A New Schedule-Based Transit Assignment Model With Travel Strategies And Supply Uncertainties

Elsevier Editorial System(tm) for Transportation Research Part B
Manuscript Draft
Manuscript Number: TRB-D-13-00278R2
Title: A new schedule-based transit assignment model with travel strategies and supply uncertainties
Article Type: Research Paper
Keywords: User equilibrium; Schedule-based transit assignment; Strategy; Supply uncertainty
Corresponding Author: Dr. Wai Yuen Szeto, PhD
Corresponding Author's Institution: The University of Hong Kong
First Author: Younes Hamdouch
Order of Authors: Younes Hamdouch; Wai Yuen Szeto, PhD; Y. Jiang

Statement of contributions
A transit assignment model is useful in estimating or predicting how passengers utilize a
given transit system. In the literature of transit assignment studies, these models used
either frequency-based (static) or schedule-based (dynamic) approach to model transit
route choice. The optimal strategy approach is one of the commonly adopted
formulations in these approaches. However, most of existing related studies did not
consider the effect of uncertainty in transit networks on route choice.
In fact, due to supply side uncertainty, in-vehicle travel times and waiting times,
especially for buses and mini-buses, are highly uncertain. Studies such as Jackson and
Jucker (1982) and Szeto et al. (2011b) found that travel time uncertainty does affect the
route choice of passengers. It is essential to capture this realistic travel behavior into the
transit modeling framework. Therefore, transit assignment models have recently
emphasized the influence of uncertainties in frequency-based frameworks and their
transit network design applications (Yang and Lam, 2006; Li et al. 2008, 2009; Sumalee
et al., 2011, Szeto et al., 2011a, b). These transit assignment models can be used to study
aggregated stochastic effects of a specific line from a static perspective. However,
uncertainties exist in both vehicle running and dwelling process in line operation and the
schedule-based models provide a means to investigate uncertainties within the vehicle
process (Zhang et al., 2010). Hence, Zhang et al. (2010) developed a schedule-based
transit assignment model to capture the uncertainties. Nevertheless, they proposed a path-
based model and hence path enumeration or column generation is needed to obtain
solutions. Optimal strategies and hence the concept of set of attractive lines are also not
explicitly considered in their model.
The objective of the paper is to extend the schedule-based transit assignment model in
Hamdouch and Lawphongpanich (2008) to consider supply uncertainties in the transit
network and optimal strategies. This extension is not straightforward, as the resultant
problem is a stochastic and dynamic optimization problem. We propose an analytical
model that captures the stochastic nature of the transit schedules and in-vehicle travel
times due to road conditions, incidents or adverse weather. We adopt a mean variance
approach that can consider the covariance of travel time between links in a space time
graph but still lead to a robust transit network loading procedure when optimal strategies
are adopted. The method of successive averages (MSA) is adopted to solve the model.
Numerical studies are performed to illustrate the properties of the model and the
effectiveness of the algorithm. This paper differs from Zhang et al. (2010) in threefold.
First, this paper adopts a mean-variance approach to consider strategies while they adopt
effective travel cost as the factor affecting passengers’ line choice. Second, their model is
path-based and requires path enumeration and column generation, but ours is strategy-
based and relies on Bellman’s recursion principle to deal with network loading. Third, we
consider hard capacity constraints but they consider a chance constraint for dealing with
the capacity.
The contributions of this paper include the following:
1. Statement of contribution/potential impact

1. This paper proposes a schedule-based transit assignment model with the
consideration of both supply uncertainties and optimal strategies.
2. The proposed solution method does not rely on path enumeration or column
generation technique. The transit network loading procedure relies on the usage of
Bellman’s recursion principle, and is quite robust.
3. The model and the solution method allow us to evaluate the performance of
transit systems under supply uncertainties, assess the effectiveness of operational
strategies, and develop a larger model to plan transit schedules.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
A new schedule-based transit assignment model with
travel strategies and supply uncertainties
Younes Hamdoucha
, W.Y. Szetob∗
, Y. Jiangb
a
Department of Business Administration, United Arab Emirates University, Al Ain, UAE
b
Department of Civil Engineering, The University of Hong Kong, Pokfulam Road, Hong Kong, PR China
May 4, 2014
Abstract
This paper proposes a new scheduled-based transit assignment model. Unlike
other schedule-based models in the literature, we consider supply uncertainties and
assume that users adopt strategies to travel from their origins to their destinations.
We present an analytical formulation to ensure that on-board passengers contin-
uing to the next stop have priority and waiting passengers are loaded on a first-
come-first-serve basis. We propose an analytical model that captures the stochastic
nature of the transit schedules and in-vehicle travel times due to road conditions,
incidents, or adverse weather. We adopt a mean variance approach that can con-
sider the covariance of travel time between links in a space-time graph but still lead
to a robust transit network loading procedure when optimal strategies are adopted.
The proposed model is formulated as a user equilibrium problem and solved by an
MSA-type algorithm. Numerical results are reported to show the effects of supply
uncertainties on the travel strategies and departure times of passengers.
Keywords: User equilibrium; Schedule-based transit assignment; Strategy;
Supply uncertainty
1 Introduction
A transit assignment model is useful in estimating or predicting how passengers utilize
a given transit system. In the literature of transit assignment studies, these models
∗
Corresponding author, Email: ceszeto@hku.hk
1
*2. Manuscript
Click here to view linked References

