A Strategic Model For Dynamic Traffic Assignment

A STRATEGIC MODEL FOR
DYNAMIC TRAFFIC ASSIGNMENT∗
by
Younes Hamdouch1
Patrice Marcotte1
Sang Nguyen1
23 octobre 2002
∗
This research was supported in part by the Natural Sciences and Engineering Re-
search Council of Canada (NSERC) and by the Fonds pour la formation de chercheurs
et l’aide à la recherche (FCAR).
1
Département d’informatique et de recherche opérationnelle et Centre de recherche sur
les transports, Université de Montréal, C.P. 6128, succursale Centre-ville, Montréal,
CANADA H3C 3J7

ABSTRACT
We propose a model of dynamic traffic assignment where strategic choices are an
integral part of user behaviour. The model provides a discrete-time description of
flow variation through a road network involving arcs with rigid capacities. A driver’s
strategy assigns, to each node of the network, a set of arcs in the forward star of
that node, sorted according to a preference order. The main element of the model
is a ‘within-day’ submodel where strategic volumes are loaded onto the network in
accordance with the first-in first-out discipline and user preferences. An equilibrium
assignment is achieved when expected delays are minimal, for every origin-destination
pair. We prove the existence of such an assignment and provide numerical results on
test networks.
Keywords: dynamic traffic assignment, strategy, hyperpath, capacities.
RÉSUMÉ
Cet article propose un modèle dynamique discret pour le trafic routier où le com-
portement des usagers est dicté par des stratégies. Le modèle permet d’analyser la
variation des flots de véhicules dans les réseaux dont les arcs sont munis de capacités
rigides. Chaque stratégie associe à tout nœud du réseau un ensemble d’arcs inci-
dents, ordonné en ordre de préférence décroissante. L’élément central du modèle est
un sous-modèle où les volumes stratégiques sont chargés sur le réseau en respectant
les priorités temporelles (premier arrivé premier servi) ainsi que les préférences. Un
équilibre est atteint lorsque tous les automobilistes sont affectés à des plus courts
chemins reliant leur origine et leur destination. Nous démontrons l’existence d’une
affectation respectant ce principe d’équilibre et présentons des résultats numériques
sur des réseaux test.
Mots-Clés: trafic routier, affectation d’équilibre dynamique, stratégie, hyperchemin,
capacités.

1 Introduction
Traffic assignment has been the topic of many studies in the last four decades, and it
can be safely argued that the static case is settled by now. However, extending the
results from a static to a dynamic environment is far from trivial and, as no agreement
on an ideal model has been reached within the scientific community, dynamic traffic
assignment remains a topic of active research.
Macroscopic dynamic models belong to one of three classes: models using exit
functions (Merchant and Nemhauser [10], Carey [2], Drissi-Kaı̈touni and Gendreau [4]),
models based on space-time networks (Drissi and Hamada-Benchekroun [3], Zawack
and Thompson [15]) and models using arc travel delays (Friesz and al. [5], As-
tarita [1]). In all these models, users travel along paths of the underlying network,
which path cannot be modified en route. If congestion levels become very high, how-
ever, users may choose to switch path rather than experience unduly delays on a
given route. Our proposal for dealing with this situation is to assume that users be-
have strategically and modify their route choice according to congestion conditions.
More precisely, a strategy associates with each node of the network an ordered set of
successor nodes, whose access is controlled by three factors: the arrival time at the
node, the capacity of the corresponding arc, the number of users with overlapping
strategies.
This model adapts the strategic approach of Marcotte, Nguyen and Schoeb [9] and
Hamdouch [6] to a time-varying environment. As in most dynamic models, the main
challenge consists in performing efficiently the ‘loading’ operation, whereby strategic
volumes are converted into path and arc volumes. It is a key feature of our approach
that the loading preserves the order of arrival of users at the nodes of the network,
and thus automatically fulfills the FIFO (First-In First Out) rule.
The remainder of this paper is organized as follows. In the next section, we
introduce the strategic concepts underlying the dynamic assignment model and for-
mulate the equilibrium problem as a variational inequality. Based on these results,
we develop an algorithmic framework and provide numerical results. Throughout the
paper, we assume that time-dependent demand and departure times are known, and
that congestion is the sole consequence of waiting time at the tails of congested nodes.
2 A strategic dynamic assignment model
Let G = (N, A) be a network with node set N, arc set A and time-varying demand
dt
qr (t = 0, 1, . . . , T′
≤ T), for users departing origin node q at t and bound for
destination node r. The objective of our research is to determine strategic volumes
and travel times over a fixed time interval T whose expected delays are consistent
with the travel costs ca and the arc capacities ua. We assume that the time interval
is sufficiently large to allow all vehicles to reach their destination by T and that,
whenever the number of drivers that wish to access an arc exceeds its capacity, a
proportion of those commuters will either travel on an alternative (unsaturated) arc,
or wait until an arc become available later in time. All operations take place on a
1

time-space network R = (V, E) which is constructed as follows:
- each node i ∈ N is expanded into T + 1 nodes it, t = 0, 1, . . . , T:
V = {it : i ∈ N, 0 ≤ t ≤ T}
- each link a ∈ A is expanded to T + 1 − ca links (it, jt+ca ) such that t + ca ≤ T;
we refer to these links as travel links. The waiting time at the nodes is
represented by means of waiting arcs of infinite capacity denoted (it, it+1)
where t = 0, 1, . . . , T − 1.
The set of arcs is partitioned into the set E1 of access arcs and its complement E2,
the set of waiting arcs:
E1 = {(it, jt+cij
) : (i, j) ∈ A, 0 ≤ t ≤ T − cij}
E2 = {(it, it+1) : i ∈ N, 0 ≤ t ≤ T − 1}
E = E1 ∪ E2
As an example, the time-space network R of 5 periods corresponding to the example
of Figure 1 is illustrated in Figure 2. Note that R is a standard acyclic graph with
|V | = n(T + 1) and |E| ≤ (m + n)T where n = |N| and m = |A|.
♠ ♠ ♠
i j k
✲ ✲
cij = 1
uij = ∞
cjk = 2
ujk = 5
Figure 1: A simple network
2.1 Strategies
A strategy s is associated with an origin-destination pair (q, r)(t) and maps every
node jt of the time-space network into an ordered set of successor nodes that embodies
the arrival instant at jt. More precisely, let Et
j denote the set of all subsets of the
forward star j+
t associated with node jt. A strategy s assigns to each node jt ∈ V an
element Estt′
j (t′
≤ t) in Et
j. The set Estt′
j is the preference order at node jt for users
adopting strategy s and having reached node j at instant t′
. Three strategies for the
simple network of Figure 2 are listed below:
preference order: Es00
i Es11
i Es22
i Es11
j Es21
j Es22
j Es33
j Es55
k
s1 : [ j1 ] [ ] [ ] [ k3, j2 ] [ k4 ] [ k4, j3 ] [ ] [ ]
s2 : [ ] [ j2 ] [ ] [ ] [ k4, j3 ] [ k4 ] [ k5 ] [ ]
s3 : [ ] [ ] [ j3 ] [ ] [ ] [ ] [ k5 ] [ ]
2

