Bicriterion Traffic Assignment Basic Theory and Elementary Algorithms

Bicriterion Traffic Assignment: Basic Theory and Elementary Algorithms
Author(s): ROBERT B. DIAL

Source: Transportation Science, Vol. 30, No. 2 (May 1996), pp. 93-111
Published by: INFORMS
Stable URL: http://www.jstor.org/stable/25768710
Accessed: 23-02-2018 07:23 UTC
REFERENCES
Linked references are available on JSTOR for this article:
http://www.jstor.org/stable/25768710?seq=1&cid=pdf-reference#references_tab_contents
You may need to log in to JSTOR to access the linked references.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms
INFORMS is collaborating with JSTOR to digitize, preserve and extend access to

Transportation Science
This content downloaded from 14.139.128.17 on Fri, 23 Feb 2018 07:23:07 UTC
All use subject to http://about.jstor.org/terms
Bicriterion Traffic Assignment: Basic
Theory and Elementary Algorithms1
ROBERT B. DIAL
Volpe National Transportation Systems Center (DTS-49), Kendall Square, Cambridge, Massachusetts 02142
This paper describes a bicriterion equilibrium traffic assignment model that accurately fore
casts path choices and consequent total arc flows for a stochastically diverse set of trips. Called
T2, its develops around a linear generalized cost model, which generalizes classical traffic
assignment by relaxing the value-of-time parameter from a constant to a random variable with
an arbitrary probability distribution. For the case where arc time and lor cost are flow depen
dent, this paper formulates conditions and algorithms for stochastic bicriterion user-optimal
equilibrium arc flows, which reflect every trip's exclusive use of a path that minimizes its
particular perception of generalized cost.
X he old saw "time is money" connotes the very instead of the usual simpler formulation, which as
raison d'etre of transportation systems. Every trans sumes all user have the same VOT. His reward was
portation plan evaluation includes a user's value of an amazing improvement in goodness-of-fit (BEN
time (VOT), every modern mode-choice model hosts AKIVA, 1994).
a VOT parameter, and every congestion pricing This paper proposes a similar remedy to the same
(Roth, 1967) scheme wagers that some travelers deficiency in classical traffic assignment. It presents
take slower routes to avoid tolls while others choose the basic theory and elementary algorithms of a
toll roads to save time. traffic assignment model that admits VOT distribu
Furthermore, it is a fact that each traveler has a tions. Called T2, the model is a bicriterion user
different VOT, depending on how much money or optimal equilibrium traffic assignment model,
which generalizes classical traffic assignment, by
time he is willing or able to spend on a particular
relaxing the VOT parameter in the generalized-cost
trip. Conventional transportation planning models, function from a constant to a random variable with
however, acknowledge this fact poorly. Instead of
an arbitrary probability density function (PDF).
using a realistic distribution of the VOT, these mod A traffic assignment model that is more robust
els use an average VOT. As a consequence, they than its conventional ancestor, T2's potential use
invariably produce large estimation errors and inac spans a wide spectrum of current and difficult prob
curate forecasts.
lems. These include mode-route choice, parking pol
In a recent exception to this common practice, icy planning, and congestion pricing. T2 can model
Professor Moshe Ben Akiva of M.I.T. estimated his mode choice by assigning trips to paths in a multi
logit mode choice model's parameters by assuming a modal ("hyper") network, which combines walking,
distribution f( ) of the VOT a G si. That is, letting a riding, transit and highway links. It can selectively
be the VOT and m a choice option, he fit the expected route auto trips to parking lots that are cheap but
value: require a long walk and to others that are closer to
work but expensive. It is useful model for determin
ing where to place tolls and what prices to levy in
order to reduce congestion.
Related Work
T2's roots are in QUANDT (1967) and later SCHNEI
Accepted by Mark S. Daskin. DER (1968); however, this paper develops the model
93
Transportation Science 0041-1655/96/3002-0093 $01.25
Vol. 30, No. 2, May 1996 ? 1996 Institute for Operations Research and the Management Sciences
94 / R. B. DIAL
very differently. Its mathematical programming ap 100

proach adds generality regarding the value-of-time
probability density function (PDF) and formulates
and solves the bicriterion user-optimal traffic equi 90 -j ^helicopter
librium problem. It extends considerably the short
note by DIAL (1979), which addressed only path find 80
ing. DAFERMOS (1981) addresses T2's equilibrium
problem; however, this paper makes no direct use of
70 H ^hot-air balloon
her results. Besides these earlier papers, this paper
builds on the variational inequality work of MAG gondola
NANTI and PERAKIS (1993) presented at the 1994 60 taxi
CD
ORSA/TIMS conference in Boston, which inspired
algorithms T2-ETA and T1.5-ETA. C ^dial-a-ride icksha
0 50 H
The insightful work of LEURENT (1994) merits spe
.train
cial comment. Leurent's formulation is at once more
and less general than T2's. He deals with elastic
40 H &.express bus/park'n ride
demand, while T2 assumes a fixed total trip matrix;
however, Leurent only permits one criterion, i.e.,
^car/tollroad express bus/walk'n ride
time, to be flow dependent, while T2 permits both. 30
Moreover, his Monte Carlo solution approach
ocal bus
(SHEFFI and DAGANZO, 1977) brings sampling er
rors, greatly restricts the VOT PDF and, compared 20 H
\ ^car/free road
to T2's algorithm, converges glacially (SLAVTN, ^covered wagon
1993). 10
bicycle
Overview of the T2 Model gp(0A3) = Cp + 0.43V .walk
-4*
Combined traffic assignment and mode choice 0 30 60 90 120 150 180 210 240 270 300
serves as a concrete example to introduce T2. Con TIME
sider a trip from origin o to destination d. We wish to
Fig. 1. Feasible paths and #^(0.43).
estimate the mode choice for this trip. Assume for
the moment it is possible to enumerate all feasible
paths for this trip and to know the time and cost of Path-Choice Probability
each. Now plot each path at a point in a graph
Answering this first question requires knowing
according to its time and cost.
how the trip decides among its feasible paths. Pre
Figure 1 plots these time-cost points for fifteen tend for the moment that arc times and costs are not
feasible paths?ignore the dashed line for now. The
flow dependent. Assume that a trip chooses a path p
horizontal axis is path time; the vertical axis is path
cost. Each path label suggests its so-called "mode." that minimizes its perceived generalized costgâ),
where:
Helicopter is the fastest and most expensive. Walk
ing is the slowest and cheapest. Gondola is very gp(a) =cp + atp;
cp = out-of-pocket ("dollar") cost of path p;
expensive and very slow.
This example begs three general questions: tp =time of path p;
a E [0, ??) = value-of-time, a random variable.
1. path-choice probability: what are the odds The probability of a trip choosing a certain path
that a trip will choose each of these paths? depends on that path's time and cost compared to all
2. path-finding algorithm: how can a computer others'?together with the trip's particular VOT a.
find these paths without enumerating all zillion Referring again to Figure 1, the dashed line rep
feasible paths? resents the objective function, which a trip with a
3. traffic equilibrium: how do we model the fact VOT of $0.43/min aims to minimize. The slope of
that path times depend on path choices, and vice this line is negative a. The generalized cost#p(a) of
versa? path p is the line's intercept with the cost axis.
Therefore, for a given a, the best path is the first
We introduce these questions now and answer them point that this line touches as it translates from left
later in greater detail. to right. Figure 1 shows path 4, "car/tollroad," is a
BICRITERION TRAFFIC ASSIGNMENT / 95
0 .25 .50 .75 1.00 1.25 1.50 1.75 2.00