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
used either the frequency-based (static) or the schedule-based (dynamic) approach to
model transit route choice. Similar to the traditional static user equilibrium assignment
models, frequency-based transit assignment models (Spiess and Florian, 1989; De Cea and
Fernandez, 1993; Cantarella, 1997; Lam et al., 1999, 2002; Kurauchi et al., 2003; Cepeda
et al., 2006; Schmöcker et al., 2009; Sumalee et al., 2009; Schmöcker et al., 2011; Cortés
et al., 2013; Trozzi et al., 2013; Szeto and Jiang, 2014) often assume that passengers
select transit routes to minimize their perceived expected travel cost, and departure time
is not the concern. These static transit assignment models are commonly adopted for the
strategic and long-term planning/evaluation of transit networks.
Schedule-based transit assignment models (Wilson and Nuzzolo, 2004; Poon et al.,
2004; Hamdouch and Lawphongpanich, 2008; Hamdouch et al., 2011; Zhang et al., 2010;
Nuzzolo et al., 2012) are typically dynamic and are better suited to short-term tran-
sit operations and service planning such as transit timetabling and vehicle scheduling.
In a schedule-based model, the temporal dimension is the most important part as it is
assumed that transit passengers choose not only their transit routes, but also their de-
parture times for minimizing their individual generalized cost. Researchers incorporate
this time dependent choice in different ways which is classified by Poon et al. (2004) as
(a) diachronic graph representation (Nuzzolo et al., 2001); (b) dual graph representation
(Moller-Pedersen, 1999); (c) forward star network formulation (Tong and Wong, 1998),
and; (d) space-time formulation (Nguyen et al., 2001; Hamdouch and Lawphongpanich,
2008; Hamdouch et al., 2011). In the last representation, the schedule-based transit net-
work is represented by a time-expanded graph. This graph has an explicit representation
of single runs and allows a more straightforward treatment of congestion when capacity
constraints are considered. Moreover, it can explicitly represent passenger movements
through the in-vehicle and waiting links in the space-time network. This representation
and the first one both consider space-time nodes and links. However, a time-expanded
network is built on a two dimension graph with one time axis and one space axis. A
diachronic network is built in a three dimension graph with two space axes and one time
axis.
To model the route choice, one commonly approach is to adopt the concept of optimal
strategy. In the frequency-based approach, the core idea for an optimal strategy is that a
traveler selects, at each node of the network, a set of attractive lines that allows him/her
to reach his/her destination at a minimum expected cost (Spiess and Florian, 1989; Wu
et al., 1994; Cepeda et al., 2006; Schmöcker et al., 2009). Different from the previous
static models, Hamdouch and Lawphongpanich (2008) developed a dynamic schedule-
based transit assignment where the choice of strategy is an integral part of user behavior.
2

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
In that study, passengers specified their individual travel strategy by providing, at each
transit station and each point in time, an ordered list of transit lines they preferred to use
to continue their own journey. For a given passenger, the user-preference set at each time-
expanded (TE) node collectively yielded a set of potential paths that departed from the
passenger’s origin at the same time and generally arrived at the destination at different
times. Also, when loading a transit vehicle at a station, on-board passengers continu-
ing to the next station remained on the vehicle and waiting passengers were loaded in a
first-come-first-serve (FCFS) basis. To explicitly consider vehicle capacities, the model
assigned the fail-to-board passengers to the wait arc to wait for their next preferred tran-
sit services with residual capacities. Hamdouch et al. (2011) extended the model in
Hamdouch and Lawphongpanich (2008) to differentiate the discomfort level experienced
by the sitting and standing passengers. Each class of passengers, grouped by their re-
maining journey lengths and times already spent on-board, was assigned success-to-sit,
success-to-stand, and failure-to-board probabilities. These probabilities were computed
by performing a dynamic network loading. The stimulus of a standing passenger to sit
increased with his/her remaining journey length and time already spent on-board. When
a vehicle was full, passengers unable to board must wait for the next vehicle to arrive.
The above studies do not consider the effect of the uncertainties of transit networks
on route choice. In fact, due to supply side uncertainties, in-vehicle travel times and
waiting times, especially for buses and mini-buses, are highly uncertain. Studies such as
Jackson and Jucker (1982) and Szeto et al. (2011b) found that travel time uncertainty
does affect the route choice of passengers. It is essential to capture this realistic travel
behaviour into the transit modelling framework. Therefore, transit assignment models
have recently emphasized the influence of uncertainties in the frequency-based framework
and their transit network design applications (Yang and Lam, 2006; Li et al., 2008, 2009;
Sumalee et al., 2011; Szeto et al., 2011b, 2013) as in traffic assignment (Shao et al., 2006;
Szeto et al. 2011a). These transit assignment models can be used to study the aggregated
stochastic effects of transit lines from a static perspective. However, uncertainties exist in
both the vehicle running and dwelling processes in line operation and the schedule-based
models provide means to investigate uncertainties within the vehicle processes (Zhang et
al., 2010). Hence, Zhang et al. (2010) developed a schedule-based transit assignment
model to capture the uncertainties, wherein they adopted the effective travel cost as
the factor affecting the route choice of passengers and considered chance constraint for
dealing with the capacity. Nevertheless, they proposed a path-based model and hence
path enumeration or column generation is needed to obtain solutions. Optimal strategies
and hence the concept of the set of attractive lines are also not explicitly considered in
3