♠ ♠ ♠
i5 j5 k5
t = 5
t = 4
t = 3
t = 2
t = 1
t = 0
15
❅
❅
❘
♠ ♠ ♠
i4 j4 k4
♠ ♠ ♠
i3 j3 k3
♠ ♠ ♠
i2 j2 k2
♠ ♠ ♠
i1 j1 [5]
[5]
[5]
[5]
k1
♠ ♠ ♠
i0 j0 k0
❅
❅
❅
❅
❅
❘
❅
❅
❅
❅
❅
❘
❅
❅
❅
❅
❅
❘
❅
❅
❅
❅
❅
❘
❅
❅
❅
❅
❅
❘
❆
❆
❆
❆
❆
❆
❆
❆
❆
❆
❆
❯
❆
❆
❆
❆
❆
❆
❆
❆
❆
❆
❆
❯
❆
❆
❆
❆
❆
❆
❆
❆
❆
❆
❆
❯
❆
❆
❆
❆
❆
❆
❆
❆
❆
❆
❆
❯
❄
❄
❄
❄
❄
❄
❄
❄
❄
❄
❄
❄
❄
❄
❄
Figure 2: Expansion of the simple network
With any given strategy s, we associate the strategic graph H(s) = (V s
, Es
),
where Es
is the union of the set Es
1 of travel arcs {(jt, kt+cjk
) ∈ E1 kt+cjk
∈ Estt′
j , t′
≤
t} and the set Es
2 of waiting arcs {(jt, jt+1) ∈ E2 jt+1 ∈ Estt′
j , t′
≤ t}. Every node
jt ∈ V s
is connected to the destination r(s). The strategic graph induced by strategy
s1 is shown in Figure 3.
The set of all strategies associated with OD pair (q, r)(t) is denoted by St
qr and
we denote by
S =
[
0≤t≤T
St
=
[
0≤t≤T
[
(q,r,t)
St
qr,
the set of all strategies.
2.2 Strategic volumes and strategic costs
Let xt
s represent the number of drivers using strategy s and leaving their origin q(s)
at instant t. We denote by x = (xt
s)s∈S the vector of strategic volumes and by X
the set of all demand-feasible such vectors, i.e.,:
X = {x :
X
s∈St
qr
xt
s = dt
qr ∀(q, r), ∀t ∈ T} (1)
3

✒✑
✓✏
k4 r4(s1)
✒✑
✓✏
k3
✒✑
✓✏
j2
✒✑
✓✏
j1
✒✑
✓✏
i0
q0(s1)
❅
❅
❅
❘
❅
❅
❅
❅
❅
❅
❘
❅
❅
❅
❅
❅
❅
❘
❄
Figure 3: The strategic graph of s1 on simple network
The cost of strategy s depends directly on the arc access probabilities πst′
jk (x)
and πst′
j (x), the latter corresponding to a waiting arc; πst′
jk (x) (respectively πst′
j (x))
is the probability that drivers adopting strategy s access node kt′+cjk
(respectively
jt′+1) from node jt′ . These access probabilities are computed by a dynamic loading
process that will be the focus of Section 3. The arc access probabilities induce node
access probabilities, computed recursively as follows:
τst′
j (x) =









0 if jt′ /
∈ V s
1 if jt′ = qt(s)
τ
s(t′−1)
j π
s(t′−1)
j +
P
ktc ∈j−
t′
τstc
k πstc
kj if jt′ ∈ V s
{qt(s)},
(2)
where tc = t′
− ckj is the instant when a user of strategy s leaves node k through arc
(k, j).
In a natural fashion, we define the strategic cost of a strategy s with departure
time t and destination q(s) as the weighted sum:
Ct
s(x) =
X
t′≥t
½ X
(jt′ ,kt′+cjk
)∈Es
1
τst′
j (x)πst′
jk (x)cjk +
X
(jt′ ,jt′+1)∈Es
2
τst′
j (x)πst′
j (x)
¾
. (3)
An equilibrium strategic flow x∗
is then readily characterized as a solution of the
variational inequality
hC(x∗
), x∗
− xi ≤ 0 ∀x ∈ X. (4)
If the cost function C were continuous, the existence of an equilibrium solution
would follow directly from classical fixed point theorems, such as Brouwer’s or Kaku-
tani’s. A technical difficulty arises in the degenerate situation where a null strategic
volume wants to access an arc with null residual capacity. This can be resolved ex-
actly as in Hamdouch [7], i.e., by showing that defining the access probability to zero
in the degenerate case yields a lower semi-continuous cost mapping C. The proof will
not be repeated here.
4

2.3 The FIFO rule and flow priority
A main feature of our model is that the FIFO rule is strictly enforced. At a a node
jt′ ∈ V , this is achieved dividing each group of users zst′
j into subgroups according
to their arrival instant at node j. We denote by zst′t′′
j the strategic volume at node
jt′ adopting strategy s and having reached node j at instant t′′
≤ t′
and regroup all
active strategies into a class S
t′t′′
restricted to users having reached node j at instant
t′′
:
S
t′t′′
= {s ∈ S : Est′t′′
j 6= ∅, zst′t′′
j > 0}. (5)
Based on the access probabilities πst′
jk (x) and πst′
j (x), the volume zst′t′′
j is then com-
puted according to the recursion
zst′t′′
j =