a: VALUE OF TIME
(b)
1 I 1 M ' 1 I 1 1 I 1 1 I 1 1 I 1 1 I 1 1 I I 1 ' I Fig. 3. Probability of using path 4. (a) EF in neighborhood of
0 30 60 90 120 150 180 210 240 270 300 path 4. (b) VOT PDF: share on path 4.
TIME
Fig. 2. Efficient frontier, LikelyPath probabilities.
4 is the probability that the random variable a is
between these values.
For example, if the probability density of a for
trip's best choice if a = 0.43. If the dashed line had
trips going from o to d is fod( ), then as Figure 3
a different slope, then path 4 might not be the best
shows, the probability of using path 4 is 0.63. In
path. general, because the generalized cost is a linear
Refer now to Figure 2. Whatever the slope a, the function of a, the probability of using a path off the
only paths with any chance of selection are 1 EF is zero, while for a path p on the EF:
through 6?those shown delimiting line segments.
Among all feasible paths, only these six have a gen
eralized cost (GC) that could be the minimand of any
Prob[p]= \ hfJia)&a9 (1)
linear function of a. The line segments that connect KP
such path points, in decreasing sequence by path
time, form the efficient frontier (EF). Points where where ap is the slope of the line segment connecting
the EF changes direction are called extreme points. p to its right neighbor on the EF, and bp is the slope
As we shall see momentarily, in the case of a con of the segment connecting p to its left neighbor. Note
tinuous VOT PDF, only paths forming extreme that ap = 0 for the cheapest path, and 6p = 00 for the
points represent rational path choices. fastest path.
Thus, for fixed arc times and costs, Figure 2 re Any path that might be used we call likely; other
veals how to compute the probability of using a path. paths we call unlikely. Technically, likely paths have
For example, consider path 4, the toll road in Figure a nonzero probability of use; unlikely paths have
3(a), detailing the neighborhood of path 4 in Figure probability zero. A likely path is always on the EF
2. Assume that the line segment connecting path 5 but being on the EF does not guarantee a path is
and path 4 has a slope of -2.8, and the line segment likely. For a continuous PDF, a likely path is always
connecting path 4 and path 3 has slope ?1.2. Then, an extreme point on the EF, and a path not having
path 4 will minimize any GO whose VOT is between distinct bounding a's is an unlikely path?whether
1.2 and 2.8. Hence, the probability of selecting path on the EF or not: for Prob[p] to be greater than
96 / R. B. DIAL
zero, bp must be strictly greater than ap. The table

inset in Figure 2 lists the probabilities for Figure 2's 01000 ^ 1000
-30-?- ^
30 -0
likely paths; all other, unlikely paths have probabil
ity zero. 30 1000 55 1300 30 1000
In the case of a discrete PDF, other paths on the
EF are equally likely, and entropy considerations >k 1300 As 1300
(KAPUR and KOSAVAN, 1992) dictate that any path
with a minimum GC merits an equal share of trips;
?-io^P-gT
we ignore this concern here; the "all-shortest-paths" 30 1000 55 1300 30 1000
algorithm in DIAL (1971) handles this condition.
Furthermore, as will be seen later, if path time is a 1000
function of arc flow, then these shares are not in 30 30
general equal.
(a)
Path-Finding Algorithm
By considering likely paths only, we reduce an
otherwise daunting challenge to a practical chore: of 0> 10.00 ^ 10.00 ^
all the paths connecting o to d, we need to find only
those few that shape the EF. Presuming that a 10.00 $5 5.45 10.00
path's cost and time are the sums of the cost and
time of the path's arcs, then the GC of a path is just
the sum of the GC of its arcs. Consequently, finding
these likely paths can be easy. 5
$5 45 kIHf??
Each likely path p on the EF minimizes gp(a) for 10.00 $5 5.45 10.00
some a. Let the GC of an arc e be the sum of its cost
plus a times time:ge = ce + ate. If we label each arc
with its ge for a fixed a and find the minimum GC <$>
path, that min-GC path will be on the EF. Thus, to
find the EF is equivalent to finding min-GC paths (b)
for an appropriate set of a's. Fig. 4. Arc attributes at free-flow volumes, (a) Capacity and
Remarkably, the computational complexity of an free-flow speed, (b) Free-flow time and cost.
algorithm that finds all likely paths, is the same as
one that finds a single min-GC path. This latter
algorithm is virtually a standard min-path rou
tine?its only complication is tie-breaking?to guar represents these differences in a by user-supplied
antee extreme-point paths on the EF. (Details for all probability distributions. T2's generalization of
algorithms appear in Appendix A.) Wardrop's principle (1952), states that each trip
uses only paths that minimize their particular per
Traffic Equilibrium ceived GC. A simple example makes this clear.
A traffic assignment model accumulates trips for
all o-d pairs on each arc comprising their connecting
path. When all trips use a min-GC path, we call the
result a user-optimal traffic assignment. When arc 140 -i
times are fixed, a single min-path (a.k.a., "all-or
nothing") assignment performs this calculation by 100 H
*(*) = f(i-o
assigning each trip to its min-GC path. However, arc
times may change, depending on how many trips use
the arc?which in turn depends on arc times. The
solution to this circular problem has the name traffic
equilibrium.
All classical assignment models permit only a sin
IIIIIIIIIIIIIIIIIIIIIIIIIII1II
gle, fixed VOT a: they presume the perceived GC of 0 1000 2000 3000
a path is to all trips identical. In contrast, T2 as
sumes different trips perceive the same path as hav FLOW
ing different GCs by virtue of different VOTs. It Fig. 5. Arc time vs arc flow (BPR).
1.0 2000 881

0.6-i r2400 34.00 10.90 <D
0.5
0.4 - 34.00 2000 5.90 1119 10.90 881
P1600 J
/19(a) 0.3 2.0 I 1119 1119
0.2 ih 800 <3> 5.90 5.90
0.1 -
10.90 881 5.90 1119 34.00 2000
0.0 -
l I I I | I I I I I ' I I I I I f I I I I I]
0.0 0.5 1.0 1.5 2.0 881 , \ 2000
(z> 10.90 ^ 34.00 <8)
a ($/min)
Fig. 6. Value-of-time probability density: /19().
(a)
413
$1.09
$3.40 387 $5.591 $1.09 413