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
their model.
The objective of the paper is to extend the schedule-based transit assignment model
proposed by Hamdouch and Lawphongpanich (2008) to consider supply uncertainties in
the transit network, optimal strategies, and hard capacity constraints. This extension is
not straightforward, as the resultant problem is a stochastic and dynamic optimization
problem. We propose an analytical model that captures the stochastic nature of the
transit schedules and in-vehicle travel times due to road conditions, incidents, or adverse
weather. We adopt a mean variance approach that can consider the covariance of travel
time between links in a space-time graph but still lead to a robust transit network loading
procedure when optimal strategies are adopted. We formulate the problem as a user
equilibrium problem. We adopt a user equilibrium (UE) framework instead of a stochastic
user equilibrium (SUE) framework because of the following:
i) It is easier to illustrate the concept of travel strategy and the model formulation
clearly and analyze the model properties without being smeared by other factors
such as the perception error of passengers on travel costs.
ii) SUE transit assignment models require a probabilistic choice model to depict the
travel choice behavior of passengers. However, a realistic choice model always has
some limitations. For example, the Probit model used in SUE transit assignment
(e.g., Nielsen, 2000 and Nielsen and Frederiksen, 2006) relies on simulation that
suffers from computational burden. The Logit model used in transit assignment
models (e.g., Lam et al., 1999; 2002) suffer from the path overlapping issue. Solving
C-Logit (Cassetta et al., 1996) and other path-based choice models often requires
a path set generation or path enumeration algorithm, and an efficient link based
algorithm that obviates the path set generation or enumeration procedure has not
yet been developed to solve these models.
iii) A UE framework has a good mathematical property that allows the dynamic pro-
gramming technique to be used during the solution process. The technique does not
rely on path set generation or path enumeration during that process.
The proposed model is formulated as a variational inequality (VI) model, unlike the
nonlinear complementarity problem (NCP) model (e.g., Lo et al. 2003) and the fixed
point (FP) model (Cantarella, 1997) in the transit assignment literature. Nevertheless,
according to Nagurney (1993), our proposed VI model can be reformulated into an NCP
model and a FP model so that other solution techniques developed for solving NCP and
FP models can be used. In this paper, the method of successive averages, which is often
4

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
used to solve FP models, is adopted to solve our model. Numerical studies are given
to illustrate the effects of supply uncertainties, vehicle capacity, and early/late arrival
penalty parameters on travel strategies and/or departure times of passengers. The effects
of the value of travel time variability (which was termed by Jenelius (2012) and Brjesson
et al. (2012)) or equivalently the degree of risk aversion (termed by Jackson and Jucker
(1982)) are also investigated.
The contributions of this paper include the following:
i) This paper proposes a schedule-based transit assignment model with the consider-
ation of both supply uncertainties and optimal strategies.
ii) The solution method developed does not rely on any path enumeration or column
generation technique. The transit network loading procedure relies on the usage of
Bellman’s recursion principle, and is quite robust.
iii) The model and the solution method allow us to evaluate the performance of transit
systems under supply uncertainties, assess the effectiveness of operational strategies
under these uncertainties, and develop a larger model to plan transit schedules to
cope with these uncertainties.
For the remainder, Section 2 presents the network representation, notations, and as-
sumptions of the proposed model. Section 3 depicts how to determine the mean and
variance travel times and arrival probabilities. Travel strategies and the computation of
the effective strategy costs are described in Section 4. Section 5 formulates the transit
assignment problem as a variational inequality and proposes an MSA-based solution al-
gorithm. Section 6 presents numerical results and Section 7 discusses the applicability of
our model in real-life applications. Finally, Section 8 concludes the paper.
2 Network representation, notations, and assump-
tions
2.1 Network representation
Consider a transit network that consists of nodes and arcs. Nodes include origins, desti-
nations, and station nodes where a transit vehicle stops to load and unload passengers.
Arcs are used to connect nodes. They consist of walk arcs and in-vehicle arcs. An example
is given in Figure 1 that displays a transit system with two origin nodes q and o, two
5

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
destination nodes r and y, and three transit lines l1, l2, and l3. Nodes labeled a, b, c,
and d are station nodes. In this example, there are four walk arcs: two access arcs (q, a)
and (o, b), and two egress arcs (d, r) and (c, y). The remaining arcs correspond to route
segments of the three transit lines. As an example, Line 1 or l1 begins its route at node
a, travels to node b, then to node c, and finally terminates at node d. Thus, {a, b, c, d} is
the route sequence associated with line l1.
→ → →
→
→
Figure 1: A small network with three transit lines
In the transit network, the number next to each arc (j, k) is the “travel time” Tjk. For
walk arcs, Tjk is assumed to be constant (Tjk = tjk) and represents the time to walk from
j to k. When (j, k) corresponds to a transit-line segment, {Tjk} is assumed to follow a
discrete distribution with the probabilities Pjk(t), a mean E(Tjk) = µjk and a variance
V ar(Tjk) = σ2
jk.
As in Hamdouch and Lawphongpanich (2008) and Hamdouch et al. (2011), we use a
time-expanded (TE) approach to model transit supply in a schedule-based setting. The
time horizon is represented as a set of discrete points of the form Γ = {t0, t0 + δ, t0 +
2δ, · · · , t0 + nδ}, where δ is the duration of each time interval and Ω = {0, 1, 2, 3, · · · , n}
is the set of time intervals. All time related variables in the model are then specified as
a multiple of δ. In general, each node j in the transit network is expanded into multiple
nodes jτ , where τ ∈ Ω, in the TE network. Similarly, an in-vehicle arc (j, k) in the
transit network is expanded into multiple in-vehicle arcs (jτ , kτ′′ ) where τ′′
denotes the
time interval to reach node k. Similarly, arcs (q, k) and (j, r) are expanded into multiple
access arcs (qτ , k(τ+Tqk)), and egress arcs (jτ , r(τ+Tjr)), respectively. These two types of arcs
represent walking from an origin to a station and from another station to a destination,
respectively. In addition, there are arcs of the form (jτ , jτ+1) that represents passengers
having to wait at station j from time τ to (τ + 1).
6