0 if jt′ /
∈ V s
xt
s if jt′ = qt(s) (t′′
= t′
)
π
s(t′−1)
j z
s(t′−1)t′′
j if jt′ ∈ V s
{qt(s)} and t′′
≤ t′
− 1
P
ktc ∈j−
t′
πstc
kj zstc
k if jt′ ∈ V s
{qt(s)} and t′′
= t′
(tc = (t′
− cjk).
(6)
In order to control access priorities, we assign the highest priority to users of the first
class S
t′0
. The strategic volume of the remaining classes are then loaded in increasing
order according to their arrival instant at node j. This ordering is an integral part of
the loading process described in the next section.
3 The dynamic loading process
In static models, the arc-path incidence matrix is flow independent. As a conse-
quence, link information can be readily recovered from path flow information. In
dynamic models, the situation is different and calls for a mechanism that maps path
(or strategic) flows into link flows. In our model, this operation is performed by a
loading procedure that assigns users to their preferred available arc, until their first
choices becomes saturated. Saturated arcs are then removed from the preference sets
and the procedure is iterated until all users are assigned to an outgoing arc.
This procedure will be explained in detail on the example corresponding to figure 4,
where capacities of the physical arcs (j, k) and (j, l) are set to 10 and 15, respectively.
We focus on the loading process at node j4, i.e., the physical node j at instant 4. A
total of 20 users are already in the waiting queue at j; among these, 15 users have
been in the queue for two periods, and the remaining 5 for one period. They are joined
in the queue by 20 users that access node j from node i. All users adopt one of either
two strategies: s1 or s2, which are specified by the preference sets Estt′
j displayed in
the figured. Note that the preference set of a strategy can be time dependent, which it
is in our example. Next to each arc is shown its volume, together with the repartition
among the two classes.
We initiate the process by loading the strategic flow associated with class S
42
, i.e.,
flow that arrived at node j at instant 2. Since capacity is not exceeded, 3 users from
5

t = 2
t = 3
t = 4
t = 5
t = 6
♠
♠
♠
k6
j5
l6
♠
j4
♠
j3
♠ ♠
i2 j2 Es142
j = {k6, j5}
Es143
j = Es144
j = {k6, l6, j5}
Es242
j = {l6, j5}
Es243
j = Es244
j = {l6, k6, j5}
❆
❆
❆
❆
❆
❆
❆
❆
❆
❆
❆
❯
❆
❆
❆
❆
❆
❆
❆
❆
❆
❆
❆
❯
❅
❅
❅
❅
❅
❅
❅
❅
❅
❅
❅
❘
❄
5 (s1:3, s2:2)
15 (s1:10, s2:5)
20 (s1:12, s2:8)
∞
(10) (15)
❄
❄
Figure 4: An example of strategic dynamic loading
s1 and 2 from s2 head for nodes k6 and l6, respectively. Next, residual capacities are
updated and the loading ends for drivers of class. At the second iteration, strategic
volumes zs143
j and zs243
j of the second class S
43
are assigned to their respective preferred
node at a uniform rate. Since the ratio 7/10 is less than 13/5, arc (j4, k6) gets
saturated first. At that instant, 7
10
of s1 and s2 flows are assigned, i.e., 7 users from s1
reach node k6 and 3.5 from s2 reach l6. At iteration 3, the saturation of (j4, k6) forces
users of strategy s1 to adopt their second best choice l6; 4.5 units of volume (3 from
s1 and 1.5 from s2) vie for arc (j4, l6) whose residual capacity is 9.5. Since capacity
exceeds demand, the loading of class S
43
terminates. At the fourth iteration, 12 units
of the strategic volume zs144
j and 8 of zs244
j compete for the residual capacity (5) of
arc (j4, l6). According to the ratio 1 to 4, 3 units from s1 and 2 from s2 reach node
l6. A fifth iteration is required to load the unassigned flow to the waiting arc (j4, j5).
The arc probabilities are then obtained by dividing the amount of strategic volume
having accessed an arc by the total corresponding strategic volume. More details are
provided in Table 1.
A pseudocode of the algorithm, which extends that presented in Hamdouch et
al [7] for static networks with priority rules, is given in the appendix. It provides a
formal description of the loading of a strategic flow vector x = (xt
s)s∈Wt⊂St, t≤T with
respect to a subset of strategies Wt
working set. The working set includes at least
one strategy for every OD pair (q, r)(t).
6

arc: (j4, k6) (j4, l6) (j4, j5)
iter. 1 res. cap. 10 15 ∞
flow 3(zs142
j ) 2(zs242
j ) 0
flow 3(zs142
j )+7( zs143
j ) 2(zs242
j ) + 3.5(zs243
j ) 0
iter. 3 res. cap. 0 9.5 ∞
flow 3(zs142
j )+7( zs143
j ) 2(zs242
j ) + 3.5(zs243
j ) 0
+ 3(zs143
j ) + 1.5(zs243
j )
flow 3(zs142
j )+7( zs143
j ) 2(zs242
j ) + 5(zs243
j ) 0
+ 3(zs143
j ) + 3(zs144
j )
+ 2(zs244
j )
flow 3(zs142
j )+7(zs143
j ) 2(zs242
j ) + 5(zs243
j )+ 9(zs144
j )
+ 3(zs143
j ) + 3(zs144
j ) + 6(zs244
j )
+ 2(zs244
j )
access probabilities πs14
jk = 10
25
πs14
jl = 6
25
πs14
j = 9
25
πs24
jl = 9
15
πs24
j = 6
15
Table 1: Dynamic loading process at node j4
7

4 Computation of an optimal strategy
The algorithms that we implemented are extensions to the dynamic case of algorithms
presented in [6]. At each iteration, one moves from the current strategic vector toward
another vector that incorporates a ‘best’ response to current traffic conditions, for
each origin-destination pair. More precisely we construct, for every OD pair (q, r)(t)
and instant t, a strategy s∗
qrt(x) such that:
Ct
s∗ (x) = min
s∈St
qr
{Ct
s(x)}.
The construction of the solution s∗
to the above problem is achieved by an al-
gorithm that plays the role of the shortest path algorithm in standard static traffic
assignment. It is based on dynamic programming and uses the information (strategic
volumes, access probabilities) generated by the loading procedure. One conceptual
difficulty is that access probabilities associated with strategies not yet in the working
set are unknown. However, a micro loading procedure can be devised to perform
that very operation.
The knowledge of the strategic vector z allows us to start the micro-loading at
node jt′ which must comply with the FIFO rule. At each instant p (t ≤ p ≤ t′
), the
optimal preference set Es∗t′p
j is constructed by running the micro-loading over the sets
S
t′t
, S
t′t+1
, . . . , S
t′p−1
and S
t′p
∪ {s∗
} given that the users of the optimal (unknown)
strategy s∗
have reached node j at instant p. The dependence of the preference set on
the arrival instant at node j is key to enforcing the FIFO condition at node jt′ . The
recursive process of Nguyen and Pallottino [11] for computing a shortest hyperpath
can then be applied to the space-time network R. The expected travel cost from
node jt′ to the destination rT , denoted ωs∗t′
pj , is obtained from Bellman’s generalized
Bellman:
ωs∗t′
pj =