Figures 5-7 show an example of T2 equilibrium in
a nine-node grid network, a single o-d pair, and a
simplistic three-spike VOT PDF. The network is in Y "$5.59 " ^ 15-.59 "
Figure 4. All arcs are two-way and of length (d) one I
mile. The eight peripheral arcs are free roads; the $1.09 387 $5.591 $3.40 413
four interior arcs are toll roads costing $5.00. Figure
387 387
5 shows the time-flow (BPR) curves of the network's
$1.09-<?)
^ $3.40 -0
two arc types, toll arcs having higher practical ca
pacity (k) and speed limit (s) than free arcs. (b)
Figure 6 is the VOT PDF, consisting of three 468
spikes, at $0.10, $1.00, and $2.00 per minute and ??r222?<2>
$34.00 $10.90
corresponding volumes of 800, 2400, and 800 trips.
All trips go from node 1 to node 9. Figure 7(a) shows $34.00 1077 $10.90 855 $10.90 468
the associated T2 equilibrium flow. It labels each arc
781 ,
<D $10.90 ^
583
with an integer representing the total number of
$10.90
trips on the arc and a decimal number indicating the
arc traversal time. For example, arc (1, 2) has 2000 $10.90 494 $10.90 657 $34.00 1249
trips and requires 34.00 mins.
Figure 7(b) shows an example of the component of 494 1151
the total flow that represents trips with VOT of $10.90 w $34.00 <?)
$0.10/min. The dashed lines indicate unused arcs. (c)
The labels on the arc show trips and generalized
cost. Note that all used paths from node 1 to node 9 264
have the same generalized cost, which is less than $68.00 $21.81 ^
that of all unused paths: trips with VOT $0.10/min
$68.00 536 $16.81 264 $21.81
are in equilibrium. As seen in Figures 7(c) and (d),
I
the same is true for the trips with VOT $1.00 and 536 338
$2.00. $16.81 $16.81
Figures 7(bMd) sum to those in 7(a). Note how
ever that unlike the latter, these former flows are $21.81 $16.81 462 $68.00 338
not unique: any reallocation of individual flows
462
among their min-GC paths that summed to the
W(7>--<g
$21.81 ^ $68.00
same total in Figures 7(a) would also be in equilib
rium for every individual VOT classes. For example, (d)
if in Figure 7(b) 413 trips shift from path 1-2-3-6-9 to
Fig. 7. A T2 equilibrium traffic assignment, (a) T2-ETA total
path 1-4-7-8-9, while at the same time in Figure 7(c), arc flow and time, (b) Flow and GC for a = $0.10/ram. (c) Flow and
413 trips make the opposite shift; the result is an GC for a = $1.00/mm. (d) Flow and GC for a = $2.00/mm.
98 / R. B. DIAL
other equilibrium flow. The total arc flows would be work. As classical static traffic assignment provided
the same, but individual arc flows for VOTs $.01 and ground work for recent work in dynamic traffic as
$1.00 would differ greatly. signment (Le Blanc et al., 1992; Friesz et al.,
1993; jansen, 1994), the results of this paper, by
Remark: The VOT PDF their extension into the time domain, invite the de
Continuous vs Discrete. T2 theory and algo velopment of a dynamic T2. Although discussed no
rithms developed below permit each directed o-d further in this paper, this extension is natural and
pair can have it own PDF fod( ), which can be con its development urgent.
tinuous, discrete or any mixture of either. The inte
gral sign used in this paper always means Riemann Organization of this Paper
Stieltjes integration (APOSTOL, 1968). Except in the
The next two sections of this paper provide theo
simplistic case of a PDF consisting of only a few
retical justification and algorithmic detail for the
discrete points, T2's algorithms' complexity is unaf
fected. claims and procedures introduced above. Section
Estimation. The best way to estimate a particu covers T2 min-path traffic assignment, which incor
lar instance of the VOT PDF is a worthwhile re porates VOT as a random variable. It develops a
search topic. As a first cut, we could estimate it by model and algorithm for loading trips with a sto
chastic value of time on a network whose arc times
pairing likely paths with their observed probability
of use, and fit a cumulative distribution function and costs are fixed. Building on the concepts intro
(CDF). Fortunately, likely paths are a function of duced in prior sections, Section 1.4.2 develops math
the network only: they are independent of the dis ematical theory and algorithms for bicriterion equi
tribution of a. The path finding algorithm discussed librium traffic assignment for trips with an
later finds every likely path p on the EF and the arbitrary VOT PDF, where the network's arc costs
range [ap, bp) of a associated with each. and/or times vary with flow. The final section briefly
An observed value of the probability of using any addresses a few proposed T2 R&D topics, then ends
likely path p is an estimate of the CDF's differential the paper by acknowledging people who helped pro
duce it.
between ap and bp. That is, if Fod( ) is the (unknown)
cumulative distribution function of a, then
Prob[p] = Prob[ap < a < bp] = FJJbp) - F?,(ap). 1. T2 MIN-PATH ASSIGNMENT
Given a set of these observations, we may estimate given a set of paths connecting all origin-destina
Fod( ) with special statistical methods (SlLVERMAN, tion o-d node pairs and given a trip matrix, which
1986). To forecast Fod( ) for other locations in space specifies the number of trips between each o-d pair,
time, its estimate can be correlated to traveler char a traffic assignment algorithm simply accumulates
acteristics such as income. on each arc the trip volume of every path using that
arc. The output of a traffic assignment algorithm is
Remark: Dynamic Traffic Assignment this vector of accumulated arc flows, which we call a
The "bad news" is the model presented here is "traffic assignment." Depending on the paths the
static: T2 does not forecast arc-flow time profiles. trips use, this vector has many possible values. We
Static traffic assignment models cannot accurately call the set of all (feasible) traffic assignments for a
forecast congestion effects (i.e., travel time); there fixed trip table the ground set. Let
fore, they seriously compromise a planner's fore
casts. This shortcoming is particularly enfeebling 2ft+= {positive real numbers}
when planning tolls for congestion pricing, since? M = {nodes in the network}
besides making inaccurate link time estimates?the % = {arcs in the network}
static model cannot forecast trips shifting their vod = (given) number of trips going from o to d
travel times to avoid or reduce toll costs (BERNSTEIN, e =arc (ie,je), directed from node ie to node je
1993). =flow on arc e that originated at node o.
The "good news" is that the temporal dimension is Xe ~ ôeJV Xoe
literally orthogonal. All the theory and algorithms = total flow on arc e
presented in this paper support the development of x =(xoe) E ft+MxW = a "traffic assignment."
even more accurate dynamic models. For example,
all of the algorithms can be applied with minor mod In classical min-path traffic assignment, x is in the
ification to a discrete time-staged expanded net ground set iff it is a non-negative solution to the
following \Jt|2 equations in |^| \N\ unknowns (AHUJA 1.3. T2 Min-Path Traffic Assignment
ET AL., 1993):
Clearly, if arc costs and times are fixed then a T2
min-path traffic assignment is a flow that reflects
Vod = 2 xoe ~ E for all o, d G Jf. each trip taking its particular min-GC path, which
fee* | ./.=<*} {eG?|ic=d}
would minimize the total generalized trip cost (4):
1.1. T2 Ground Set X
Lemma 1.1. Flow y G is a T2 min-path traffic
T2 is analogous to the classical model; however, it assignment iff
introduces the random variable a, which partitions
arc and o-d flows by qualifying the variables and
augmenting the constraints given above as follows:2 y G arg min X (ce + ate)xe(a)da. (5)
si = {values of time (VOTs)} C 2ft+
vod(ct) = number of trips going from o to d with
VOTa Proof (=?: Because ce and te are fixed, the right
hand side of (4) is separable in a. A T2 min-path
xoe{ct) = flow on e of trips that originated at o with
VOTa traffic assignment certainly results in minimizing
each integrand. Hence the integral is minimized.
xe(a) =2oeJyr xoe(a) = total flow on arc e of trips
with VOT a. ?=): Proved by contradiction: if the integral is min
imized, so is each integrand; hence, each must be a
Thus for all e G %, o, d G >f, and a G i: min-path assignment, which contradicts the hypoth
esis.
Uod(a) = X *<*(<*) ~ S (2)
1.4. T2 Min-Path Assignment Algorithm
where as discussed in Section 1, Referring to Lemma 1 and recalling from Section
1.1 the definitions
Uod(a) = VoJUa)- (3)
y = (y(a)) = ((yja)))
Now letting
*(a) = (..(a)) G ?f?
x = (x(a)) G ?WXW*M
(an "infinite vector"), we now present a simple algorithm for finding (ye).
That is, the algorithm produces only total arc flows
the ground set for T2 becomes on each arc?not flows for every VOT. The following
two subsections are informal descriptions, which
3? = {x | equations (2) and (3) are satisfied}. summarize the formal definitions in Figures Al and
A2 of Appendix A. (Ignore for now all references to
1.2. Total Generalized Trip Cost variables u and u? and the function H( ). They con
In the case where arc costs and times are fixed, a cern T2 equilibrium traffic assignment, and Section
very useful statistic describing a traffic assignment 1.4.2 explains them.).
x G is its total generalized trip cost, which is the
sum of products of each trip times the GC of the path 1.4.1. Algorithm T2-MPA (Appendix A,
it uses. For now, let these costs and times be fixed at Figure A2)
c = (ce) and t = (te). For any traffic assignment x G Compared to the classical model, a min-path as
9P, we define
signment algorithm for T2 is more involved. Each
o-d pair may have several distinct min-GC paths
G(x \c,t)= 2 (c? + ate)xe(ct)da, (4) depending on the value of a. Each path must receive
ee.% its fair share of trips, and the sum of the flow on
each path must be the given vod. Algorithm T2-MPA
which is the total generalized trip cost of the x. executed for each o-d pair yields a T2 min-path
traffic assignment. It uses algorithm LikelyPaths to
find the paths to load.
2The infinitesimal notation vod(a)da, xoe(a)da, xe(a)da, etc.
might seem a more rigorous notation, but the visual burden To load the network with trips going from o to
exceeds its utility. d and having a VOT density fod( ), do the following:
100 / R. B. DIAL
1. [likely paths] Using Algorithm LikelyPaths, section moves on to flow-dependent arc costs and
find all likely paths from o to d. times. Accordingly, we augment the notation ce and
2. [trip assignment] For each likely path p: te to allow for functions as well as scalars:
2.1 [a-ranges] associate with each likely path p
its VOT range: ap ^ a < bp for which p ce(xe) = cost to use arc e when arc's total flow is xe
minimizes gp( a); te(xe) =time to traverse arc e when arc's total flow is
2.2 [path probability] compute the probability
of each p:
2.1. T2 Adaptation of Wardrop's Principle
(T2-WP)
Prob[p] = Prob[ap <?<&,]= fMda; Wardrop's principle (1952) extends readily to the
stochastic-a case:
2.3 [path share] multiply vod by Probfp] to ob In a bicriterion equilibrium traffic assignment,
tain p's share of trips, vp; where the VOT is a random variable a, no trip
2.4 [arc load] add vp to the flow xe of each arc e (with its particular a) has another path with a
on p. smaller GC, i.e., gp(ct) = cp + atp than the one
which it is using.
1*4.2. Algorithm LikelyPaths (Appendix A,
Figure A3) That is, a T2 equilibrium traffic assignment (T2
ETA) is the vector of arc flows x* E 9? reflecting a
The following procedure finds all likely paths and traffic assignment that puts every trip on a path
determines the exclusive range of a for which each having minimum generalized cost with respect to
path is optimal. that trip's particular VOT a. This generalization of
classical traffic equilibrium admits a large, usually
1. [cheapest path] Let a? = 0 and solve this infinite, number of categories of trips in simulta
min-GC path problem; this finds the cheapest neous equilibrium.
path?path 1 in Figure 2.
2. [fastest path] Let a? = oo and solve this min-GC
2.2. T2 Equilibrium Traffic Assignment
path problem; this finds the fastest path?path 6 Theorem
in Figure 2.
3. [between paths] Let a? be the slope of the line All the above conditions reduce to a simple fact:
connecting paths 1 and 6, and solve this min-GC
T2-WP implies that a T2 equilibrium traffic as
path problem; this finds path 4. Now consider the
signment has the same total generalized cost as a
two line segments connecting path 1 with 4, and
T2 min-path traffic assignment with the arc costs
path 4 with 6; their slopes find paths 5 and 3.
and times fixed at the level implied by the equi
Continue recursively until unable to find any librium flow.
more likely paths.
We restate this fact as the
Caveat. Algorithm T2-MPA is elementary, intend
ing only to demonstrate T2's solvability. For exam T2-ETA THEOREM. The flow x* E<%isaT2 equilib
ple, subroutine LikelyPaths() is a reasonably effi rium flow iff
cient path algorithm: its complexity is comparable to
a single min-path algorithm. However, its enumer G[x* | c*, i*] = G[y | c*, f*] (6)
ation of min paths renders Algorithm T2-MPA im
practical for solving real-world problems?popular where the fixed arc cost and time vectors are fixed at
claims of limitless free computer time not with c* = (ce(x*)) and t* = (te(x*)), and as before:
standing. A practical T2-MPA algorithm would not
enumerate paths. Such an algorithm is the subject
of a future paper.
2. T2 EQUILIBRIUM ASSIGNMENT
y E arg min G[x \ c*, ?*].
UP TO NOW, our technical discussion has addressed
only the simple case, where the arc costs and times Proof We first state and prove a lemma regarding
are fixed?i.e., do not depend on the arc flow. This equilibrium for trips of a single given VOT a? 6 i:
LEMMA 2.1. Define a 9?(a) as a "projection" ofX: proaching x* and in doing so, solves a series of T2
min-path traffic assignment problems?each solu
tion being called x.
= {feasible assignments of trips with VOT a]
Lemma 2.2. Ifx? e % and
y(a) E arg min G(x | c, ?).
xE?(a) *? =(*.(*;)), c?=(ce(x?))
Afou; consider a single VOT a? and assume that for x G arg min G[y | c?, t?] (8)
all the remaining a G i - {a?}, the flows x(a)
EX(a)'s are equilibrium flows for their respective
a's. Then the partial traffic assignment x(a?) E 9?(a?)
is in equilibrium iff w? = ax?(a)da, ue = I axe(ct)dct (9)
G[x(a?) | c*, i*] = G[y(a?) | c*, **]. (7)