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
2.2 Notations
Sets
N set of nodes (with i, j, k ∈ N)
A set of arcs (with a ∈ A)
Ω set of time intervals (with τ, τ′
, τ′′
∈ Ω, where τ stands
for the time interval considered; τ′
and τ′′
, respectively,
represent the arrival time interval not later and earlier
than the current time interval, i.e., τ′
≤ τ,τ′′
≥ τ
L set of transit lines (with l ∈ L)
Lj set of transit lines that traverse node j
Ljk set of lines traversing on arc (j, k) with Njk its cardinality
I+
(jτ ) set of successor nodes for the time-expanded node jτ
I−
(jτ ) set of predecessor nodes for the time-expanded node jτ
S(q,r) set of strategies for OD pair (q, r) (with s ∈ S(q,r))
Es,τ
j user-preference set for strategy s, node j, and time τ
Wτ,τ′
j set of passengers who have reached node j at time τ′
≤ τ
Wτ,0,1
j set of passengers who have continuance priority at node j
at time τ and travel on the run with the highest
probability to reach node j at time τ
{j1(l), j2(l), . . . , jNl
(l)} set of route sequence nodes associated with line l
{DT1,jn(l), DT2,jn(l), . . . , DTMl,jn(l)} set of the departure/arrival times at transit node jn(l)
with the first subscript is for run
Mjτ ,l set of runs of line l that have positive probabilities
to reach node j at time τ
7

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Parameters
δ duration of a time interval
dg
(q,r) travel demand for OD pair (q, r) and group g
[t−
(q,r)(g), t+
(q,r)(g)] desired arrival time interval for OD pair (q, r) and group g
tjk travel time for access/egress arc a = (j, k) ∈ A
Tjk random travel time for in-vehicle arc a = (j, k) ∈ A
Y lm
jk random travel time for arc (j, k) and transit line lm
under no effects from the previous arc
µjk mean travel time for in-vehicle arc a = (j, k) ∈ A
σ2
jk variance of the travel time for in-vehicle arc a = (j, k) ∈ A
φ, φl
0, φl
m coefficients used in the autoregressive model
cjk, cl
jk constants used in the autoregressive model
Pjk(t) probability that the travel time Tjk is equal to t
Ml number of runs for transit line l (with 1 ≤ m ≤ Ml)
Nl number of transit nodes for line l (with 1 ≤ n ≤ Nl)
Pm,jn(l)(τ) probability that the departure/arrival time for the mth
transit vehicle at node jn(l) is equal to τ
uτ
jk transit capacity for arc (j, k) at time τ
uτ
jkml transit capacity for the mth
run of line l
serving arc (j, k) at time τ
vτ
jk transit fare on arc (j, k) at time τ
ηg
1 early arrival penalty (in monetary units) for group g
ηg
2 late arrival penalty (in monetary units) for group g
ηg
3 value of travel time variability for group g
η4 crowding penalty (in monetary units)
γtravel value of time for travelling
γwait value of time for waiting
eτ,g
jr late penalty cost for egress arc (jτ , r(τ+Tjr)) and group g
b
eτ
jk crowding cost function on arc (j, k) at time τ
Decision variables
xs,τs
(q,r,g) number of passengers for OD pair (q, r) and group g
assigned to strategy s and who leaves q at time τs
(starting time of strategy s)
X strategy assignment (SA) vector (with its components xs,τs
(q,r,g))
8

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Functions of decision variables
fs,τ,τ′′
jkml number of passengers using strategy s and traveling on arc
(jτ , kτ′′ ) and run m of line l
fs,τ,τ′′
jk number of passengers using strategy s and traveling on arc
(jτ , kτ′′ ) (fs,τ,τ′′
jk =
∑
l∈Lj
∑
m∈Ml
fs,τ,τ′′
jkml )
fs,τ
jk number of passengers using strategy s and traveling on arc
(j, k) at time τ (fs,τ
jk =
∑
τ′′≥τ
fs,τ,τ′′
jk )
fτ
jk number of passengers traveling on arc (j, k) at time τ
πs,τ,τ′′
jk probability that a passenger using strategy s travels on arc (jτ , kτ′′ )
πs,τ
jk probability that a passenger using strategy s
accesses arc (j, k) at time τ (πs,τ
jk =
∑
τ′′≥τ
πs,τ,τ′′
jk )
πs,τ
j probability that a passenger using strategy s
waits at node j from time τ to time τ + 1
zs,τ,τ′
jml number of passengers using strategy s, travelling on run m of line l,
and having reached node j at time τ′
≤ τ; τ′
= 0 represents
the case that these passengers have continuance priority
zs,τ
jml number of passengers using strategy s and travelling on run m of line l
who reach node j at time τ
zs,τ,τ′
j number of passengers using strategy s
and having reached node j at time τ′
≤ τ; τ′
= 0 represents
the case that these passengers have continuance priority
zs,τ
j number of passengers using strategy s
who reach node j at time τ (zs,τ
j =
∑
l∈Lj
∑
m∈Ml
zs,τ
jml)
Y s,τ
j random variable representing the node selected from the preference set Es,τ
j
Cg
jk random cost associated with link (j, k) for passenger group g
Cs,τs
(q,r,g) cost for passenger group g reaching destination r from origin q
using strategy s at time τs
C vector of strategy costs (with its components Cs,τs
(q,r,g))
ECs,τs
(q,r,g) effective cost for passenger group g reaching destination r from origin q
using strategy s at time τs
EC vector of effective strategy costs (with its components ECs,τs
(q,r,g)).
9

