of set N and fo is an arbitrary real-valued function
defined by the outcomes of xj given an influence diagram G = (N, A), where N is a
node set and A is an arc set. It can also be represented as P(x i|x j), where i is an arbitrary
node in the node set N. A probabilistic database for the variables xN is assumed to
have been assembled in the form of a partially specified influence diagram (Shachter,
1986). For example, suppose the OPEC policy is changed and the raw material buyer
of a textile cooperation wants to know the likelihood that the oil price in the spot
market will increase given that condition. If such a problem is represented as in
Figure 1 except for the decision and value node, he/she should solve the probability
inference problem to obtain P(Spot Price | OPEC Policy). Shachter (1986) and Rege
and Agogino (1988) developed a very similar reduction algorithm for probabilistic
inference problems. They use node reduction and arc reversal procedures, originated
from Howard and Matheson (1984) and Olmsted (1983). See Matzkevich and
Abramson (1995) for a survey of related studies.
KNOWLEDGE AC QUISITION
Knowledge Acquisition Process
The notion of mental representation is the principle that unifies research refer-
ences to problem setting, problem conceptualization, locating the problem, require-
ment determination, and framing and structuring decisions (Smith, 1988, 1993).
Each of these activities involves development of a mental representation of a
situation that deserves attention, thought, and action. Cognitive scientists claim that
having a mental representation of a problem is a necessary condition for trying to
solve it (Newell & Simon, 1977; Smith, 1988). Such representations can be exter-
nalized using various symbolic media. An influence diagram is one method for
external representation of uncertain, complex decision problems. Decision analysis
is regarded as a systematic procedure for transforming opaque decision problems
into transparent decision problems by a sequence of transparent steps (Howard,
1988). Opaque means ª hard to understand, solve, or explain; not simple, clear, or
lucid,º and transparent means ª readily understood, clear, obvious .º In other words,
decision analysis offers to a DM the possibility of replacing confusion by clear
insight. The transformation is generally composed of formulating, evaluating, and
appraising the decision problems.
The first step of a formulation fits a formal model to the DM’ s opaque real
situation. This formal representation of the decision problem is usually represented
as an influence diagram. The formulation process (or other process) is usually
conducted through conversation between the DM (and domain experts) and a
decision analyst (or knowledge engineer). During the formulation process, the DMs
or domain experts teach the details of the decision at hand (e.g., available resources,
Knowledge Acquisition for Knowledge-Based Decision Systems 137
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
attitude toward risks, or conflicting objectives) to the decision analyst, who learns
by building an appropriate influence diagram (Holtzman, 1985, 1988). This conver-
sation seems to be inevitable when DMs confront difficult decision problems, and
they have an uncanny ability to list the factors relevant to their problems or their
relationships. However, people who face similar decision problems again and again
have their own heuristic decision-making process by learning from experience, right
or wrong.
Based on previous research, the key points of knowledge acquisition of KBDS
may be summarized by the following two facts. In contrast to decision-analytic
knowledge, which is almost permanent, domain-specific knowledge is changed very
rapidly from situation to situation. First, domain-specific knowledge of a problem
should be obtained from human domain experts or DMs. Second, after the
knowledge is stored in the knowledge base of KBDS, similar decision problems
should be solved from the existing knowledge base after a slight extension or
modification. This research thus suggests a knowledge acquisition process based on
VPA and MADM. The main purpose of VPA is to obtain a clear understanding of
the current decision-making process. Designing a better process starts from a clear
understanding of the current process. The VPA technique has been extensively
applied to understanding human problem solving in order to identify the problems
and current decision-making process (Ericsson & Simon, 1984; Newell & Simon,
1977). VPA is not only a powerful tool extracting knowledge from human experts
but it also makes it easy to transfer the extracted knowledge to a knowledge base
using a well-defined coding scheme. MADM can treat well the relationship and
trade-off of objectives, so an influence diagram combined with MADM is very
useful. Furthermore, in a hierarchical structure, two similar decision problems are
identical at a higher level even though they are slightly different at a lower level.
Thus a similar decision problem can be easily formulated from the high-level
influence diagram in the knowledge base of KBDS.