Axe = xe - x?e, Aue = ue - u?e\
Proof. Obvious: By definition, this is just the clas then
sical single fixed-a case: at equilibrium, all trips
with the same o-d use a paths with same, minimum G[x | c?, t?] - G[x? | c?, *?]
generalized cost.
= X [ce(*3A*e + fe(*;)Ai*e].
We now prove Theorem T2-ETA. (=?: Follows
from Lemma 2.1. ?=): Follows from the fact that the
Proo/i
left-hand side of (7) can never be smaller than its
right-hand side. Assume (6) is true and (7) is false
for one or more a's; that is, one or more of the sums G[x | c?, f?] = f X [cM + ate(x?)]xe(a)da
on the right are less than their partner on the left.
Then the integral of the left-hand side of (7) has to
be greater than that on the right-hand side: a con
tradiction. -2
eG?
ce(x?)xe(ct)dct + afe(x2)xe(a)da
'a *^ a
COROLLARY 2.1. The T2-ETA Theorem and Lemma
2.1 imply that the flow x* E % is a T2-ETA, i.e. all
trips for all VOTs are simultaneously in equilibrium,
iff
eG?
ce(x?)j" *e(a)da + fc(x?)J axe(a)da
= S [c.(jc;)x? + a*?)"J- (10)
eG?
2 [c?(?t) + a^(*t)]*i(a)da
G[x | c?, f?] - G[x? I c?, f?]
= X [C?(*:)(*. - <) + - "e)]

2 l>e(*?) + ate(x*)]y (a)da.
= X [cc(*?)A*e + ?.(*3AmJ. (11)
eG?
2.3. T2 Equilibrium Traffic Assignment
Algorithm COROLLARY 2.2. The T2-ETA Theorem says when the
right-hand side of (11) vanishes, x? is a T2-ETA.
T2-ETA entails the simultaneous equilibration of
an infinite number of trip classes. Fortunately, only COROLLARY 2.3. T2-ETA conditions can be stated as
the total flow xe on each arc is interesting, since as a variational inequality: the flow x* E is T2-ETA
seen in the Introduction, the individual class arc iff
flows xe(a) are not unique. This permits the con
struction of an efficient solution algorithm. 2eE.%
[ce(x$(xe -x^ + te(xt)(ue - u*)] s 0.3 (12)
NOTATION. For purposes of symmetry, we effectively
rename x* as x? and y as x. More precisely, the 3Letting g( ) = (c( ), t{ )) and v = (Jc, u) gives (12) a more
T2-ETA algorithm generates a sequence of x?'s ap familiar VI look: g(v) Au > 0.
102 / R. B. DIAL
Proof. Define (for this proof only) x? as 1. [ascent direction] Fix all arc costs and times:
x?G argmin G[x | c*, **]. (13)

and perform a T2 min-path assignment to obtain:
That is, for all x G 9?,
x G arg min G[y | c?, t?]
X \.ce{xe)xe+ te(xe)ue]
-2 [c.(*5*:+ *.(*>;]*<). d4) A*. xe(a)da - x?e, Aue axe(a)da -
By the Corollary 2.2, the flow x* is T2-ETA iff 2. [termination test] Let:
2 [ct(x$x?t + tt(x*)u?\ ?[A, (xe), (ue)]
ee.%
-2 [c.(*:s)*:+a*t)?a = o. us) = X l>e(*e + kAxe)xe + ^(xj + AAxe)i*J.
Adding (14) and (15) yields If
X [ce(**)x.+ te(x*>Ue] L[0, (A*e), (AiQ]

L[0, (xe), (ue)]
- eG?
2 [c.(*5?:+*.(*5?3ao. (16) quit: (x?e) implies an (e-approximate) T2-ETA;
else go to Step 3.
Or, [step size:] Determine step size A*: Let
2 [c.(x*)(*. - x% + te(xt)(ue - ?*)] ^ 0. (17) if L[l, (A*e), (Am.)] ^ 0)

eG?
^* ^ arg L[A, (Axe)* (Awe)] = 0 otherwise
Ae(0,l)
Thus, the T2-ETA problem reduces to a variational
inequality problem in only two state variables per
arc, xe and ue. This permits its solution with existing 4. [improved solution] Update all x? and u?:
VIP algorithms. One such algorithm is MAGNANTI
and PERAKIS'S (1993) Generalized Frank-Wolfe algo
x?e ?- xe? + A*Axe, <- + A*Awe
rithm (GFW), which they have shown to converge if and return to Step 1.
the variation's Jacobian is symmetric. The following
elementary algorithm adapts GFW to T2. Although 2.4.1. The Calculation of ue
the Jacobian of (17) is not necessarily symmetric, all
The calculation of the integral ue (9) warrants
my tests tend to converge?if slowly. (Later, in Sec
comment. Recall that for each o-d pair, the likely
tion 2.5, we provide another algorithm, for a slightly
paths induce a partitioning on the range of a, asso
easier problem, which always converges.)
ciating with each likely path p the exclusive range of
2.4. Algorithm T2-ETA (Appendix A, a for which that p is a min-GC path. Because there
Figure A4) are a finite number of likely paths, this partitioning
permits the exact calculation of the ue integral as the
This algorithm starts with x? being a T2 min-path
sum of discrete quantities. Let:
assignment for c? = c(0) and t? = t(0), then modi
fies it to drive (11) towards 0.
fod(?) = VOT PDF for trips from o to d
0. [initialization] Do a T2 min-path assignment to vod = total trips from o to d
obtain: op = path p's origin node
dp = path p's destination node
x? G arg min G[x | c?, t?] <3>e = set of likely paths that use arc e
ap = lower limit of VOT for which path p is
optimal
xe^~ I *e(a:)da, w"<- I ctxe(a)da. bp = upper limit of VOT for which path p is
optimal
xep(a) = trips with VOT a assigned to path p E 2Pe