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
2.3 Model assumptions
Seven main assumptions are made within the model as in the literature and are presented
below.
i) The demand for each OD pair and group g is assumed to be fixed. However, network
uncertainties are incorporated in the model through the stochastic nature of the
transit schedules and in-vehicle travel times due to road conditions, incidents or
adverse weather.
ii) The dwelling time (time for passengers to board and alight) is negligible and the
mean travel time, µjk, denotes the difference between the scheduled departure times
(arrival times) at stations j and k.
iii) When loading a vehicle, on-board passengers continuing to the next station remain
on the transit vehicle and waiting passengers are loaded on a First-Come-First-Serve
(FCFS) basis.
iv) Transit fares are collected based on arcs. This assumption is reasonable for the cases
of additive or distance-based fare structures. (i.e., the fares are directly proportional
to the travel distance or time.) However, if the fares are not directly proportional
to the travel distance or the fares are non-additive over arcs (such as the zone-based
fare), one can construct a direct in-vehicle arc between each pair of connected nodes
in the TE network. The drawback is an increase in the number of arcs in the TE
network (see, e.g., Lo et al. (2003) for more details).
v) All wait arcs have zero fares, zero penalties, and infinite capacities.
vi) All access and egress arcs have zero fares and infinite capacities. However, there
are penalties associated with egress arcs to account for lost opportunities associated
with arrivals outside the desired interval. Typically, these penalties are different for
various groups because of their different values of time or trip purposes.
For egress arcs (jτ , r(τ+Tjr)), one form of such penalty is as follows:
eτ,g
jr = ηg
1 max{0, t−
(q,r)(g) − (τ + Tjr)} + ηg
2 max{0, (τ + Tjr) − t+
(q,r)(g)}. (1)
vii) All in-vehicle arcs have transit fares and transit capacities. In addition, there is a
discomfort penalty for having too many passengers on board. For example, such a
discomfort function can be defined as follows:
b
eτ
jk(fτ
jk) = η4
(
fτ
jk
uτ
jk
)2
, (2)
10

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
where uτ
jk is the capacity of arc (j, k) at time τ, and fτ
jk is the total number of
passengers on arc (j, k) at time τ.
3 Travel time and arrival probabilities
Let Lj ⊂ L be the set of transit lines that traverse node j. For each line l ∈ Lj, node j
can be viewed as j = jn(l), (1 ≤ n ≤ Nl). To consider correlation between arcs belonging
to the same transit line l, we adopt a first-order discrete autoregressive (DAR(1)) model
(see Brockwell and Davis, 1991; Biswas and Song (2009) ) that accounts for the travel
time’ effects of an arc on its subsequent one within transit line l. For each 2 ≤ n ≤ Nl −1,
given {Tjn−1(l)jn(l)} = {Tij} with the probabilities Pij(t), a mean µij and a variance σ2
ij,
{Tjn(l)jn+1(l)} = {Tjk} can be determined as as a mixture distribution of {Tij} and {Yjk}:
Tjk = (Tij, φ) ∗ (Yjk, 1 − φ) + cjk, (3)
with cjk = φ(E(Yjk) − µij) and the marginal probability function given by:
P(Tjk + cjk = t) = φP(Tij = t) + (1 − φ)P(Yjk = t), (4)
where {Yjk} are i.i.d with given probabilities, a mean E(Yjk), and a variance V ar(Yjk).
Yjk represents the travel time for arc (j, k) under no effects from the previous arc (i, j)
and cjk is a constant added in the model to ensure that the mean travel time µjk is not
affected by the mean travel time of arc (i, j) and the correlation between arcs (i, j) and
(j, k) is measured by the variance travel time. φ (0 ≤ φ < 1) is the coefficient in the
autoregressive model that measures the effects of the previous arc (i, j) on the travel time
Tjk. If φ is close to 0, then the travel time Tjk is not affected by the previous arc (i, j)
but as φ approaches 1, the travel time Tjk gets a larger contribution from the previous
arc (i, j).
Using (4), we have:
µjk =
∑
t
tP(Tjk = t) + cjk (5)
=
∑
t
t(φP(Tij = t) + (1 − φ)P(Yjk = t)) + cjk
= φµij + (1 − φ)E(Yjk) + φ(E(Yjk) − µij)
= E(Yjk).
Also, we can compute the variance and covariance terms (see Appendix A):
σ2
jk = φσ2
ij + (1 − φ)V ar(Yjk) + φ(1 − φ)
(
µij − E(Yjk)
)2
(6)
Cov(Tin(l)in+1(l), Tin+n′ (l)in+n′+1(l)) = φn′
σ2
in(l)in+1(l), 1 ≤ n ≤ Nl − 2, 1 ≤ n′
≤ Nl − 1 − n.
11

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
In the case we have overlapping lines from node j to k, the travel time Tjk depends on not
only the travel time of the previous arc (i, j) but also the travel times of all transit lines
serving arc (j, k). If Ljk denotes the set of all transit lines lm traversing arc (j, k) with Njk
its cardinality and Tl
jk is the travel time associated with arc (j, k) and transit line l ∈ Ljk,
{Tl
jk} can be determined as a mixture distribution of {Tl
ij}, {Y l1
jk }, {Y l2
jk }, · · · {Y
lNjk
jk }:
Tl
jk = (Tl
ij, φl
0) ∗ (Y l1
jk , φl
1)∗, (Y l2
jk , φl
2) ∗ · · · ∗ (Y
lNjk
jk , φl
lNjk
) + cl
jk,
with cl
jk = E(Y l
jk)−φl
0E(Tl
ij)−
Njk
∑
m=1
φl
mE(Y lm
jk ) and the marginal probability function given
by
P(Tl
jk + cl
jk = t) = φl
0P(Tl
ij = t) +
Njk
∑
m=1
φl
mP(Y lm
jk = t),
where {Y lm
jk } are i.i.d with given probabilities, a mean E(Y lm
jk ), and a variance V ar(Y lm
jk ).
For each 1 ≤ m ≤ Njk, Y lm
jk represents the travel time for arc (j, k) and transit line lm
under no effects from the previous arc (i, j), and φl
m (0 ≤ m < Njk) are the coefficients
in the autoregressive model with
Njk
∑
m=0
φl
m = 1. Following the proofs of (5) and (6), we can
show that
E(Tl
jk) = E(Y l
jk),
V ar(Tl
jk) = φl
0V ar(Tl
ij) +
Njk
∑
m=1
φl
mV ar(Y lm
jk ) +
Njk
∑
m=1
φl
0φl
m
(
E(Tl
ij) − E(Y lm
jk )
)2
+
Njk
∑
m=1
Njk
∑
m′=m+1
φl
mφl
m′
(
E(Tlm
jk ) − E(Y
lm′
jk )
)2
, and
Cov(Tin(l)in+1(l), Tin+n′ (l)in+n′+1(l)) = (φl
0)n′
σ2
in(l)in+1(l), 1 ≤ n ≤ Nl − 2, 1 ≤ n′
≤ Nl − 1 − n.
To illustrate the discrete autoregressive model, Table 1 displays the input data of all
in-vehicle arcs in Figure 1. Using equations (5) and (6) and setting φ = 0.3, we can
compute all probability distributions and all mean and variance/covariance terms (see
Tables 2 and 3).
Using the probabilities Pjk(t), we can calculate the arrival probabilities Pm,jn(l)(τ)
associated with the mth
transit vehicle at node jn(l). We first set all arrival probabilities
12