Verbal Protocol Analysis
To elicit knowledge from DMs or (domain) experts, it is essential to gain a clear
understanding of the current decision situation and decision-making process. Verbal
protocols are verbal reports obtained from experts who have been instructed to
ª think out loudº while solving problems. DMs and experts are requested to express
their thoughts step-by-step while solving the decision problems presented. VPA
identifies the states that are explored and the knowledge that is accessed and applied
to transform one state into another. The results of VPA are used to constrain the
normally unconstrained task of knowledge representation design. They are given a
decision problem, which may be new or similar to one seen before. They are
requested to express their thoughts step by step into a tape recorder while solving
the decision problems provided. If there are aspects of the current decision-making
138 J.-K. Kim
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
process with which a DM is not satisfied, it is important to identify these in order to
design a new process. It is also important that the DM mention instances of decisions
made, largely or in part, based on rules of thumb, intuition, emotions, or external or
political constraints rather than by using a predefined decision-making process.
After protocols from DMs are collected, they are transcribed and analyzed
according to a well-defined developer-specific coding scheme. An illustrative
example of using protocol analysis is shown in the next section. The following
guidelines are suggested to DMs for the protocol analysis of this research.
1. Look carefully at the attached decision problems. You may have encountered
this decision previously, yet attempt to look at it as if you were seeing it for the first
time.
2. Do not think through the problem fully and then begin recording your
problem-solving process. It is the process of ª thinking through the problem fullyº
that we wish you to record. Look at it once, and then express your thinking process
precisely and step by step into the tape recorder. Please be as frank and thorough as
possible.
3. The thinking process that is to be recorded begins when you first look at the
problem, and ends once you have made your decision for a given problem.
4. The most important point is to express what you are thinking. W hat are you
taking into consideration? How do these factors relate to each other?
The VPA is frequently used to obtain information regarding a subject’ s thoughts
while focused on a problem-solving task (but not on the subjects’ interpretation of
these thoughts) (Ericsson & Simon, 1984; Newell & Simon, 1977). This method is
frequently used to gather information about the cognitive processes of a subject.
System developers (the author, in this research) analyze the verbal protocol obtained
from the recording machine. In such cases the data are ª encoded.º Encoding is the
process for translating raw protocols into a form suitable for analysis by the various
models being tested. In this research, the data were encoded mainly to seek answers
to the following questions: W hat is the decision problem at hand? What is being
decided? What are the objectives that need to be achieved through the decision?
What information is needed (and is it obtainable) as one proceeds? What and where
are the main uncertainties and risk factors? What is the approach used in weighing
the issues in relation to each other?
The respondents were asked to talk into a tape recorder while solving the
decision problem. The decision problem used in this case differs from tasks typically
used in VPA because the subjects cannot entirely solve the task at hand, since its
solution depends on interactions with other sources. The respondents could, how-
ever, arrive at their own conclusions without difficulty. The VPA provides useful
knowledge of the disjoined nature of thinking out loud. Something a DM says at one
time may contradict something said earlier or later. It is a very helpful technique to
Knowledge Acquisition for Knowledge-Based Decision Systems 139
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
elicit knowledge about the problem domain and the decision-making process. It also
helps to draw a high-level influence diagram easily and obtain some unconnected
detailed information.
ENVIRONMENTAL REVIEW DEC ISION MAKING: AN ILLUSTRATIVE EXAMPLE
Endangered Species Problem
The Minnesota Department of Natural Resources (DNR) considers a project
proposal in the environmental review process. Their considerations when they
review a project proposal are as follows: the legal status of the species, the viability
of that instance of the species, and effects of changes and species loss on the larger
ecological community picture. They usually also consider the social and economical
benefits and the consequences or legal repercussions for denying the permit. But
their main concerns are what effects the project will have on the species on site and
on the state of Minnesota. They encounter this decision about two or three times a
year. DNRs in other states consider similar problems. Some of the previous and
current decision situations are summarized in Table 1.
The decision situations described in this research are those that assist in deciding
between two possible, yet mutually exclusive courses of action, both resulting in a
win-lose situation. Periodically, the Minnesota DNR Natural Heritage and Nongame
Research Program encounters situations that are similar but vary in location, project
type, and number and type of species threatened (Table 1 depicts some of these
decision situations). In previously encountered cases, the decision-making process
occurred on a basis of heuristic knowledge. The DMs did not always follow an
explicit set of criteria but have their own criteria, which are slightly different from
person to person. This made it difficult to consider the multiple criteria effectively
and efficiently, and to defend the process. Some DMs also did not have an explicit
approach for assessing the relative importance of various factors or for evaluating
Table 1. Diverse decision problems of the Department of Natural Resources
Project Considerations
Landfill Expansion Existing landfill expanded onto 63 acre site
· Proposed site 17 acres of prairie, 20 acres of wetlands, 1 endangered and 1 threatened plant species
· Crucial considerations Quality and rarity of site, social pressures
Burnsville Flood Control Impound a lake in order to dam an area to prevent flooding in downstream areas
· Proposed site Very rare MN endangered plant; habitat found also very rare
· Crucial considerations Many people involved, large momentum of project, precedent the decision will set
Giants Ridge Build a large-scale development and golf course next to existing ski resort area
· Proposed site 3 endangered plant species, 2 threatened plant species, 1 plant of special concern
· Crucial considerations Forest fragmentation and water quality; no permits granted to date for a
development project
140 J.-K. Kim
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
trade-offs. The decision of whether to grant or deny can be solved using decision
analysis methodologies. The DNR DM has domain-specific knowledge and decision-
analytic knowledge, but the whole decision-making procedure is not well developed.
