Role of Computing in Urban Travel Forecasting:

How Transportation Planning Practice Shaped Software

and Software Impacted Transportation Planning Practice

-------------------------David Boyce

Northwestern Universityat the

University of DelawareMarch 21, 2008

• In 1956, Beckmann, McGuire and Winsten published their seminal formulation and analysis of integrated models of network equilibrium and network efficiency.

• At the same time, transportation engineers and planners were seeking to solve the same travel forecasting problem in order to evaluate future road and transit plans.

• This seminar examines Beckmann’s model formulation and subsequent developments, and their relation to solving computationally the Sequential Procedure in practice.

Outline of the Seminar• Origins of Travel Forecasting in Research• Origins of the Sequential Procedure in Practice• Software Development and Travel Forecasting

Practice • Remarks on the interplay of research, practice

and software development

Context of Transportation Research in the Early 1950s

• Studies during and after World War II, including research of Koopmans and Dantzig

• First symposium on linear and nonlinear programming in 1949 led by Koopmans

• “Nonlinear Programming” by Kuhn and Tucker (1951)

• Wardrop’s “two alternative criteria based on these journey times” for determining the distribution of flows on routes (1952)

Transportation Research at theCowles Commission, 1951-54

• Rand project on Theory of Resources Allocation, with applications to transportation, location and dispersal problems, motivated partly by civil defense concerns and a desire to estimate the capacity of the USSR railways.

• In 1952, Beckmann was constructing a model of “flows of transport to a given network of roads or railroads connecting a finite number of locations at which supply and demand is concentrated. …this introduces the new problem of congestion.”

• Beckmann, with McGuire, and Winsten, drafted a manuscript on Highway Transportation containing a new formulation based upon the Kuhn-Tucker optimality conditions; a parallel effort examined the efficiency of Railroad Transportation;

• William Vickrey reviewed the book manuscript, and made extensive suggestions for revisions.

• Studies of Economics of Transportation was issued by the Rand Corporation in May 1955, and published by Yale University Press and Cambridge University Press in early 1956.

• Beckmann’s formulation of the user-equilibrium problem with variable demand was expressed as follows:

Origin-destination flows: ( )pqpqpq tgd =

Link flows for pq: pqijf with ∑=

∈C),( qp

pqijij ff

( )( )

( )( )

( ) ⎥⎦

⎤⎢⎣

⎡∑ ∫−∑ ∫∈∈

−

AC ji

f

ijqp

d

pq

ijpq

dsstdssg, 0, 0

1maxfd,

N

P

BA

BA

BA

∈∀=∑−∑

∈∀−=∑−∑

∈∀+=∑−∑

∈∈

∈∈

∈∈

iff

Qqdff

pdff

ii

qq

pp

h

pqhi

j

pqij

pqh

pqhq

j

pqqj

pqh

pqhp

j

pqpj

,0

,

,

By applying the Kuhn-Tucker theorem, Beckmann demonstrated that maximizing an artificial function, subject to node conservation of flow and non-negativity constraints on d and f, results in the following optimality conditions:

• all used routes from p to q have equal travel times, and no unused route has a lower time;

• origin-destination flow from p to q is a function of this equilibrium travel time.

• They also studied efficiency by minimizing total travel time subject to the same constraints.Surviving documents show these results were obtained in a short period during 1953-1954.

Contributions of Beckmann et alDiscussion Paper 1953

Completed equilibrium and efficiency formulations 1954

1951Project initiation

1955Publication of SET by Rand

1959Publication of Studies in Spanish

1967Beckmann paper, Traffic Quarterly

Recognition at Montréal Symposium 1974

Publication of SET by Yale University Press 1956

1952Working Papers

Origins of the Sequential Procedure• Traffic Assignment – “the estimated allocation of

traffic to a proposed facility” (Campbell, 1952)• Diversion curve vs. all-or-nothing assignment• No concept of a model of route choice; instead a

procedure for loading OD flows onto a road network, somehow involving shortest routes.

• No references to the principles of Beckmann et al, or the criteria of Wardrop, in the US literature.

• Even so, some usable results were obtained using iterative procedures on early mainframe computers.

• Descriptions of the elements of travel forecasting methods date from Voorhees (1955), Carroll and Creighton (1956) and Carroll and Bevis (1957).

• Irwin and Von Cube (1962) described a “sequence of operations,” which included “feedback” from Traffic Assignment to Trip Distribution.

• British travel forecasting followed a similar approach, but with a more solid foundation for trip distribution models pioneered by Wilson (1967).

Carroll and Bevis (1957) Predicting local travel in urban regions, Papers, Regional Science Ass’n.

Irwin and Von Cube (1962) Capacity Restraint in Multi-Travel Mode Assignment Programs, Bulletin 347, Highway Research Board.