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Line 1 Line 2 Line 3
Tab Ybc Ycd Tac Tbd
Time Prob Time Prob Time Prob Time Prob Time Prob
4 0.25 3 0.25 3 0.1 9 0.25 8 0.1
5 0.5 5 0.5 4 0.15 10 0.5 9 0.15
6 0.25 7 0.25 5 0.4 11 0.25 10 0.4
6 0.35 11 0.35
Table 1: Input data for in-vehicle arcs in Figure 1
Tab Tbc Tcd Tac Tbd
Time Prob Time Prob Time Prob Time Prob Time Prob
4 0.25 3 0.175 3 0.1225 9 0.25 8 0.1
5 0.5 4 0.075 4 0.1275 10 0.5 9 0.15
6 0.25 5 0.5 5 0.43 11 0.25 10 0.4
6 0.075 6 0.2675 11 0.35
7 0.175 7 0.0525
Table 2: Probability distributions for in-vehicle arcs in Figure 1
Pm,jn(l)(τ) to 0 and then update them recursively as follows:
Pm,jn(l)(τ) =







1 if n = 1 and τ is the starting time
of the mth
run of line l;
∑
τ′<τ
Pm,jn−1(l)(τ′
)Pjn−1(l)jn(l)(τ − τ′
) otherwise.
(7)
Using equation (7), we can obtain the probability distributions of all transit lines in
Figure 1 (as shown in Table 4).
In our example, the time horizon is [7h00, 8h00], Ω = {0, 1, 2, · · · , 60}, δ = 1 min. We
assume line l1 has 4 runs (Ml1 = 4) and lines l2 and l3 have 3 runs (Ml2 = Ml3 = 3).
Associated with each line l, there are fixed departure times, DTm,j1(l), at which each mth
transit vehicle must leave its starting station j1(l). At node a, there are four departure
times (DT1,a(l1) = 5, DT2,a(l1) = 15, DT3,a(l1) = 25 and DT4,a(l1) = 35) corresponding to
transit line l1 and three departure times (DT1,a(l2) = 5, DT2,a(l2) = 20 and DT3,a(l2) = 35)
corresponding to transit line l2. At node b, there are three departure times (DT1,b(l3) = 10,
13

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Tab Tbc Tcd Tac Tbd
Mean 5 5 5 10 10
Var/Cov
Tab 0.5 0.15 0.045
Tbc 0.15 1.55 0.465
Tcd 0.045 0.465 1.095
Tac 0.5
Tbd 0.9
Table 3: Mean and variance/covariance terms for in-vehicle arcs in Figure 1
DT2,b(l3) = 20 and DT3,b(l3) = 30) associated with transit line l3.
4 Travel strategies and effective strategy costs
In this section, we show how the concept of travel strategies is adopted in the TE networks
with supply uncertainties and illustrate how to compute the effective cost of a strategy.
4.1 Travel strategies
As in previous studies (Hamdouch and Lawphongpanich, 2008 and Hamdouch et al.,
2011), we assume that passengers use strategies when travelling. To specify a strategy
(denoted as s), passengers must provide, at each node jτ , a preference set Es,τ
j of sub-
sequent nodes at which they want to reach via a transit line, walking, or waiting at a
station. The order in which nodes are listed in Es,τ
j gives the passengers’ preference, i.e.,
the first node in the set is the most preferred and the last is the least. To each node k
in the preference set that can be reached via a walking or a wait arc, we associate a time
interval index representing the actual time interval to reach node k. To each node k that
can be reached via an in-vehicle arc, we associate an index representing the corresponding
transit line. It is important to note that this strategy definition is different from the one
used in previous studies with fixed timetables. Indeed, while we can identify the actual
time passengers reach node k via a walking or a wait arc, the time to reach node k via
an in-vehicle arc is random and passengers can only include transit line indices in their
preference set. For example, Table 5 displays one valid strategy s1
for OD pair (q, r).
14