Knowledge acquisition for a landfill expansion project is performed below using
VPA and MADM.
Knowledge Acquisition Using VPA
DMs are requested to express their decision-making process and domain-specific
knowledge in as much detail as possible into a tape recorder. Protocols from four DMs
are collected for this study. They are analyzed from sentence to sentence, the sentences
are grouped by their similarities, and some assertions are elicited. These assertions are
helpful in understanding the problem and the process. The following are some assertions
and supporting protocol fragments. The number preceding each sentence represents
sentence sequences by the author’ s encoding scheme.
Assertion 1: First, DMs want to understand the problem, its decisions, and their
objectives.
DM 1:
01 I assume that the problem we are looking at is where to issue the permit of
the species that are listed . . . to be specific, what we are deciding is whether to issue
a permit for the taking of two state-threatened plant species that are known to occur
on this site.
02 The objectives that have to be taken into consideration are the protection of
endangered species in this state, preventing them from becoming extirpated but also
the integrity of the endangered species act so that it will be available for future use
as well as present use. . . . umm.
DM 3:
01 A landfill expansion . . . with 17 acres of prairie, 15 plant, 20 acres wetland.
02 It seems obvious that in the metro area, a 20-acre wetland to be proposed to
be filled would require a 404 Water Quality Permit from the Corps of Engineers and
quite likely a permit from the DNR under the Wetlands Conservation Act. . . .
04 The 17 acres of prairies have no legal protection; there is no jurisdiction here
by the DNR.
Assertion 2: DMs extract major uncertainty factors first, and other related fac-
tors step by step.
DM 4:
08 Um. The main uncertainties and risk factors. . . . I am using the bullets from
the sheet to talk about this process. . . . One big uncertainty is what the long-term
Knowledge Acquisition for Knowledge-Based Decision Systems 141
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
viability of the prairie and the species on this site would be if the project doesn’ t go
forward.
13 Another uncertainty . . . is what sort of mitigation to ask for . . . how to go
about determining what the appropriate mitigation is, and how to successfully
negotiate with the project proposer to get it, are uncertainties that have to be dealt
with. . . .
DM 2:
15 It is of extremely high quality and it is the habitat of two threatened plant
species both of which have very well established and viable populations at this site.
16. And this landfill would not simply take part of the prairie and few of the
plants but would destroy all of the plants and all the habitat.
Assertion 3: DMs consider several objectives and related factors of each objec-
tive, and synthesize the relationships among objectives using trade-offs.
DM 1:
27 . . . my approach in weighing the different issues in relation to one another. . . .
28 . . . look at the significance of the plant species on the site and the prairie in
the context of the state as a whole.
DM 3:
32 So my first reaction is just to say no.
34 Ah . . . however, certainly there has to be considered, there is a need for more
landfill space . . . it would be economically and politically easier to simply expand
the current landfill into this area than it would be to sight an entirely new landfill
somewhere else. Opening landfills are very, very difficult.
Besides the above decision-making assertions, the elicited properties of this
problem are listed as follows:
· Main decision is made by several DMs after considering the trade-off between
viability at that instance of that species with social and economic benefits.
· The major uncertainty of the project that should be considered by DNR is the
probability of existence of species on site or in state.
· Another uncertainty is the projected viability of the population on site proposal
is permitted or not.
· Decisions of a similar type are to be made frequently, and these decisions strongly
depend on the specific situation of the site and project. Thus, if an influence
diagram is structured, it can be used repeatedly with a slight change of structure
and information.
· The species information, such as threatened, endangered, of special concern, total
number at site and in state, and density, can be stored in the database or knowledge
base of KBDS.
142 J.-K. Kim
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
Building an Influence Diagram from VPA and MADM
MADM is well suited for comparing options against multiple objectives.