= {v0pdf0pdp(a) if ap < a < bp
0 otherwise.
When Algorithm T2-MPA assigns a path's flow to its

arcs, it knows these values. This permits an alter
native expression for ue:
ue = ctxe(a)da
= X ctxep(a)da
P ?e J ?
= I ? Vopd/opdpMÔL
?i?1?r
0.0 0.5 1.0
= 1 v0pdp 'P af0pdp(a)da. (18) /12(a) = 2a

Pê J ap
(b)
This last summand is the product of a known originFig. 8. Two-arc equilibrium example: input, (a) Two-Arc net
destination trip volume times the first moment of awork, (b) VOT PDF.
from ap to bp. Computing (ue) is, therefore, merely to
accumulate "link loads" of this summand. This ex
plains the reference to ue in the procedure load in this simple example with only two arcs and one
triangular
Path() and shows how to obviate storing any xe(a). PDF, there are easy formulae for the
decision variables:
2.4.2. Example 1: Algorithm T2-ETA
Figure 8 poses an example simple enough for xa = v12 fuMda = 10 2a da = 10a2,
manual calculation. The network has only two arcs: '0 ^ 0
the upper arc is free and has a time equal to its flow;
the lower arc has a fixed cost of 1 unit and has a time xh = 10 - x?
equal to twice its flow. Ten trips go from node 1 to
node 2; their VOT PDF is triangular between zero Too 2
and one. ua = xa I ocf12(a)da = x
' 0 J 0
Figures 9(a-c) show three iterations of Algorithm
T2-ETA applied to the input data in Figures 8(a and ub = 6.7 - ua.
b). Each iteration is described with four figures: The
leftmost figure shows the beginning state of the Furthermore, since c = (0, 1) and t = (xa, 2xb), the
algorithm: x?, u? and t(x?) (c is constant). The next solution to
figure is the efficient frontier implied by the arc
times and costs at the beginning state: the dot on the X [ce(x?e + A*Axe)Axe + te(x?e + A*Axe)Aae] = 0
eE.%
?-axis represents the upper arc and the dot on the
c-axis the lower arc. The slope of the line connecting is simply
these two dots give aQ, the threshold VOT at which
trips shift from the upper arc to the lower. The third xaAwa + Axh + 2Awfcx6
figure shows the PDF (not the CDF, hence its range
A* =
AaaAxa + 2Aw6Ax6
is [0, 2]) and the fraction of trips that the ascent
direction assigns to the upper arc (the remainder The subtitle shows the relative disequilibrium of the
goes to the lower arc.) The fourth figure is the ascent iteration's beginning state and the optimal fraction
direction, its arc flows x and moments u. Note that, A* for combining the beginning state with the ascent
104 / R. B. DIAL
/12(c)
xa = 0.1
0.0 0.5 1.0 xb - 9.9

L(0}x?,u?) = 66.7 EF 10/?-10/(<*)** = 0.10 L(0,x,Ti) = 10.0
(a)
< = 7.2 <a = 7.2
0.0 5 10 0.0 0.5 1.0 Xft = 5.5

x?b = 2.8 *6 = 5.7
L(0>x?>ti?) = 47.7 EF 10 J00 67 f(a)da = 4.53 L(0, x, u) = 46.3
x? = 7.0 ta = 7.0
1 I 1
x? = 3.0 tb = 6.0 0.0 5 10 0.0 0.5 1.0 x& = 0
-1.0
L(0,x?,u?) = 47.6 EF 10 Jo f(<*)d<* = 100 MO, = 4^.6
(c)
Fig. 9. Ex. 1: Three iterations of Algorithm T2-ETA. (a) Iteration 1: s = 5.69; A* = 0.287. (b) Iteration 2: s = 0.03; A* = 0.066. (c) Iteration
3: s = 0.02; A* = 0.053.
direction to obtain an improved beginning state for Similarly,

the next iteration.
For example, the first iteration starts with all 10
trips on the upper arc, implying arc times of 10 and
ua = ^ xaa30 = 0.0
0 (with fixed costs of 0 and 1) for the upper and
ub= 6.7 - ua = 6.7
lower arc respectively, and producing points (10,
0) and (0, 1) on the EF. A line connecting these 10(-6.7) + 9.9 + 2(6.7)(9.9)
points yields a threshold aQ = 0.1. Thus, the ascent
direction has
A (-6.7)(-9.9) + 2(6.7)(9.9)
By adding 0.287Ax to x? and 0.287AiZ to u?, we
xa = 10a2o = 0.1 obtain the beginning state of Iteration 2, whose rel
ative error of 0.03 is significantly lower than the
xb = 10 - xa = 9.9 5.69 of Iteration 1.
Table I shows the results of the first ten iterations
total trips on the upper and lower arc respectively. of this example in Figure 9. Equilibrium occurs at
TABLE I.
Ten Iterations of T2-ETA (Two-Arc Example)
x?a x?b u?a u% L(0,x?, u?) c?a c?b t? t%
n xa xb ua ub L(0, x, u) a s A*
10.0 0.0 6.7 0.0 66.7 0 1 10.0 0.0
0.1 9.9 0.0 6.7 10.0 0.10 5.69 0.287
7.2 2.8 4.8 1.9 47.7 0 1 7.2 5.7
4.5 5.5 2.0 4.6 46.3 0.67 0.03 0.066
7.0 3.0 4.6 2.1 47.6 0 1 7.0 6.0

10.0 0.0 6.7 0.0 46.6 1.00 0.02 0.053
7.1 2.9 4.7 2.0 47.6 0 1 7.1 5.7
4.8 5.2 2.2 4.4 46.4 0.69 0.03 0.070
7.0 3.0 4.5 2.2 47.5 0 1 7.0 6.0

10.0 0.0 6.7 0.0 46.6 1.00 0.02 0.050
7.1 2.9 4.6 2.0 47.6 0 1 7.1 5.7

5.1 4.9 2.4 4.2 46.5 0.71 0.02 0.077
7.0 3.0 4.5 2.2 47.5 0 1 7.0 6.0

10.0 0.0 6.7 0.0 46.5 1.00 0.02 0.048
7.1 2.9 4.6 2.1 47.5 0 1 7.1 5.8

5.4 4.6 2.6 4.0 46.6 0.73 0.02 0.087
7.0 3.0 4.4 2.3 47.4 0 1 7.0 6.1

10.0 0.0 6.7 0.0 46.5 1.00 0.02 0.047
10 7.1 2.9 4.5 2.2 47.4 0 1 7.1 5.8

5.6 4.4 2.8 3.8 46.6 0.75 0.02 0.102
xa = 7.06?which can be verified by noting that at for the skeptic an alternative algorithm, which
the corresponding arc times, arc a is used by all trips solves a simpler problem, but whose convergence is
with a < 0.84; and when a = 0.84, 10a2 = 7.06 =readily proved. This simpler problem allows only arc
xa = ta. Thus, the algorithm converges quickly to times to vary with flow; arc costs are fixed; further
within a few percent of the right answer, but thenmore, VOTs are strictly positive. Due to the major
begins to "tail." This is a common problem with the restriction on arc costs and minor constraint on a,
Frank-Wolfe algorithm. I have had the same expe the model is not quite bicriterion; so, we call it T1.5
rience running Algorithm T2-ETA on larger net equilibrium traffic assignment?or T1.5-ETA.
works. LEMMA 2.3. IfO<aGs& then x* is a T1.5-ETA iff
2.4.3. Example 2: Sampled Iterations of T2 for all xGSf,
ETA
To give a sense of Algorithm T2-ETA's behavior, [xe(a) -x*(a)]da. ^ 0. (19)
Figure 10 shows the results of a sampling of four
iterations of the algorithm applied to the example in
Proof Corollary 2.1 implies equilibrium occurs at
Figures 4 and 5. Notice the error in Figure 10(d):
x* iff
Iteration 10 is within about 4 percent of the precise
equilibrium flows in Figure 7(a). Due to the infa
mous "tailing problem/' to duplicate them would:+*?(*?) [ye(a)-x*(a)]da = 0. (20)
require several more iterations. Fortunately, for
most practical applications an approximate equilib
rium suffices. Where
2.5. T1.5 Equilibrium Traffic Assignment
Because the proof of convergence of Algorithm y G arg min X xe(a)da,
T2-ETA is outside the scope of this paper, we offer xESe J a eE%
106 / r. b. dial
i.e., for all x E 3200
<2) 167.29
3200
167.29 ifi7 oq vy ?
10.61 800 5.45 0 167.29 3200