∞ if j 6= r, t = T
0 if j = r
P
kt′′ ∈Es∗t′p
j
πs∗t′
jk (cjk + ωs∗t′′
t′′k ) + πs∗t′
j (1 + ωs∗t′+1
pj ) if j 6= r, t < T,
(7)
where t′′
= t′
+ cjk is the arrival instant at node k ∈ j−
, and where the optimal
preference order Es∗t′p
j is constructed efficiently by sorting the labels cjk + ωs∗t′′
k in
increasing order, with
ωs∗t′′
k =
(
ωs∗t′′
pk if kt′′ = kt′+cjk
= jt′+1
ωs∗t′′
t′′k otherwise.
An optimal solution s∗
, together with its cost vector ω∗
, is then obtained by solving
Bellman’s equations in reverse topological order, i.e., starting at instant T.
We close this section with an illustration of one iteration of the algorithm within
the space-time network. Let us consider the situation depicted in Figure 5, where
the relevant data comprises the arrival instants, the arc flows, their associated strate-
gies, the preference orders associated with strategies s1 and s2, delays, capacities (in
8

t = 6
t = 5
t = 4
t = 3
t = 2
♠
♠
♠
k6
ωs∗5
3j = 3.7
ωs∗6
6k = 2
j5
ωs∗5
2j = 3.4
ωs∗5
5j = 4.2
l6
ωs∗5
4j = 4
ωs∗6
6l = 2.5
♠
j4
♠
j3
♠ ♠
i2 j2
Es142
j = {k6, l6, j5}
Es143
j = Es144
j = {k6, l6, j5}
Es242
j = {l6, j5}
Es243
j = Es244
j = {l6, j5}
❆
❆
❆
❆
❆
❆
❆
❆
❆
❆
❆
❯
❆
❆
❆
❆
❆
❆
❆
❆
❆
❆
❆
❯
❅
❅
❅
❅
❅
❅
❅
❅
❅
❅
❅
❘
❄
8 (s1)
10 (s1)
15 (s2)
(∞)
delay = 1
(10)
delay = 2
(15)
delay = 2
❄
❄
Figure 5: Computing an optimal strategy at node j4
parenthesis) and cost-to-go (bottom of picture) that are available at nodes j5, k6 and
l6, since their topological index is higher than that of node j. We proceed to compute
the best strategic response, starting at the physical node j.
The optimal preference orders are obtained by minimizing the sum of arc delays
and the corresponding cost-to-go. Accordingly, the inequalities
cjk + ωs∗6
6k = 2 + 2 < 1 + ωs∗5
2j = 1 + 3.4 < cjl + ωs∗6
6l = 2 + 2.5
yield Es∗42
j = [k6, j5, l6]. In a similar fashion, we derive Es∗43
j = [k6, l6, j5] from
cjk + ωs∗6
6k = 4 < cjl + ωs∗6
6l = 4.5 < 1 + ωs∗5
3j = 1 + 3.7
and Es∗44
j = [k6, l6, j5] from the inequalities
4 < 4.5 < 1 + ωs∗5
4j = 1 + 4.
The algorithm must now determine the values of ωs∗4
2j , ωs∗4
3j and ωs∗4
4j at node j4.
We start with instant 2 and the strategy set S
42
∪ {s∗
} = {s1, s∗
}. Since flow does
not exceed available capacity, all users from s1 and s∗
head for node k6. This yields
the access probabilities πs∗4
jk = 1, πs∗4
j = πs∗4
jl = 0 and the cost
ωs∗4
2j = πs∗4
jk (2 + ωs∗6
6k ) + πs∗4
j (1 + ωs∗6
2j ) + πs∗4
jl (2 + ωs∗6
6l )
= 1(4) + 0(4.4) + 0(4.5) = 4.
9

We now repeat the loading operation for users that arrived at instant 3. We first
perform the assignment of the 8 users that arrived at instant 2 to their preferred
node k6. Next, the residual capacity of arc (j, k) is updated (it becomes 10-8=2)
and the loading ends for drivers of strategies S
42
. Next, the strategic volumes zs143
j
and zs∗
j associated with the strategies in S
43
are assigned to their respective preferred
node at the uniform rate 2/10. At that instant, 1
5
of s1 and s∗
flows have reached
node k6, and the arc (j4, k6) is saturated. This forces users of strategy s1 and s∗
to
adopt their second best choice l6; 8 units of volume from s1 and s∗
access arc (j4, l6),
whose residual capacity (15) exceeds demand (8). This yields the access probabilities
πs∗4
jk = 1/5, πs∗4
jl = 4/5 and πs∗4
j = 0, from which we derive the cost-to-go
ωs∗4
3j = πs∗4
jk (2 + ωs∗6
6k ) + +πs∗4
jl (2 + ωs∗6
6l ) + πs∗4
j (1 + ωs∗6
3j )
=
1
5
(4) +
4
5
(4.5) + 0(4.7) =
22
5
= 4.4.
Then, we proceed with users that reached node j at instant 4. This involves
loading the flows associated with strategy sets S
42
, S
43
and S
44
∪ {s∗
}. This yields
(we skip details) πs∗4
jk = 0, πs∗4
jl = 7/15, πs∗4
j = 8/15 and ωs∗4
4j and
ωs∗4
4j = πs∗4
jk (2 + ωs∗6
6k ) + +πs∗4
jl (2 + ωs∗6
6l ) + πs∗4
j (1 + ωs∗6
4j )
=
7
15
(4.5) +
8
15
(5) =
71.5
15
= 4.76.
5 Numerical results
Numerical results have been obtained using algorithm DSTRATEQ, which mimics
the method of successive averages (MSA) for solving monotone variational inequali-
ties. Since the cost mapping C may fail to be monotone, this method is heuristic in
our context. At each iteration, we set
xk+1
= (1 − θk
)xk
+ θk
x̄(xk
), k = 1, 2, . . .
where θk
∈ (0, 1) and
x̄t
s(xk
) =
(
dt
qr if s = s∗
0 otherwise
, ∀(q, r) ∀s ∈ Wt
qr, ∀t ≤ T.
For improved convergence, the stepsize was chosen adaptively and independently for
each OD pair. More precisely:
θs = 1 −
Cs∗
Cs
t
.
The algorithm was halted as soon as the relative gap function became smaller than
some predetermined tolerance. We report numerical results on one small and one
medium-sized network.
10