Competing objectives, such as ª to maintain biodiversity of Minnesota and its social
benefitsº and ª to increase economic and other social benefits,º are common in
environmental problems. MADM establishes attributes for measuring the degree to
which identified objectives are achieved. On the basis of the protocol analysis, we
set up three objectives and related attributes of DNR decision problems.
1. To maintain biodiversity of the state
· Ensure the continued existence of species on site
· Ensure the continued existence of species in state
2. To increase the social benefit of state residents
· Increase social benefit from this site
· Increase economic benefit from this site
3. To abide by practical considerations
· Save time and cost of knowledge elicitation and decision making
· Avoid lawsuits
· Give consistent response to applicants.
On the basis of this hierarchical value structure and VPA, the following steps
are used to develop an initial high-level influence diagram in this research.
Step 1: Isolate objectives from verbal protocols
· Biodiversity in state (to maintain biodiversity of state)
· Benefit (to increase the social benefit of state resident)
· DNR (to abide by practical considerations)
Step 2: Isolate decision elements from verbal protocols
· Permit decision
· Viability in state
· Viability of the population on site
· Lawsuit
· Result of suit
Step 3: Establish relationships between elements from verbal protocols (the relation-
ships between elements are listed in pairs)
· Permit decision influences benefit
· Permit decision influences viability of the population on site
· Lawsuit influences result of suit
Step 4: Draw an influence diagram. Each node is examined, and its type either
decision, chance, or value node, is determined.
As a result, Figure 2 is constructed.
Knowledge Acquisition for Knowledge-Based Decision Systems 143
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
Expansion Procedure of an Influence Diagram
The major uncertainties of this problem (see Figure 2) are represented as the
following inference problem: P(viability of the species on site|proposal permit) and
P(viability of the species on site|proposal denial). Because this probability is not
easy to elicit directly from DMs, we expand this node into a lower level through an
interview process with DMs. Figure 3 represents part of such an expanded low-level
diagram.
Figure 4 represents the final version of the influence diagram of the landfill
project problem after the interview with decision makers. The main purpose of the
interview is to obtain the exact knowledge of a probability distribution and preferen-
ces from DMs. But to fulfill such a purpose , a more detailed influence diagram is
helpful. Building a final diagram currently occurs by obtaining all the knowledge.
A major benefit of influence diagrams comes when it is time to analyze the model
after building. Because an influence diagram is logically precise, there is no need to
transform the diagram into a form appropriate for analysis. The same graphical
representation that is so effective for modeling is also effective for analysis. Because
Figure 2. High-level influence diagram of the Department of Natural Resources.
144 J.-K. Kim
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
transformations of influence diagrams correspond to applications of probability
calculus, graphical techniques can be used to transform the diagrams to more
convenient forms without getting bogged down in mathematical details. See Olm-
sted (1983) and Shachter (1986) regarding transformations.
Transfer Knowledge Between Decision Problems
The Minnesota Natural Heritage Program must periodically make decisions
requiring transfer of knowledge from other problems. Analysis of three previous
decisions (in Table 1) indicates that the situations are similar, varying primarily in
the project type and location and in the number and type of species that would be
affected.
The Burnsville Flood Control Project and the Giants Ridge Project can be
modeled with ease from the decision model shown in Figure 4. The models are
relationally very similar because many variables are common between the problems.
However, the functional level or the numerical level will be somewhat or very
different. For example, we are to model the Burnsville Flood Control Project from
the Landfill Expansion Project. The influence diagram of the Burnsville project is
identical to that of the landfill project on the relational level. Thus Figure 4 can be
used for the Burnsville project. Furthermore, the options of the decision node are
same. The species database and knowledge base for the entire state will be used
without modification, but the species to be considered are different in the two
Figure 3. Expansion of one-chance node.
Knowledge Acquisition for Knowledge-Based Decision Systems 145
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
Figure 4. Influence diagram of landfill problem.
146
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
projects. Thus possible state and probability of chance node and value equations are
totally different from those of the landfill problem. DMs or domain experts may feel
confident in giving a subjective probability to a specific situation given the whole
structure when they face a similar problem. The DMs can solve the problems without
the help of decision analysts or knowledge engineers. The KBDS helps the DMs to
see the problem in a global sense and provides guidelines to elicit a situation-specific
knowledge (Kim, 1991; Park et al., 1993).
GUIDELINES TO KBDS
The following four guidelines are suggested for using KBDS for managerial
decision problems. The basis of these guidelines comes from existing studies.