+ *?(*?) (xe(a) - ye(a))da > 0.
e(=% 800 800
(21) 0> 5.57 5.57
0 5.45 0 394.00
Adding (20) and (21) proves the lemma.
0
T1.5-ETA Theorem (Leurent[16]). Flow x* is a 10.00 w 10.00
T1.5-ETA if 5 = 17.17 A* = 0.469
x* E arg min Q(x), (a)
1699 1699
22.51 22
where
2301 5.45 0 22.51 1699
Xe(a)
a da + t(v)dv
(22) 800 425
Jo
17.61 1501 5.46 375 40.54
Proof Suppose inequality (19) were the direc
tional derivative of Q( ) at x* in the direction x - x*;
v!> FTei ^ 2T57 ^
then if Q( ) were convex in %y (19) would be a nec
5 = 0.53 A* = 0.152
essary and sufficient for x* to be Q( )'s global mini
mand. Inequality (19) is Q( )'s specified derivative if (b)
(bazaraa et al., 1979) 1931 /-s 1352
30.84 15.01 0
dQ _ ce 37.51 2069 5.49 579 15.01 1352

J^?a) = ~a~+te{Xe)> 917
which is clearly the case:

0> 5.66
12.65 1153 5.60 845 34.09 2002

cx(a) \Xe
dxe(a) dxe(a)
da + ^(i;)di; 1153 1998
s = 0.08 A* = 0.038
CeXe(a) (c)
da\ ^ 2026 1243 x-n
dxe(a) 35.27 13.58
1974 5.56 783 13.58 1243

32.77
+ 909 750
t(v)dv
dxe(a) 5.65 5.55
1065 5.( 942 33.68
Ce d lXe \ dXe Ce
= ~ + ? t(v)dv ?= -+ te(xe). 1065
wn.93 ^ 34.33
2007
a dxe\) Jdxe(a) a 3 = 0.04 A* = 0.020
That Q( ) is convex in 9? is also clear. Its Hessian is (d)

zero everywhere except on its main diagonal, which Fig. 10. Ex. sampled iterations of Algorithm T2-ETA (flow and
has tfe(xe) as element (e, e). Assuming t( ) is nonde time): (a) Iteration 1; (b) Iteration 2; (c) Iteration 6; (d) Iteration
creasing and twice differentiable for (xe(a)) > 0, the 10.
Hessian is symmetric and positive semidefinite,
making Q( ) convex in % (bazaraa et al., 1979).
This completes the proof.
2.6. Algorithm T1.5-ETA (Appendix A5) (Frank and Wolfe, 1956; LeBlanc, 1973; Nguyen,
Since Q( ) is convex, we can find its minimand 1974). However, let us use an indirect approach that
using the well-know Frank-Wolfe (FW) algorithm reuses Algorithm T2-ETA. As before, rename x* as
x? and y as x, but now redefine ue as: Proof. Follows directly from Lemma 2.3 and the
T1.5-ETA Theorem. Since Algorithm T2.5-ETA uses
GFW in a case where GFW and FW are equivalent,
it behaves like FW.
3. SUMMARY AND CONCLUSIONS

In the manner of Lemma 2.2, we can easily prove
thatx?isaT1.5-ETAiff THIS PAPER COVERED the basic theory and elemen
tary algorithms for a bicriterion user-optimal equi
X [ceAue + te(x?)Axe] = 0. librium traffic assignment model dubbed T2. It de
fined T2's behavioral assumptions, derived its
Changing the definitions of the procedures L() and mathematical formulation, and designed its solution
H() in Section 2.4 (and Appendix A4 to those in A5), procedures. Specifically, it extended Wardrop's Prin
and changing (18) to read: ciple to include a stochastic value-of-time, produced
a simple equation that reflected an equilibrium
state, and presented four novel and space-efficient,
bpf0pdp(<x)
X v0pt a
da but slow, algorithms:
(i) a bicriterion parametric min-GC path algo
we can run this slightly revised Algorithm T2-ETArithm, LikelyPaths,
to perform in effect the appropriate FW. (ii) a bicriterion min-GC path assignment algo
Algorithm T1.5-ETA is an instance of what per rithm, T2-MPA,
(iii) a bicriterion equilibrium assignment proce
akis and magnanti (1993a, 1993b, 1994) call Gen
eralized Frank-Wolfe (GFW), which solves varia dure, T2-ETA, that solved problems where both
tional inequalities directly, obviating a formula for a arc time and cost are flow-dependent, and
function being minimized. If the Jacobian of the (iv) a one-and-a-half-criteria equilibrium procedure,
GFW functional (i.e., Hessian of the FW objective) isT1.5-ETA, that permits only time to depend on
symmetric, GFW is equivalent to FW. flow.
Note that T1.5-ETA's ascent direction is the neg T2-ETA is more robust than T1.5-ETA but lacks the
ative of FW's descent direction, and that its line
latter's proof of convergence. LikelyPaths and T2
search minimizes Q(x? + AAx) by finding a zero for
MPA provide support for either of these two equilib
its derivative for A E (0, 1):
rium assignment procedures, which produce esti
dQ a mates of the total arc flow accruing when each trip
uses its particular min-GC path.
3.1. Additional Research and Development