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
a b c d a c b d
Time Pr Time Pr Time Pr Time Pr Time Pr Time Pr Time Pr Time Pr
Run 1 5 1 9 0.25 12 0.044 15 0.005 5 1 14 0.25 10 1 18 0.1
10 0.5 13 0.106 16 0.019 15 0.5 19 0.15
11 0.25 14 0.206 17 0.058 16 0.25 20 0.4
15 0.288 18 0.119 21 0.35
16 0.206 19 0.181
17 0.106 20 0.224
18 0.044 21 0.195
22 0.122
23 0.058
24 0.017
25 0.002
Run 2 15 1 19 0.25 22 0.044 25 0.005 20 1 29 0.25 20 1 28 0.1
20 0.5 23 0.106 26 0.019 30 0.5 29 0.15
21 0.25 24 0.206 27 0.058 31 0.25 30 0.4
25 0.288 28 0.119 31 0.35
26 0.206 29 0.181
27 0.106 30 0.224
28 0.044 31 0.195
32 0.122
33 0.058
34 0.017
35 0.002
Run 3 25 1 29 0.25 32 0.044 35 0.005 35 1 44 0.25 30 1 38 0.1
30 0.5 33 0.106 36 0.019 45 0.5 39 0.15
31 0.25 34 0.206 37 0.058 46 0.25 40 0.4
35 0.288 38 0.119 41 0.35
36 0.206 39 0.181
37 0.106 40 0.224
38 0.044 41 0.195
42 0.122
43 0.058
44 0.017
45 0.002
Run 4 35 1 39 0.25 42 0.044 45 0.005
40 0.5 43 0.106 46 0.019
41 0.25 44 0.206 47 0.058
45 0.288 48 0.119
46 0.206 49 0.181
47 0.106 50 0.224
48 0.044 51 0.195
52 0.122
53 0.058
54 0.017
55 0.002
Table 4: Probability distributions for transit lines in Figure 1
For a passenger using s1
, the order of nodes in the user-preference set at node a5, i.e.,
[bl1 , cl2 , a6], indicates that the passenger prefers Line 1 over Line 2 and Line 2 over waiting.
Using this strategy, there are several directed paths emanating from q0 and reaching node
r at different times. The arrival time at the destination depends on the probabilities to
access various lines at nodes a, b, c, and d as well as the probabilities associated with the
random travel times Tab, Tac, Tbc, Tbd, and Tcd.
The effective cost of a strategy s depends directly on the arc probabilities πs,τ
jk and
πs,τ
j associated with in-vehicle and wait arcs at time τ. The procedure for computing
this strategy cost comprises two main steps. In the first step, a stochastic loading of
the TE network is performed according to a given strategy assignment vector X and
15

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Strategy: s1
Node Pref. Node Pref. Node Pref. Node Pref. Node Pref.
q0 [a5] a5 [bl1
, cl2
, a6] b9 [cl1
, b10] c12 [dl1
, c13] d15 [r20]
a15 [bl1
, a16] b10 [cl1
, dl3
, b11] c13 [dl1
, c14] d16 [r21]
a20 [cl2
, a21] b11 [cl1
, b12] c14 [dl1
, c15] d17 [r22]
a25 [bl1
, a26] b19 [cl1
, b20] c15 [dl1
, c16] d18 [r23]
a35 [bl1
, cl2
, a36] b20 [cl1
, dl3
, b21] c16 [dl1
, c17] d19 [r24]
b21 [cl1
, b22] c17 [dl1
, c18] d20 [r25]
b29 [cl1
, b30] c18 [dl1
, c19] d21 [r26]
b30 [cl1
, dl3
, b31] c22 [dl1
, c23] d22 [r27]
b31 [cl1
, b32] c23 [dl1
, c24] d23 [r28]
b39 [cl1
, b40] c24 [dl1
, c25] d24 [r29]
b40 [cl1
, b41] c25 [dl1
, c26] d25 [r30]
b41 [cl1
, b42] c26 [dl1
, c27] d26 [r31]
c27 [dl1
, c28] d27 [r32]
c28 [dl1
, c29] d28 [r33]
c32 [dl1
, c33] d29 [r34]
c33 [dl1
, c34] d30 [r35]
c34 [dl1
, c35] d31 [r36]
c35 [dl1
, c36] d32 [r37]
c36 [dl1
, c37] d33 [r38]
c37 [dl1
, c38] d34 [r39]
c38 [dl1
, c39] d35 [r40]
c42 [dl1
, c43] d36 [r41]
c43 [dl1
, c44] d37 [r42]
c44 [dl1
, c45] d38 [r43]
c45 [dl1
, c46] d39 [r44]
c46 [dl1
, c47] d40 [r45]
c47 [dl1
, c48] d41 [r46]
c48 [dl1
, c49] d42 [r47]
d43 [r48]
d44 [r49]
d45 [r50]
d46 [r51]
d47 [r52]
d48 [r53]
d49 [r54]
d50 [r55]
d51 [r56]
d52 [r57]
d53 [r58]
d54 [r59]
d55 [r60]
Table 5: One travel strategy for OD pair (q, r)
16