First, it should be identified whether the problems are to be solved by decision
analysis, to determine if an influence diagram would be a suitable modeling tool.
Similar decision problems should occur frequently. For these kinds of problems, the
KBDS is a decidedly efficient system, saving effort, money, and time for decision
analysts and DMs (Kim, 1991).
Second, similar problems can be conveniently treated as one unit: a class of
decisions. Holtzman (1985, 1989) first suggested the concept of decision class
analysis. He defines a class of decisions as a set of decisions having some degree of
similarity that can be treated as a single unit. The decisions in a class share a common
domain and, perhaps, also some common situation features. Just as the end result of
a specific decision analysis is a decision, the end result of a decision class analysis
is a specific decision analysis. This is appropriate in small-frame situations. How-
ever, Holtzman does not explain how to construct a class of decisions and how to
decide whether to include a new problem in a certain class or not. The concept of
decision class analysis is unclear in some respects. The concept of a class of
decisions is very important but very difficult to define. The next problem is the
methodology of implementing the DCA. Holtzman used a rule-based approach,
but a rule-based approach is well suited to a narrow decision scope in the author’ s
experience (Chung et al., 1992; Kim et al., 1992). His idea of decision class
analysis is, however, thought to be the first stage in automating the decision
analysis process.
Third, the knowledge elicitation of a problem should be conducted by a
goal-oriented approach. In this article a mixture of protocol analysis and interview
process is suggested. A MADM is thought to be a good methodology for a complex
goal structure. A protocol analysis represents defining current problems, scope,
relationship of factors, needed information, preference, goal structure, and view-
point of the problem. Using a verbal analysis, it is possible to construct a high-level
influence diagram and other knowledge, as we have shown before. If necessary, a
node can be expanded to make it easy to elicit knowledge from DMs. This procedure
Knowledge Acquisition for Knowledge-Based Decision Systems 147
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
saves time and effort and avoids the need for trial and error if the problem is too
complex.
Fourth, decision problems of the same class are solved with ease using the initial
influence diagram because class problems have similar boundaries and share many
of the problem parts. After defining and solving one class of decision problems,
case-based reasoning (CBR) is thought to be a useful tool if KBDS are required to
seek automatic classification procedures (Beck et al., 1994). CBR is a technique that
compares new individual decision problems with an existing class of problems
stored in a knowledge base. This comparison is made on the basis of similarity
defined in terms of the problem structure. Individual decision problems may be
grouped into a new class based on similarities in their structures, regardless of any
existing class description, if they cannot fit into existing classes. Thus this kind of
classification algorithm may be a good alternative to the knowledge acquisition
process of KBDS.
CONC LUSION
This article presents a knowledge acquisition process using verbal protocol
analysis for the development of KBDS. The influence diagram is used as a
knowledge representation tool. VPA is used in the knowledge acquisition process
of KBDS because it is a powerful tool extracting knowledge from human experts
when they solve a real managerial decision problem with their heuristic knowledge
combined with decision-analytic knowledge and domain-specific knowledge. VPA
makes it especially easy to transfer the extracted knowledge to a knowledge base
using a well-defined coding scheme. The influence diagram combined with MADM
is also used to make it easy to elicit knowledge of a complex decision problem and
to apply the knowledge to similar problems. To elicit more accurate probabilities of
an influence diagram, an expansion procedure is introduced. It includes a discussion
transferring knowledge between similar decision problems. Some guidelines are
suggested for the development of KBDS.
Defining a class and storing it efficiently for similar problems is a promising
research area. A CBR algorithm combined with defining a class is also a possible
area for further research.
REFERENC ES
Beck, H. W., T. Anwar, and S. B. Navathe. 1994. Conceptual clustering algorithm for database schema design.
IEEE Transactions on Knowledge and Data Engineering 6(3):396±411.
Chung, T. Y., J. K. Kim, and S. H. Kim. 1992. Building an influence diagram in a knowledge-based decision system.
Expert Systems with Applications 4:33±44.
Ericsson, K. A., and H. A. Simon. 1984. Protocol analysis: Verbal reports as data. Cambridge, Mass.: MIT Press.
Holtzman, S. 1985. Intelligent decision systems. Ph.D. thesis, Dept. of Engineering-Economic Systems. Stanford,
Calif.: Stanford University.
148 J.-K. Kim
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
Holtzman, S. 1989. Intelligent decision systems. Reading, Mass.: Addison-Wesley.
Howard, R. 1984. The evolution of decision analysis. In The principles and applications of decision analysis, vol.