x?e(a) + A*Axe(a)
2 ce I -da Besides dynamic assignment and VOT PDF esti
mation, which were mentioned at the end of the
Introduction, several other problem areas come to
X?+\*bxe mind for T2-related additional research and devel
+ I t(v)dv opment.
o
Algorithm T2-ETA Convergence Proof. The con
vergence of Algorithm T1.5-ETA's was easy to prove;
Axe(a)
=e<=%2cP I ?-? da + t(x? + A*Axe)Ajc however, it is beyond this writer's ability to prove
that T2-ETA converges. Accordingly, I trust that a
later paper will prove it or some variant does?as my
= 2 [CeÛe + *(*; + A*A*e)A*J tests would indicate.
ee.%
Algorithm Speedup. Intended merely to show that
T2 makes sense and has an accessible solution, the
= L(A*) = 0.
algorithms presented here are "elementary," they
T1.5-ETA Convergence Theorem. Algorithm T1.5 are unsuitable for real-world applications. Accord
ETA, in Appendix A5, converges to an equilibrium ingly their performance received no discussion. The
traffic assignment (in perhaps an infinite number of complications attending more sophisticated algo
iterations). rithms appropriate for "productions code" would ob
108 / R. B. DIAL
scure this paper's aim. Speedups from sophisticated value-of-time a and a value-of-inconvenience /3. A
algorithms are the subject of a future paper. For path's GC becomes gp = cp 4- atp + ($dp, where dp
example, to find a feasible descent direction, Algo is pathp's inconvenience measure. For each o-d pair,
rithm T2-MPA enumerates likely paths, as opposed T3 uses a joint PDF fod(a, /3). The efficient frontier
to using an approach base on min-path trees. T2 can, is a triangulated convex surface, with each vertex
however, can be implemented with a parametric tree representing a path. The path-choice probability is
building algorithm (ORLIN, 1993). This improvement the double integral of fod(a, j3) over the values of a
not withstanding, the descent direction still con and j8 that define the vertex's subgradient. T2's ex
sumes more computer cycles than the simpler clas tension to handle more than two criteria is a fruitful
sical model; therefore, T2 implementation must subject for research.
eliminate steepest descent's tailing problem, proba
bly by replacing the GFW algorithm with simplicial 3.2. Applications
decomposition. The use of T2 and/or Tn to solve actual transpor
Upper Bounds on Arc Flows. If T2's model and tation planning problems presents many interesting
algorithm incorporated prespecified upper bounds potentialities for applications-oriented research and
on arc flows in the manner proposed by HEARN development. Four examples are the following:
(1980), it would greatly improve its unique suitabil
ity for analyzing congestion pricing and, most espe Road pricing is the most obvious application of
cially, parking policy. T2. To plan toll roads, T2 could be used with an
auto-driver trip table to determine traffic volume
Nonlinear Generalized Cost. Every interesting on arcs at various toll levels. It is the tool neces
result in this paper depends on the generalized cost sary to apply congestion pricing: using cost as a
of a path being a linear function of its cost and time.
flow-dependent variable, an arc's "toll" can be re
When GC is nonlinear, the GC of a path is no longer
lated to the marginal cost of an additional trip;
the sum of its arcs' GCs, making its min-path algo thus, users would pay an amount that reflected
rithm an interesting and worthy challenge.
not only the direct cost to themselves but also the
Elastic Demand. T2 ignores the fact that the total delay they cause other users.
trips vod between an o-d pair usually depend on the
Parking receives simplistic treatment in the tra
generalized cost of the trip, which in T2 is a random
ditional approach to transportation planning,
variable. Speaking theoretically, the inverse to this
which ignores the fact that auto trips begin and
demand function maps vod(a) into god(a); and the
end with a walk to and from a parking place. In
expected value of this inverse provides the expected
most cities, there are several different opportuni
generalized cost E[god \ vod] among all trips from o to
d. Hence, at equilibrium ties to park, each with a different cost and walk
time. The choice among these opportunities de
2 W.g.\x*\x*= 2 E[ge\x*]ye pends on the trip maker's value-of-time with re
e e
spect to that trip. Besides walk links, T2 would
allow links representing parking spaces with var
od ious costs and capacities. (An elastic parking
costs, which increased with demand, could fore
where y E arg min^. G[x\c(x*), t(x*)]. These equa cast a particular parking cost.) It would, then
tions could be solved with a two-stage algorithmreasonably route a trip that parks and then walks
using T2-ETA. Built upon the work of EVANS (1976) to its destination.
and LEURENT (1994), the resulting algorithm would,
Mode choice, as already mentioned, could be
for example, yield simultaneous destination-mode
come more streamlined by using a T3 model with
route choice equilibrium.
time, cost and inconvenience. The VOT PDF could
Tn: Beyond Two Criteria. While time and cost are be correlated with origin-destination parameters
paramount, they certainly do not comprise all path to provide necessary socio-demographic sensitiv
choice criteria. In various applications, other deter ity. The model would be run as a combined mode
minants of discomfort and inconvenience are influ choice traffic assignment, which would expedite
ential, e.g., schedule reliability, number of transfers, processing and assure continuity of travel time
safety, etc. Even the simplest mode-choice model estimates. It would be interesting to compare the
would probably require time, cost, and inconve results of this simple model with a conventional
nience. Such a "T3" model uses two parameters: alogit model.
Land use forecasts typically use a gravity-model type trips Set data describing trip matrix
approach. It is well known that a gravity model Node o origin
can be integrated into an equilibrium traffic as Node d destination
Real v number of trips from o to d
signment model (evans, 1976). These forecasts type FlowPair a traffic assignment (or
would likely improve if they used T2's path find Vector differential)
ing for determining the expected cost, time, and RealArray[|^|] arc total-flow vector (e.g., x, x?,
generalized cost separating zones Furthermore,
Real Array[\%\ ] arc first-moment vector (e.g., u,
the cost of "settling" in a zone could be made u?, Au)
elastic, depending on zone's attraction potential,
and T2-ETA would solve this equilibrium prob A2: ALGORITHM T2-MPA
lem.
FlowPair T2MPA(Net n, Trips Y, Real Function Array H)
(i?, c?) <- (i(x?), c{x?))
ACKNOWLEDGMENTS (x, a) <- (0, 0)
Individuals from several institutions contributed for (o, d, v) G V
(&, a) <? (likely Paths(n, o, d), 0) (see A3)
directly and indirectly to the production of this pa for m G {1, . . ., fc>
per. Academe: I thank Tom Magnanti and Rob
Freund of the Massachusetts Institute of Technol loadPath(n, o, d, aPm_lf a, b, v)
a <? 6
ogy for their lucid math programming lectures; loadPath(n, o, d, aPQ, a, o?, y)
David Boyce of the University of Illinois at Chicago, return (x, S)
who read several drafts this paper; David Bernstein loadPath(Net n, Node o, d, VOT a, a, 6, Real u)
of Princeton University for his helpful comments; minPath(n, o, <i, a) (see A3)
Mike Florian and Patrice Marcotte of the University while e 0
of Montreal, whose counter example (florian and
(xc, ue, e) <- (xc + vFod(a, b), ue + vHod(a, b), q^)
marcotte, 1994) prompted my finding and repair return.
ing a serious error in an earlier draft; and particu
larly Georgia Perakis of MIT for introducing me to A3: ALGORITHM LIKELYPATHS
the generalized Frank-Wolfe algorithm and for her
Integer likely Paths(Net n, Node o, d, Path Array p)
many helpful suggestions. Volpe Center: I am very k <? 1; (count of likely paths)
grateful for my excellent working environment and p <- minPath(n, o, d, oo) (fastest path)
generous colleagues: the assistance Sarah Maccal Pi minPath(n, o, d, 0) (cheapest path)
midPaths(n, o, d, p, p, k) (likely paths between)
ous, the opinions of Mark Bucciarelli, the insights of return^ + 1) (count of p found)
Peter Mengert, and most importantly the encour
midPaths(Net n, Node o, d, Path p, Path Array p, Intege
agement of Mike Jacobs. Transportation Science: ref k)
Feedback from the Editor, Associate Editor, and two ol <? (cp - cPk)/(tPk - tp) (slope of connecting line)
astute referees greatly improved this paper's form p <? minPath(n, o, d, a) (find path in between)
and content. I am indebted to their thoughtful re if gp < Cp + (check if done)
midPathsin, o, d, p, p, fc) (paths between)
views and very helpful suggestions. Department of midPaths(n, o, d, p, p, &) (paths between)
Transportation: Finally thanks go to the U.S. DOT, else & ?- & + 1 (bump path count)
who with the Environmental Protection Agency, pk <? p (stash p in its spot)
sponsored this work as part of its Transportation Path minPaMNet Of, ?), Node o, d, VOT a)
Model Improvement Program (TMIP); and to its for i G >T gf <? oo (c- is cost to node i)
(c0, t0, So, <7o> SO ?- (0, 0, 0, 0, {o?
technical representative, Fred Ducca, without (^ is cost to node i)
whose support this paper would not have been writ for i G ^ is GC to node i)
ten. for e = (i, jr) G ^ (qt is i's predecessor arc)
(?, c, 2) <- + ce + ate(x?e), ct + ce, ^ + fe)
APPENDIX A: FORMAL T2 ALGORITHMS vtg <gj\/(g = gjAt <Cj)
Al: DATA TYPES FOR ALL ALGORITHMS V (6 = gj At = cj At < cj)
(gp cp tp qj, SO (?, c, *, e, SP n {/?
type Net data describing network return (a, cd, td, gd).
Node Set N nodes in network
Arc Set % arcs in network A4: ALGORITHM T2-ETA
type Path data for min path p* (for VOT a)
VOT a value-of-time used to findFlowPair
p* T2ETA(Net n, Trips V, Real e)
Cost c cost of min path: cp* ??) <- (0, 0) (start with zero flow)
Time t time of min path: tp* (jc?, <- T2MPA(n, Y, if) (T2 min-path assignment, A2)
GC g gc of min path: gp* = cp* do
+ atp*
110 / R. B. DIAL
% u) ?- T2MPA(n, Y, H) (T2 min-path assignment, A2) Dial, R., "A Model and Algorithm for Multicriteria Route
(Ax, Aw) <? (x - jc?, S - S?) (ascent direction) Mode Choice," Transportation Research, 13B, 311-316
8 <_ L(?> A*> (relative error) (1979) .
if e <^s^' (if not done, get step size) Evans, S., "Derivation and Analysis of Some Models for
if L(l, Ax, AS) < 0 Combining Trip Distribution and Assignment," Trans
A ^ 1 portation Research 10, 37-57 (1976).
else A <- argAG(0 L(A, As, AS) = 0 Florian, M. and P. Marcotte, Personal Correspondence,
(x?, S?) ?- (x? + AM, S? + AAu)
University of Montreal (1994).
(and update flows)
until s < e (quit when rel. error small) Frank, M. and P. Wolfe, "An Algorithm for Quadratic
return (x?, S?). (send back updated flow) Programming." Naval Research Logistics Quarterly 3,
Real L(Real A, x, u) 95-110 (1956).
return (2e E^ [ce(x? + AAxe)xc + *e(x? + AAxe)z/J). Friesz, T., D. Bernstein, T. Smith, L. Tobin L. and B. Wie,
Real Hod(VOT a, 6) "A Variational Inequality Formulation of the Dynamic
return (/* afod(a)da). Network User Equilibrium Problem," Operations Re
search 41:1, 179-191 (1993).
Real Fod(VOT a, 6) (Prob[a E [a, &)])
return (Jj ?d(a)da). Hearn, D., "Bounded Flow Equilibrium Problems by Pen
alty Methods," Proceedings of the IEEE International
A5: ALGORITHM T1.5-ETA Conference on Circuit and Computers 1, 162-166
(1980) .
FlowPairT2ETA(Net n, Trips Y, Real e)
Janson, B., "Dynamic Traffic Assignment for Urban Road
(x?, u?) <- (0, 0) (start with zero flow)
(x?, S?) <- T2MPA(n, Y, H) (T2 min-path assignment, A2) Networks," Transportation Research 25B, 143-161
do (1994).
(x, S) <- T2MPA(n, T, #) (T2 min-path assignment,Kapur,
A2) J. and H. Kesavan, Entropy Optimization Princi
(Ax, AS) ?- (x - x?, S - u?) ples with Applications, Academic Press, 1992.
(ascent direction)
Leurent, F., "Cost Versus Time Equilibrium over a Net
s ^_ L(?> A*> AS) (relative error) work," Presented at the 1994 Annual Conference of the
if e <Js^' *^ (if not done, get step size) Transportation Research board. Washington, DC
if L(l, Arc, AS) < 0 A <? 1
(1994).
else A <- argA6(01) L(A, Ax, As) = 0
(x?, S?) ^ (x? +' A As, S? + AAS) LeBlanc, L., "Mathematical Programming Algorithms for
(and update flows) Large Scale Network Equilibrium and Network Design
until s < e (quit when rel. error small) Problems," Ph.D. Thesis, Northwestern University,
return (x?, S?). (send back updated flow) 1973.
Real L( Real A, x, S) (different from Algo T2-ETA)
LeBlanc, L., B. Ran and D. Boyce, "Dynamic Travel Choice
return (2ee^ [ceue + ?e(x? + AAxe)xJ). Models for Urban Transportation Networks," Proceed
Real Hod(VOT a, b) (different from Algo T2-ETA)ings of the 1992 IEEE International Conference on
return (J* a-^Mda). Systems, Man and Cybernetic 1, 225-231 (1992).
Real Forf(VOT a, b) (Prob[a e [a, 6)]) Magnanti, T. and G. Perakis, "A Unifying Geometric So
return (Jj /od(a)da). lution Framework and Complexity Analysis for Varia
tional Inequalities," Working Paper OR 276-93, Oper
REFERENCES ations Research Center, MIT, Cambridge, MA, 1993a.
Magnanti, T. and G. Perakis, "On the Convergence of
Ahuja, R., T. Magnanti and J. Orlin, Network Flows, Pren
Classical Variational Inequality Algorithms," Working
tice Hall, Englewood Cliffs, NJ, 1993.
Paper OR 282-93, Operations Research Center, MIT,
Apostol, T., Mathematical Analysis, Addison-Wesley,
Cambridge, MA, 1993b.
Reading, MA, 1968.
Magnanti,
Bazaraa, M., H. Sherali and C. Shetty, Nonlinear Pro T. and G. Perakis, "From Frank-Wolfe to Steep
gramming: Theory and Algorithms, John Wiley & est Descent: A Descent Framework for Solving VIPs,"
Sons, New York, 1979. Presentation at TIMS/ORSA Joint National Meeting,
Ben Akiva, M., Personal Correspondence, Massachusetts
Boston (1994).
Institute of Technology (1994). Nguyen, S., "An Algorithm for the Traffic Assignment
Berstein, D., Personal Correspondence, Massachusetts In Problem," Transportation Science 8:3, 205-218 (1974).
stitute of Technology (1993). Orlin, J., "2-Criteria and 3-Criteria Shortest Path Prob
Dafermos, S., "A Multicriteria Route-Choice Traffic Equi lem: a Proposal," Personal Correspondence, Massachu
librium Model," Paper prepared for Frontiers in Trans setts Institute of Technology (1993).
portation Equilibrium and Supply Models Symposia, Quandt, R., "A Probabilistic Abstract Mode Model," in
University of Montreal (1981). Studies in Travel Demand, VIII, Mathematica, Inc.,
Dial, R., "A Probabilistic Multipath Assignment Model Princeton, NJ, pp. 127-149 (1967).
that Obviates Path Enumeration," Transportation Re Roth, G., Paying for Roads: the Economics of Traffic Con
search 5, 83-111 (1971). gestion, Penguin Books, Middlesex, England, 1967.
Schneider, M., "Access and Land Development," in Pro Slavin, H., Personal Correspondence, Caliper Corpora
ceedings of the conference on Urban Development Mod tion, Newton, MA, (1993).
els, Dartmouth College, 1968. Wardrop, J., "Some Theoretical Aspects of Road Traffic
Sheffi, Y. and C. Daganzo, "On Stochastic Models of Traf Research," Proc. Inst. of Civil Engineers, Part 2, pp.
fic Assignment," Transportation Science 11:3, 253-274 325-378 (1952).
(1977).
Silverman, B., Density Estimation for Statistics and Data (Received: July 1994; revisions received: March 1995, April 1995;
Analysis, Chapman and Hall, New York, 1986. accepted: May 1995)