5.1 Small network
Our first test problem is based on the bidirectional network illustrated in Figure 6.
The number of periods T is set to 65 and the latest departure instant T′
is set to 9.
Additional data for the reference scenario is shown in Tables 2 and 3. A second
scenario was constructed by decreasing the capacities of arcs (2,3) and (3,2) to 5
units, and simultaneously multiplying all demand by the factor 3/2. Key statistics
are displayed in Tables 4, 5, 6 and 7, namely: number of active strategies, gap
function, CPU time and minimum strategic costs.
In both scenarios, DSTRATEQ converged rapidly to a unique equilibrium so-
lution, from any starting point. The relative gap function decreases to 1% after
roughly 20 iterations for the both networks. Also, as expected, the increase pressure
on available capacity in the second scenario yields an increase in the number of active
strategies, and to increased CPU time. It can also be readily observed from the out-
put data that FIFO is satisfied, i.e., later departures at the origin node result in later
arrivals at the destination. Due to the demand patterns and the system’s dynamics,
travel delays increase sharply as one approaches the peak period, around instants 6
and 7.
5.2 Sioux Falls network
This classical network (see [13]) illustrated in Figure 7 comprises 24 nodes, 76 arcs
and 8 OD pairs. The number of periods T and the latest departure time T′
are set
to 100 and 4, respectively.
Numerical tests have been performed on three scenarios. Tables 8 and 9 contain
the relevant input data for the base scenario A. In scenario B, some arc capacities
are modified (see Table 10) and demand is multiplied by a factor 2 throughout. The
modifications relevant to scenario C are displayed in Table 11. For each scenario, the
control parameters ǫ1 and ǫ2 have been set to 0.1.
Tables 12, 13 and 14 show that convergence occurs at a reasonable rate, with a
relative gap inferior to 1% reached after less than 20 iterations. As was the case
on the small network, the algorithm converged toward a unique equilibrium solution.
Tables 15, 16 and 17 give the profiles of minimum travel delays for different departure
times. While the nature of the solution changes, both qualitatively and quantitatively,
in Scenarios B and C (note the sharp increase of the size of the working set for
Scenario B), the rate of convergence was not greatly influenced by the increased
congestion. In particular, the running time of each major iteration remained stable,
at around one minute per iteration.
5.3 Variation of travel delays
In the strategic model, and equilibrium is reached when average delays are equalized.
However, faced with two strategies with equal mean delays, a “risk-averse” user might
prefer the one with the smallest standard deviation σt
s, defined as
11

σt
s(x) =
· X
p(t)∈Ps
κs
p(t)(x)(Cp(t)
)2
− [Ct
s(x)]2
¸1
2
,
where κs
p(t)(x) is the probability of using path p(t) and Cp(t)
is the actual delay along
path p(t).
The inclusion of a variation criterion within the model would make it intractable
from a computational point of view. However, we computed, for Sioux Falls’ sce-
nario B the standard deviation of strategic costs for OD pair (7,20), and found a
value of 0.5 . This small value is most likely to be perceived as negligible by the
users, and validates a posteriori the choice of a single criterion in the equilibrium
model.
6 Conclusion
In this paper, we have shown that the concept of strategy, previously used in the static
transit assignment problem, can be adapted to the realm of dynamic traffic assignment
on capacitated networks. This yields a model where the FIFO rule is automatically
satisfied. The nonvacuity of the solution set was proved and the implementation
showed the applicability of the concept to small networks. Moreover, the numerical
results provided interpretable insights on the strategic behavior of users.
This opens up several avenues for further research. One is the extension of the
work of [7] on static capacitated transit systems to a dynamic environment. Another
topic of research concerns the efficiency of the loading algorithm. While its high
running time follows from the number of required elementary operations, which can
be evaluated fairly accurately, it is due primarily to the fact that all physical arcs of
our test networks are capacitated. Future research will focus on taking advantage of
the small number of congested arcs in real networks. Finally, the strategic approach
is well suited at addressing situations where random accidents or incidents occur, and
where user behaviour is dictated by the appraisal of current traffic conditions, based
on past (experience) or present (on-board vehicle) information.
References
[1] V. Astarita (1996), A continuous time link model for dynamic network loading
based on travel time function, Proc. 13th International Symposium on Trans-
portation and Traffic Theory (Tarrytown, N.Y.) (J.B. Lesort, ed.), Pergamon-
Elsevier, pp. 79-102.
[2] M. Carey (1992), Nonconvexity of the dynamic traffic assignment problem, Trans-
portation Research B26, pp. 127-133.
[3] O. Drissi-Kaitouni, A. Hamada-Benchekroun (1992), A dynamic traffic assign-
ment model and a solution algorithm, Transportation Science 26, pp. 119-128.
12