Distribution / Mode Choice / Assignment

Trip Generation

Trip Distribution

Mode Choice

Traffic Assignment

Trip Generation

Sequential Procedure

feedback

Integrated Model

Consistent levels of service with a precise

user-equilibrium solution

Functional Forms

Trip Distribution

Mode Choice

Trip Assignment

ijCjjiiij eDBOAT ⋅−= β

∑∈

⋅−

⋅−

=

nm

u

um

ij mij

mij

eeP

'

'γ

γ

0 if 0

0 if 0*

**

⎪⎩

⎪⎨⎧

=≥

>=−

r

rijr h

hCC

Gravity Model w/ exponential

Logit Model

(MNL or Nested)

Deterministic User-Equilibrium

(Wardrop)

Role of Computers in the Early Studies• Initially, urban transportation studies were performed

with electro-mechanical accounting machines using punched cards, even before the first mainframe computers were available for civilian use.

• In the late 1950s, computer programs began to be written for the IBM 704 mainframe computer and its successors by the Bureau of Public Roads, the Chicago Area Transportation Study and others.

• The principal models implemented were trip distribution and traffic assignment.

IBM 407 Accounting Machine

IBM 704 Computer

Mainframe Used in Travel Forecasting

• IBM– 1st generation, 1940-56: vacuum tubes, 704, 709– 2nd generation, 1956-63: transistor, 7090, 7094– 3rd generation, 1964-71, integ. circuit, S/360, S/370

• Control Data Corporation– CDC 3600 (1963)– CDC 6600 (1964) – the first Cray supercomputer– CDC CYBER – 1970s – academic research

• Role of the service bureaus

Comparison of Computers Over Time• How can we compare the performance of computers

used in the early days with those we use today?• Computer scientists interested in this question have

performed tests during the past 35 years. Perhaps the most noteworthy is Professor Jack Dongerra, University of Tennessee, Knoxville. Despite his ongoing efforts, comparisons are imperfect.

• One metric for comparisons is “flops” on a standard problem; one difficulty seems to be that the problem used earlier is no longer demanding enough.

• Memory should also be compared: the available memory in early mainframes was roughly 256 KB.

Performance of Mainframes and PCs, 1951 - 2007

0.001

1

1000

1000000

1950 1960 1970 1980 1990 2000 2010

Year of Initial Release

Floa

ting

Poi

nt O

pera

tions

per

Sec

ond

(milli

ons)

Dongerra, Performance of Various Computers (2007)

Strohmaier et al, Recent Trends in the Marketplace,Parallel Computing (2005)

1 KFlop/s

1 MFlop/s

1 GFlop/s

IBM 360/195

IBM 370/158IBM 7090

Intel Core 2 Q6600Cray X-MPCray Y-MP/8

Cray X-MP/14se

IBM PC w/8088

CDC CYBER 170-750

Sun SPARCstation 2

UNIVAC 1

Programming 1st and 2nd

Generation Computers, 1958 - 1967• Bureau of Public Roads (BPR)

– Trip distribution - gravity model (1963)– Assignment - begun at GE in 1958 (1964)

• CATS, PJTS, others had their own programs.• Traffic Research Corporation, Toronto (1962)• Alan Voorhees and Associates (AMV) added to the

BPR battery and built a new battery for public transit systems planning (1967).

• Control Data Corporation began work on TRANPLAN in the early 1960s for the CDC 3600.

Programming 3rd Generation Computers, 1968 - 1980

• BPR moved its battery for trip distribution and traffic assignment to IBM S/360 (1969);

• AMV converted HUD mode split and transit assignment to IBM S/360;

• Ontario DOT and AMV built a new highway package, called TRIPS;

• Robert Dial at UMTA, with FHWA, led development of a new package, the Urban Transportation Planning System (UTPS) through the 1970s.

Programming Computers in the UK• London Transportation Study performed by a team

of consultants during the 1960s;• UK Ministry of Transport team led by A. G. Wilson

built an innovative model system for the SELNEC Transportation Study, using a CDC 6600 during 1964-1972;

• AMV formed a UK subsidiary, later Martin, Voorhees and Associates (MVA), that brought TRIPS to the UK in 1968;

• Elsewhere, including New Zealand, consultants began developing systems for a mainframes.

Transition to Minis and Micros• By the late 1970s, computing technology was

changing from mainframes to smaller and cheaper minis and micros;

• Consultants and planning agencies began adapting mainframe software or building entirely new systems for minicomputers;

• University research groups (MIT, Montreal, Leeds) began developing research-based software;

• US DOT was persuaded that it should leave the software development task to the private sector.

Contributions of University Research – 1• Despite pioneering contributions of Beckmann et al.

(1956), little university research occurred in the next ten years; moreover, their book was unknown, while urban transportation studies gained in status;

• Then, Brynooghe et al. (1969) and Dafermos and Sparrow (1969) reported on solution algorithms and the properties of the user-equilibrium problem;

• LeBlanc (1973) and Nguyen (1974) proposed convergent solution methods for the fixed demand traffic assignment problem;

• Wilson, Williams, McFadden and Ben-Akiva offered novel approaches to travel modeling based on entropy and utility theories.