Bicriterion Traffic Assignment Basic Theory and Elementary Algorithms

Uploaded by

Copyright:

Available Formats

You might also like

Bicriterion Traffic Assignment Basic Theory and Elementary Algorithms

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Bicriterion Traffic Assignment Basic Theory and Elementary Algorithms

Uploaded by

Copyright:

Available Formats

Bicriterion Traffic Assignment: Basic Theory and Elementary Algorithms

Author(s): ROBERT B. DIAL

INFORMS is collaborating with JSTOR to digitize, preserve and extend access to

very differently. Its mathematical programming ap 100

0 .25 .50 .75 1.00 1.25 1.50 1.75 2.00

zero, bp must be strictly greater than ap. The table

1.0 2000 881

$3.40 387 $5.591 $1.09 413

Prob[p] = Prob[ap < a < bp] = FJJbp) - F?,(ap). 1. T2 MIN-PATH ASSIGNMENT

G[x(a?) | c*, i*] = G[y(a?) | c*, **]. (7)

= X [C?(*:)(*. - <) + - "e)]

x?G argmin G[x | c*, **]. (13)

-2 [c.(*5*:+ *.(*>;]*<). d4) A*. xe(a)da - x?e, Aue axe(a)da -

-2 [c.(*:s)*:+a*t)?a = o. us) = X l>e(*e + kAxe)xe + ^(xj + AAxe)i*J.

Adding (14) and (15) yields If

X [ce(**)x.+ te(x*>Ue] L[0, (A*e), (AiQ]

2 [c.(x*)(*. - x% + te(xt)(ue - ?*)] ^ 0. (17) if L[l, (A*e), (Am.)] ^ 0)

xep(a) = trips with VOT a assigned to path p E 2Pe

When Algorithm T2-MPA assigns a path's flow to its

= 1 v0pdp 'P af0pdp(a)da. (18) /12(a) = 2a

0.0 0.5 1.0 xb - 9.9

< = 7.2 <a = 7.2

0.0 5 10 0.0 0.5 1.0 Xft = 5.5

direction to obtain an improved beginning state for Similarly,

7.0 3.0 4.6 2.1 47.6 0 1 7.0 6.0

7.0 3.0 4.5 2.2 47.5 0 1 7.0 6.0

7.1 2.9 4.6 2.0 47.6 0 1 7.1 5.7

7.0 3.0 4.5 2.2 47.5 0 1 7.0 6.0

7.1 2.9 4.6 2.1 47.5 0 1 7.1 5.8

7.0 3.0 4.4 2.3 47.4 0 1 7.0 6.1

10 7.1 2.9 4.5 2.2 47.4 0 1 7.1 5.8

10.61 800 5.45 0 167.29 3200

dQ _ ce 37.51 2069 5.49 579 15.01 1352

which is clearly the case:

12.65 1153 5.60 845 34.09 2002

1974 5.56 783 13.58 1243

1065 5.( 942 33.68

a dxe\) Jdxe(a) a 3 = 0.04 A* = 0.020

That Q( ) is convex in 9? is also clear. Its Hessian is (d)

3. SUMMARY AND CONCLUSIONS

3.1. Additional Research and Development

You might also like

G[x(a?) | c, i] = G[y(a?) | c*, **]. (7)

= X [C?(:)(. - <) + - "e)]

-2 [c.(5:+ .(>;]<). d4) A. xe(a)da - x?e, Aue axe(a)da -

-2 [c.(:s):+at)?a = o. us) = X l>e(e + kAxe)xe + ^(xj + AAxe)i*J.

X [ce(**)x.+ te(x>Ue] L[0, (Ae), (AiQ]

2 [c.(x)(. - x% + te(xt)(ue - ?)] ^ 0. (17) if L[l, (Ae), (Am.)] ^ 0)