[4] O. Drissi-Kaitouni, M. Gendreau (1992), A new dynamic traffic assignment
model, Publication CRT-854, Centre de recherche sur les transports, Université
de Montréal.
[5] T.L. Friesz, D. Bernstein, T.E. Smith, R. Tobin et B.W Wie (1993), A varia-
tional inequality formulation on the dynamic network user equilibrium problem,
Operations Research 41, pp. 179-191.
[6] Y. Hamdouch (2002), Affectation statique et dynamique des usagers dans un
réseau de transport avec capacités rigides, Ph.D dissertation, Département
d’informatique et de recherche opérationnelle, Université de Montréal.
[7] Y. Hamdouch, P. Marcotte and S. Nguyen (2002), Capacitated traffic assignment
with priority rules, preprint.
[8] P. Marcotte and S. Nguyen (1998), Hyperpath formulations of traffic assignment
problems, Equilibrium and Advanced Transportation Modelling (P. Marcotte and
S. Nguyen, eds.), Kluwer Academic Publisher, pp. 175-199.
[9] P. Marcotte, S. Nguyen and A. Schoeb (2000), A strategic flow model of traffic
assignment in capacitated networks, Publication 2000-10, Centre de recherche
sur les transports, Université de Montréal.
[10] D. Merchant et G. Nemhauser (1978), A model and an algorithm for the dynamic
traffic assignment problems, Transportation Science 12, pp. 183-199.
[11] S. Nguyen and S. Pallottino (1989), Hyperpaths and shortest hyperpaths, In B.
Simeone, editor, Combinatorial Optimization, volume 1403 of Lecture Notes in
Mathematics, pp. 258-271, Springer-Verlag, Berlin.
[12] S. Nguyen, S. Pallottino et P. Malucelli (2001), A modeling framework for the
passenger assignment on a transport network with time-tables, Transportation
Science 55B, pp. 238-249.
[13] C. Suwansirikul, T.L. Friesz and R.L. Tobin (1987), Equilibrium decomposition
optimization: a heuristic for the continuous equilibrium network design problem,
Transportation science 21, pp. 254-263.
[14] J.G. Wardrop (1952), Some theoretical aspects of road traffic research, Proceed-
ings of the Institution of Civil Engineers, Part II, pp. 325-378.
[15] D. Zawack et G. Thompson (1987), A dynamic space-time network flow model
for city traffic congestion, Transportation Science 21, pp. 153-162.
13

♠
♠
♠
♠
♠
♠
1
2
3
4
5
6
✲
✛
✲
✛
❄
✻
❄
✻
✒
❅
❅
❅
❘
✠
❅
❅
❅
■ ✒
❅
❅
❅
❘
✠
❅
❅
❅
■
Figure 6: Small network (T = 65)
♠
1 ♠
2
♠
3
♠
7
♠
12 ♠
11 ♠
10 ♠
16 ♠
18
♠
17
♠
14 ♠
15
♠
23 ♠
22
♠
19
♠
13 ♠
24 ♠
21 ♠
20
♠
4 ♠
5 ♠
6
♠
9 ♠
8
✲
✛
❄
✻
❄ ❄
❄
❄
❄
❄ ❄
✲ ✲ ✲
✛ ✛ ✛
❄
✻
✛
✛
✲
✲
✛
✛
❄ ❄ ❄
❄
❄
❄
❄
❄
❄
❄
✲ ✲ ✲
✻ ✻ ✻ ✻
✲ ✲
✻ ✻ ✻
✲ ✲
✻ ✻ ✻
✻
✻ ✻
✻ ✻
✻
✻
✛ ✛ ✛
✛ ✛
✲
✲ ✲
❅
❅
❅
❅
❅
■
❅
❅
❅
❅
❅
■
✛ ✛
✛
❅
❅
❅
❅
❅
❘
❅
❅
❅
❅
❅
❘
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✎✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄
✄✄
✗
Figure 7: Sioux Falls network
14

Arc Cost Capacity Arc Cost Capacity
(1,2) 3 20 (3,5) 4 10
(2,1) 3 20 (5,3) 4 10
(1,3) 2 20 (4,5) 4 10
(3,1) 2 20 (5,4) 4 10
(2,3) 1 10 (4,6) 5 25
(3,2) 1 10 (6,4) 5 25
(2,4) 2 10 (5,6) 6 25
(4,2) 2 10 (6,5) 6 25
Table 2: Small network: arc data
OD(t) Demand OD(t) Demand
(1,6,0) 5 (6,1,0) 10
(1,6,1) 10 (6,1,1) 15
(1,6,2) 25 (6,1,2) 30
(1,6,3) 18 (6,1,3) 23
(1,6,4) 15 (6,1,4) 20
(1,6,5) 17 (6,1,5) 22
(1,6,6) 30 (6,1,6) 35
(1,6,7) 45 (6,1,7) 50
(1,6,8) 15 (6,1,8) 20
(1,6,9) 12 (6,1,9) 17
Table 3: Small network: OD information
15

# iter. |W| Gap (%) Cpu (sec)
0 20 26.4923 3.52
1 34 11.3380 8.47
2 34 5.5682 13.37
5 36 2.8952 28.36
10 40 1.5934 54.81
20 40 0.8830 111.27
30 40 0.5842 164.44
50 40 0.3081 275.80
100 40 0.0817 555.09
Table 4: Small network: results of first scenario
OD(t) Minimum cost OD(t) Minimum cost
(1,6,0) 10.000 (6,1,0) 10.000
(1,6,1) 10.000 (6,1,1) 10.333
(1,6,2) 10.799 (6,1,2) 11.333
(1,6,3) 11.888 (6,1,3) 12.156
(1,6,4) 12.000 (6,1,4) 12.189
(1,6,5) 12.000 (6,1,5) 12.263
(1,6,6) 12.209 (6,1,6) 12.617
(1,6,7) 13.075 (6,1,7) 13.831
(1,6,8) 13.360 (6,1,8) 14.536
(1,6,9) 13.018 (6,1,9) 14.555
Table 5: Small network: minimum costs after 100 iterations (first scenario)
16

# iter. |W| Gap (%) Cpu(sec)
0 20 40.8394 4.12
1 36 13.3624 9.61
2 36 6.1367 15.00
5 40 3.8735 32.16
10 47 2.1099 64.32
20 53 1.0616 135.83
30 55 0.7803 210.38
50 57 0.4901 367.39
100 57 0.2222 765.75
Table 6: Small network: results of second scenario
(1,6,0) 10.000 (6,1,0) 10.333
(1,6,1) 10.333 (6,1,1) 11.111
(1,6,2) 11.528 (6,1,2) 12.499
(1,6,3) 12.347 (6,1,3) 13.492
(1,6,4) 12.508 (6,1,4) 14.133
(1,6,5) 12.674 (6,1,5) 14.641
(1,6,6) 13.501 (6,1,6) 15.808
(1,6,7) 15.289 (6,1,7) 18.134
(1,6,8) 16.613 (6,1,8) 19.669
(1,6,9) 16.188 (6,1,9) 19.992
Table 7: Small network: minimum costs after 100 iterations (second scenario)
17