Contributions of University Research – 2• Beckmann’s concept of a single model of demand

and route choice was forgotten as practitioners focused on the steps of the Sequential Procedure.

• The notion of combining those steps into a single formulation occurred to Murchland (1966, 1970).

• Evans (1973) and Florian et al. (1975, 1978) explored combined formulations, leading them back to the original model of Beckmann.

• An initial test of competing algorithms led to the research by Boyce and LeBlanc during the 1980s; Safwat, Horowitz and others also contributed.

• More recently, multiclass models were implemented and validated by Lam and Huang (1992), de Cea et al. (2001) and Boyce and Bar-Gera (2003, 2004).

Evolution of Practice-oriented Software• Following CDC’s withdrawal from TRANPLAN in

1970, Fennessy continued its development, sales and support through the 1970-1995 period;

• TRIPS, and later MicroTRIPS, evolved in the UK into a well-developed and widely used software product;

• MINUTP, and later TP+, was developed by Seiders and Siu for the IBM-compatible PC with DOS, and later for Windows;

• These systems, developed by and for practitioners for use with the Sequential Procedure, had little connection to research, but gradually absorbed research formulations and solution methods.

Initiation of Research-based Software• As academic research began offering more rigorous

methods, software development was initiated also;• EMME/2 was launched by INRO in the early 1980s

based on University of Montreal research;• Caliper built TransCAD based on a wide variety of

research advances in travel forecasting and GIS;• PTV AG offered VISUM based on research findings

at German universities in the early 1980s;• Van Vliet at University of Leeds in the UK offered

SATURN, a more specialized system for traffic assignment with several innovative aspects.

Synthesis of Research with Practice• Although EMME/2, TransCAD, VISUM and SATURN

attracted interest from academics and advanced practitioners in various international markets, sales of software did not really take off until the options traditionally used by practitioners were added;

• Moreover, PC hardware needed to advance sufficiently in speed and memory to allow PCs to replace the aging mainframe software systems;

• Finally, early DOS-based systems were updated to MS-Windows in order to facilitate their productive use by a wide range of practitioners.

Travel Forecasting Software Today• Citilabs was established in 2001 consolidating

TRANPLAN, MINUTP, TP+ and TRIPS under one vendor with a unified product, CUBE;

• INRO launched a major improvement in 2007, Emme, providing a GIS-type user interface;

• TransCAD and SATURN have implemented new assignment algorithms, moving beyond the Frank-Wolfe method first applied in 1975;

• VISUM adapted its German-based approach to the US market, including recent findings on solving the Sequential Procedure with feedback.

What can we expect from vendors in advancing travel forecasting software?

• Vendors compete by gradually improving their products, with considerable risk that practitioners may not appreciate new methods;

• Vendors do listen to their customers, but often they do receive mixed signals;

• Thus, improvements come slowly, often focusing more on visualization and ease of use than on advancing model capabilities;

• Some vendors have a limited ability to develop and evaluate new methods.

What can we expect from practitioners inadvancing travel forecasting software?

• Practitioners’ understanding of their software ranges from highly advanced to almost nil;

• Some view software as a black box that does not require understanding of its properties for its use;

• Lacking a standard approach or paradigm, travel forecasting practice varies widely by agencies and consultants, with each local practice varying greatly from others;

• Practitioners’ abilities to evaluate and choose among competing products is quite variable, leading to poor choices;

• Practitioners are sometimes unwilling to invest in training in application of their selected software;

• Others, especially consultants, tend to focus on one model (or step) with inadequate attention to the entire solution procedure;

• Competition for consulting contracts, desire to innovate or to be a leading practitioner may result in poor development choices.

What can we expect from researchers in advancing travel forecasting software?

• Researchers necessarily work beyond the state of practice and may oversimplify the needs of practitioners;

• As a result, newly developed methods may be either inappropriate or not accepted in practice, leading to frustration and breakdowns in communication;

• Some researchers would benefit from increased interaction with practice, with the likely result that practitioners would benefit as well.

What can we expect from universities inadvancing travel forecasting software?

• Graduate programs ought to emphasize principles and properties of models, not training on specific software;

• At the same time, experience with applying a system in a practicum or laboratory may be highly beneficial;

• Without a suitable textbook or handbook on travel forecasting principles and practice in the US, instruction is highly varied and may lag well behind the state of the art, and even the state of practice.

What is to be done?• Should a standard travel forecasting procedure be

adopted and implemented? If so, by whom?• Should vendors be required to support basic

conventions, such as estimation, convergence, consistency of inputs and outputs, etc.? By whom?

• Should academic programs offer a standard course on travel forecasting to prepare students for entry-level practice? Who would design the course, decide the content of the text, and certify the instructors qualified to teach such a course?

• A recent TRB report, Metropolitan Travel Forecasting, addresses some of these questions. Interested persons should peruse: http://onlinepubs.trb.org/onlinepubs/sr/sr288.pdf