I, eds. R. Howard and J. Matheson. Menlo Park, Calif.: Strategic Decision Group.
Howard, R. 1988. Decision analysis: Practice and promise. Management Science 34(6):679±695.
Howard, R., and J. Matheson. 1984. Influence diagrams. In The principles and applications of decision analysis,
vol. II, eds. R. Howard and J. Matheson. Menlo Park, Calif.: Strategic Decision Group.
Kim, J. K. 1991. A knowledge-based system for decision analysis. Ph.D. thesis, Dept. of Industrial Engineering,
KAIST, Seoul, Korea.
Kim, J. K. 1995. A study on the development of intelligent decision systems using influence diagram. Journal of
the Korean OR/MS Society 20(3):77±104.
Kim, J. K., T. Y. Chung, and S. H. Kim. 1992. KIDS: A knowledge-based decision system to build an influence
diagram. Paper presented at the First World Congress on Expert Systems, Orlando, Florida, 16±19 Dec.
Kim, J. K., B. S. Kang, and S. H. Kim. 1990. Decision-theoretic approach to a knowledge-based clinical system.
Expert Systems with Applications 1(1):3±15.
Kim, S. H., and J. K. Kim. 1991. Explanation in a decision-theoretic consulting system: An axiomatic approach.
Applied Artificial Intelligence 5:393±409.
Matzkevich, I., and B. Abramson. 1995. Decision analytic networks in artificial intelligence. Management Science
41(1):1±22.
McGovern, J., D. Samson, and A. Wirth. 1991. Knowledge acquisition for intelligent decision systems. Decision
Support Systems 7:263±272.
Merkhofer, M. W. 1990. Using influence diagrams in multi-attribute utility analysisÐ improving effectiveness
through improving communication. In Influence diagrams, belief nets and decision analysis, eds. R. M. Oliver
and J. Q. Smith, pp. 297±319. New York: John Wiley.
Merkhofer, M. W., and R. L. Keeney. 1987. Multi-attribute utility analysis of alternative sites for the disposal of
nuclear waste. Risk Analysis 7:173±194.
Miller, A. C., M. W. Merkhofer, R. A. Howard, J. B. Matheson, and T. R. Rice. 1976. Development of automated
computer aids for decision analysis. SRI International Technical Report, Menlo Park, Calif.
Moore, E. A., and A. M. Agogino. 1987. INFORM: An architecture for expert-directed knowledge acquisition.
International Journal of Man-Machine Studies 26:213±230.
Newell, A., and H. A. Simon. 1977. Human problem solving. Englewood Cliffs, N.J.: Prentice-Hall.
Olmsted, S. M. 1983. On representing and solving decision problems. Ph.D. thesis, Dept. of Engineering-Economic
Systems, Stanford University, Stanford, Calif.
Owen, D. L. 1984. The use of influence diagrams in structuring complex decision problems. In The principles and
applications of decision analysis, vol. I, eds. R. Howard and J. Matheson. Menlo Park, Calif.: Strategic
Decision Group.
Park, K. S., S. H. Kim, and J. K. Kim. 1993. An artificial neural network approach to the decision class analysis.
In 93 Korean/Japan joint conference on expert systems, Seoul, Feb. 2 ± 5, pp. 717±729.
Reed, J. 1989. Building decision models that modify decision systems. Paper presented at the Thirteenth Annual
Symposium on Computer Applications in Medical Care (SCAMC).
Rege, A., and A. M. Agogino. 1988. Topological framework for representing and solving probabilistic inference
problems in expert systems. IEEE Transactions on Systems, Man, and Cybernetics 18(3):402±414.
Russel, S., S. Sampath, and A. Agogino. 1988. Creating influence diagrams from examples. Berkeley Expert
Systems Laboratory, University of California, Berkeley.
Shachter, R. D. 1986. Evaluating influence diagrams. Operations Research 34(6):871±882.
Shachter, R. D. 1988. Probabilistic inference and influence diagram. Operations Research 36(4):589±604.
Smith, G. F. 1988. Towards a heuristic theory of problem structuring. Management Science 34(12):1489±1506.
Smith, G. F. 1993. Defining real world problems: A conceptual language. IEEE Transactions on Systems, Man,
and Cybernetics 23(5):1220±1234.
Knowledge Acquisition for Knowledge-Based Decision Systems 149
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4
PAGE 150 BLANK
Dow
nloa
ded
by [
Uni
vers
ity o
f C
onne
ctic
ut]
at 0
6:40
08
Oct
ober
201
4