Arc Cost Capacity Arc Cost Capacity
(1,2) 6 60 (13,24) 4 60
(1,3) 2 60 (14,11) 3 15
(2,1) 6 60 (14,15) 4 60
(2,6) 2 60 (14,23) 2 10
(3,1) 2 60 (15,10) 4 20
(3,4) 2 60 (15,14) 4 60
(3,12) 4 60 (15,19) 2 60
(4,3) 2 60 (15,22) 2 60
(4,5) 2 60 (16,8) 2 40
(4,11) 3 10 (16,10) 2 60
(5,4) 2 60 (16,17) 2 40
(5,6) 10 60 (16,18) 2 60
(5,9) 2 20 (17,10) 6 60
(6,2) 2 60 (17,16) 2 40
(6,5) 10 60 (17,19) 6 60
(6,8) 2 60 (18,7) 4 60
(7,8) 2 60 (18,16) 2 60
(7,18) 4 60 (18,20) 16 60
(8,6) 2 60 (19,15) 2 60
(8,7) 2 60 (19,17) 6 60
(8,9) 4 60 (19,20) 4 30
(8,16) 2 40 (20,18) 16 60
(9,5) 2 20 (20,19) 4 30
(9,8) 4 60 (20,21) 4 60
(9,10) 2 20 (20,22) 6 10
(10,9) 2 20 (21,20) 4 60
(10,11) 3 10 (21,22) 8 60
(10,15) 4 20 (21,24) 4 60
(10,16) 2 60 (22,15) 2 60
(10,17) 6 60 (22,20) 6 10
(11,4) 3 10 (22,21) 8 60
(11,10) 3 10 (22,23) 2 10
(11,12) 4 60 (23,14) 2 10
(11,14) 3 15 (23,22) 2 10
(12,3) 4 60 (23,24) 2 60
(12,11) 4 60 (24,13) 4 60
(12,13) 14 60 (24,21) 4 60
(13,12) 14 60 (24,23) 2 60
Table 8: Sioux Falls: arc data
18

OD(t) Demand OD(t) Demand
(1,20,0) 25 (20,1,0) 25
(1,20,1) 40 (20,1,1) 40
(1,20,2) 55 (20,1,2) 55
(1,20,3) 30 (20,1,3) 30
(1,20,4) 20 (20,1,4) 20
(1,24,0) 20 (24,1,0) 20
(1,24,1) 35 (24,1,1) 35
(1,24,2) 50 (24,1,2) 50
(1,24,3) 25 (24,1,3) 25
(1,24,4) 15 (24,1,4) 15
(7,20,0) 15 (20,7,0) 15
(7,20,1) 30 (20,7,1) 30
(7,20,2) 45 (20,7,2) 45
(7,20,3) 20 (20,7,3) 20
(7,20,4) 10 (20,7,4) 10
(7,24,0) 15 (24,7,0) 15
(7,24,1) 30 (24,7,1) 30
(7,24,2) 45 (24,7,2) 45
(7,24,3) 20 (24,7,3) 20
(7,24,4) 10 (24,7,4) 10
Table 9: Sioux Falls: OD information (scenario A)
Arc Capacity Arc Capacity
(8,16) 40 (16,8) 40
(16,17) 40 (17,16) 40
(19,20) 30 (20,19) 30
Table 10: Sioux Falls: arc capacities for scenario B
19

Pair OD(t) Demand
(7,20,0) 60
(7,20,1) 120
(7,20,2) 180
(7,20,3) 80
(7,20,4) 40
Table 11: Sioux Falls: demand for OD pair (7,20) (scenario C)
0 40 6.3853 49.81
1 64 5.7933 118.56
2 67 4.9266 188.85
5 70 2.9340 406.69
10 77 1.5637 788.17
20 78 0.7630 1595.60
30 79 0.4816 2426.13
50 80 0.3279 4157.09
Table 12: Sioux Falls results (scenario A)
0 40 32.5518 55.39
1 75 20.2456 139.39
2 92 14.4513 236.32
5 123 7.0109 577.66
10 149 3.5233 1250.41
20 168 1.4144 2811.61
30 171 0.9505 4479.48
50 172 0.5053 7894.81
Table 13: Sioux Falls results (scenario B)
20

0 40 38.6861 56.49
1 76 24.6751 142.80
2 93 16.0157 240.55
5 130 7.5075 593.06
10 159 3.6671 1301.51
20 186 1.5922 2970.06
30 193 0.9470 4763.01
50 192 0.6316 8447.17
Table 14: Sioux Falls results (scenario C)
21

(1,20,0) 20.692 (20,1,0) 20.253
(1,20,1) 21.563 (20,1,1) 21.009
(1,20,2) 22.013 (20,1,2) 22.646
(1,20,3) 23.342 (20,1,3) 23.270
(1,20,4) 23.401 (20,1,4) 23.374
(1,24,0) 14.914 (24,1,0) 14.956
(1,24,1) 17.550 (24,1,1) 17.157
(1,24,2) 20.783 (24,1,2) 20.839
(1,24,3) 23.439 (24,1,3) 23.031
(1,24,4) 23.419 (24,1,4) 23.487
(7,20,0) 16.000 (20,7,0) 16.250
(7,20,1) 16.333 (20,7,1) 16.997
(7,20,2) 17.212 (20,7,2) 18.559
(7,20,3) 17.461 (20,7,3) 19.239
(7,20,4) 17.447 (20,7,4) 19.373
(7,24,0) 16.087 (24,0) 16.087
(7,24,1) 16.910 (24,1) 17.117
(7,24,2) 19.283 (24,2) 19.683
(7,24,3) 20.659 (24,3) 21.546
(7,24,4) 21.416 (24,4) 21.549
Table 15: Sioux Falls: minimum costs after 50 iterations (scenario A)
22

(1,20,0) 24.806 (20,1,0) 22.078
(1,20,1) 27.135 (20,1,1) 25.708
(1,20,2) 29.592 (20,1,2) 29.207
(1,20,3) 31.631 (20,1,3) 30.134
(1,20,4) 31.910 (20,1,4) 30.124
(1,24,0) 18.752 (24,1,0) 16.348
(1,24,1) 23.608 (24,1,1) 21.325
(1,24,2) 25.624 (24,1,2) 24.833
(1,24,3) 26.465 (24,1,3) 25.430
(1,24,4) 26.589 (24,1,4) 25.009
(7,20,0) 16.750 (20,7,0) 17.920
(7,20,1) 19.263 (20,7,1) 19.892
(7,20,2) 20.586 (20,7,2) 20.350
(7,20,3) 20.989 (20,7,3) 20.709
(7,20,4) 20.114 (20,7,4) 20.467
(7,24,0) 16.882 (24,0) 16.873
(7,24,1) 20.448 (24,1) 20.944
(7,24,2) 26.144 (24,2) 26.343
(7,24,3) 28.737 (24,3) 27.814
(7,24,4) 28.108 (24,4) 26.843
Table 16: Sioux Falls: minimum costs after 50 iterations (scenario B)
23

OD(t) Minimum cost OD(t) Minimum costs
(1,20,0) 25.357 (20,1,0) 22.078
(1,20,1) 27.843 (20,1,1) 25.714
(1,20,2) 30.383 (20,1,2) 29.081
(1,20,3) 32.000 (20,1,3) 30.228
(1,20,4) 31.835 (20,1,4) 30.204
(1,24,0) 18.600 (24,1,0) 16.348
(1,24,1) 23.785 (24,1,1) 21.323
(1,24,2) 25.707 (24,1,2) 24.835
(1,24,3) 26.205 (24,1,3) 25.437
(1,24,4) 26.224 (24,1,4) 25.019
(7,20,0) 17.686 (20,7,0) 17.920
(7,20,1) 20.312 (20,7,1) 19.892
(7,20,2) 21.872 (20,7,2) 20.363
(7,20,3) 22.567 (20,7,3) 20.681
(7,20,4) 22.138 (20,7,4) 20.462
(7,24,0) 17.461 (24,0) 16.873
(7,24,1) 21.144 (24,1) 20.940
(7,24,2) 26.823 (24,2) 26.318
(7,24,3) 29.808 (24,3) 27.819
(7,24,4) 29.658 (24,4) 26.880
Table 17: Sioux Falls: minimum costs after 50 iterations (scenario C)
24

APPENDIX
ALGORITHM DCAPLOAD (x)
input: x = {xt
s}s∈Wt,t≤T [strategic volume vector]
output: {πst
jk(x)}(j,k)∈A, s∈Wt [conditional arc probabilities]
C = {Ct
s(x)}s∈Wt,t≤T [vector of strategic costs]
INITIALIZATION
for t = 0 to T do
for s ∈ Wt
do
Ct
s := 0 [cost of strategy s]
zstt
q(s) := xt
s [volume at origine node qt(s)]
τst
q(s) := 1 [probability of accessing qt(s)]
for t′
= t to T do
for jt′ ∈ V (jt′ 6= qt(s)) do
for t′′
= t to t′
do
zst′t′′
j := 0 [strategic volume at node jt′
endfor if j has been reached at instant t′′
]
τst′
j := 0 [conditional probability of
for kt′+cjk
∈ Est′
j do accessing node jt′ using strategy s]
vst′
jk := 0 [strategic volume on arc (jt′ , kt′+cjk
)]
πst′
jk := 0 [conditional probability of accessing
endfor arc (jt′ , kt′+cjk
) using strategy s]
endfor
endfor
endfor
endfor
LOADING PHASE
for t = 0 to T do [time topological order]
for jt ∈ V do
for kt+cjk
∈ j+
t do
ujk := ujk [residual capacity of arc (jt, kt+cjk
)]
endfor
uj := ∞ [(jt, jt+1) is uncapacitated]
for t′
= 0 to t do
zst
j := zst
j + zstt′
j [total strategic volume zst
j ]
S
tt′
:= {s ∈ W : zstt′
j > 0} [set of active strategies at jt according
endfor to instant of accessing node j]
for t′
= 0 to t do
for s ∈ S
tt′
do
25

E
stt′
j := Estt′
j [residual preference order]
zstt′
j := zstt′
j [residual strategic volume]
endfor
while S
tt′
6= ∅ do [begin of loading for S
tt′
]
K := ∅ [set of residual first choices]
for kt+cjk
∈ j+
t do [kt+cjk
= jt+1 if waiting arc]
dk := 0 [demand for node kt+cjk
]
Wk := ∅ [set of strategies having node
endfor kt+cjk
as residual first choice]
for s ∈ S
tt′
do
kt+cjk
:= E
stt′
j (1) [first residual choice of strategy s]
K := K ∪ {k} [construction of K
k = j if (jt, E
stt′
j (1)) is waiting arc]
Wk := Wk ∪ {s} [construction of Wk]
dk := dk + zstt′
j [construction of dk]
endfor
µ := max{dk/ujk : k ∈ K}
βtt′
:= min{1/µ, 1} [proportion of affected residual volume]
for k ∈ K do
for s ∈ Wk do
t′′
:= t + cjk [instant of accessing node k]
if k = j then
zst+1t′
j := zst+1t′
j + βtt′
zstt′
j [updating volume zst+1t′
j ]
else
zst′′t′′
k := zst′′t′′
k + βtt′
zstt′
j [updating volume zst′′t′′
k ]
endif
vst
jk := vst
jk + βtt′
z̄stt′
j [updating arc strategic volume vst
jk]
zstt′
j := (1 − βtt′
)zstt′
j [updating residual volume zstt′
j ]
πst
jk := vst
jk/zst
j [updating probability πst
jk]
endfor
ujk := ujk − βtt′
dk [updating residual capacity ujk]
if ujk = 0 then
for t′′
= 0 to t do
for s ∈ S
tt′′
do
E
stt′′
j := E
stt′′
j − {kt+cjk
} [updating residual preference set]
endfor
endfor
endif
endfor
if βtt′
= 1 then
S
tt′
= ∅ [end of loading for S
tt′
]
endif
endwhile
26

endfor
endfor
for s ∈ Wt
for t′
= t to T
for jt′ ∈ V
for kt′+cjk
∈ j+
t′ do
t′′
= t′
+ cjk [period of accessing node k]
Ct
s := Ct
s + cjkτst′
j πst′
jk [update of cost Ct
s]
τst′′
k := τst′′
k + τst′
j πst′
jk [update of probability τst′′
k ]
endfor
Ct
s := Ct
s + τst′
j πst′
j [update of cost Ct
s]
τ
s(t′+1)
j := τ
s(t′+1)
j + τst′
j πst′
j [update of probability τ
s(t′+1)
j ]
endfor
endfor
endfor
endfor
27

A Strategic Model For Dynamic Traffic Assignment

More Related Content

Similar to A Strategic Model For Dynamic Traffic Assignment (20)

More from Kelly Taylor (20)

Recently uploaded (20)

A Strategic Model For Dynamic Traffic Assignment