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
is an extension to the one proposed by Hamdouch and Lawphongpanich (2008). The
stochastic loading process computes the arc flows, fs,τ
jk , and the arc probabilities, πs,τ
jk (X)
and πs,τ
j (X), by processing TE nodes one at a time and in topological and chronological
(T&C) order, i.e., a node with no predecessor and the smallest time interval index is
processed first. Given all the arc flows and probabilities, the second step computes the
effective strategy cost using a mean variance approach. This step involves scanning TE
nodes in reverse T&C order and applying Bellman’s generalized recursion. Note that
this procedure is different from the one adopted in previous studies with fixed timetables.
Using a mean variance approach, Bellman’s recursion is essential to account for both
expected and variance cost terms in calculating the effective cost of a strategy.
4.2 Stochastic loading process
In loading the TE network, we ensure that, at each node jτ , the summation of the prob-
abilities associated with outgoing arcs in the preference set Es,τ
j are equal to one:
∑
k∈Es,τ
j −{jτ+1}
πs,τ
jk (X) + πs,τ
j (X) = 1, ∀jτ , ∀s. (8)
Consider processing node j at time τ. For each line l ∈ Lj, node j can be viewed
as j = jn(l), (1 ≤ n ≤ Nl). Let Mjτ ,l be the set of runs of line l that have positive
probabilities to reach node j at time τ (ordered from the highest probability to the
smallest):
Mjτ ,l = {m ∈ Ml|Pm,jn(l)(τ) > 0}.
Note that due to the variability in travel time, more than one run of the same line l can
reach node j at time τ, resulting in bus bunching (Bartholdi and Eisenstein, 2012). This
bunching issue occurs when at least one of the transit vehicles of line l is unable to keep
to its schedule and therefore reaches node j as one or more other vehicles of the same
transit line l at the same time τ. For example, Table 4 shows a bunching issue at node
d25 with Md25,l1 = {2, 1}, P2,d(25) = 0.005, and P1,d(25) = 0.002.
For each line l such that 1 < n < Nl (i.e., jn(l) is neither the starting nor the ending
node of line l) and for each strategy s such that the first choice in the user-preference
set Es,τ
j (1) = {jn+1(l)}, the passengers using strategy s on arc (jn−1(l), j) have priority
to board line l on arc (j, jn+1(l)). In case we have more than one run of the same line l
that reach node j at time τ, it is intuitively to give priority to the passengers on the run
with the highest probability first. This assumption can be relaxed by loading together all
17

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
the passengers on the runs belonging to Mjτ ,l. Therefore, the first priority class, Wτ,0,1
j ,
consists of all the passengers who have continuance priority at node jτ and travelling on
run Mjτ ,l(1) with the highest probability:
Wτ,0,1
j = ∪s ∪l∈Lj
{
zs,τ,0
jml , Es,τ
j (1) = {jn+1(l)}, m = Mjτ ,l(1)
}
, (9)
where zs,τ,0
jml denotes the number of passengers using strategy s, travelling on run m of line
l, and having continuance priority at node jτ :
zs,τ,0
jml =
∑
τc<τ
πs,τc,τ
ij zs,τc
iml , i = jn−1(l). (10)
For each line l ∈ Lj and for each strategy s such that Es,τ
j (1) = {jn+1(l)}, the passen-
gers using strategy s on arc (jn−1(l), j) and travelling on run m of line l,
∑
τc<τ
fs,τc,τ
jn−1(l)jml,
have priority to board line l on arc (j, jn+1(l)) and the flows fs,τ,τ′′
jjn+1(l)ml and fs,τ
jjn+1(l)ml are
computed as follows:
fs,τ,τ′′
jjn+1(l)ml = Pjjn+1(l)(τ′′
− τ)
∑
τc<τ
fs,τc,τ
jn−1(l)jml
= Pjjn+1(l)(τ′′
− τ)zs,τ,0
jml , ∀τ′′
> τ (11)
fs,τ
jjn+1(l)ml =
∑
τ′′>τ
fs,τ,τ′′
jjn+1(l)ml.
Then, the residual capacities of all arcs (j, jn+1(l)), uτ
jjn+1(l)ml, are updated (uτ
jjn+1(l)ml =
uτ
jjn+1(l)ml−
∑
s
fs,τ
jjn+1(l)ml) and the process ends for class Wτ,0,1
j . We repeat the same process
for the priority classes Wτ,0,m′
j for m′
= 2, ..., maxl{|Mjτ ,l|}.
After loading all on-board passengers who want to continue their journey in the same
transit vehicle, the process loads passengers who arrive at node j at time τ on various
transit lines and want to transfer to other transit lines as well as those who have been
waiting at node jτ . To enforce the FCFS rule, we classify these passengers according
to their arrival times at node j. We denote zs,τ,τ′
j as the number of passengers using
strategy s at node jτ and having reached node j at time τ′
≤ τ and group all flows into
a class Wτ,τ′
j restricted to passengers having reached node j at time τ′
:
Wτ,τ′
j = ∪s
{
zs,τ,τ′
j , Es,τ
j ̸= ∅
}
,
where zs,τ,τ′
j =
∑
l∈Lj
∑
m∈Mjτ ,l
zs,τ,τ′
jml and zs,τ,τ′
jml is computed according to the following recur-
sion:
zs,τ,τ′
jml =







πs,τ−1
j zs,τ−1,τ′
jml if τ′
≤ τ − 1
∑
τc<τ
πs,τc,τ
ij zs,τc
iml if τ′
= τ, i = jn−1(l) and
(j, Es,τ
j (1)) ∈ l′
̸= l,
(12)
18

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
In equation (12), the first term denotes the passengers who reach node j before time τ −1
and the second term denotes those who reach node j at time τ.
As for the priority classes, we load passengers on the runs with the highest probabilities
in the sets Mjτ ,l and then we repeat the process for the subsequent runs following the
descendent order of probabilities. In loading passengers belonging to the classes Wτ,τ′
j (τ′
≤
τ), the process loads, in the FCFS order, the passengers who, according to their strategy
s, prefer to access arcs (jτ , kτ′′ ) for all τ′′
> τ, i.e., the process loads those passengers who
arrive earlier at time τ1 ≤ τ (zs,τ,τ1
j ) before those who arrive later at time τ2 > τ1, τ2 ≤ τ
(zs,τ,τ2
j ) until the remaining capacity of the arc is exhausted (uτ
jk = 0). Those who cannot
be loaded must use wait arc (jτ , jτ+1).
Once all the arcs emanating from jτ are loaded, the arc probabilities are computed as
follows:
πs,τ,τ′′
jk =
∑
l∈Lj
∑
m∈Mjτ ,l
fs,τ,τ′′
jkml
∑
l∈Lj
∑
m∈Mjτ ,l
zs,τ
jml
=
fs,τ,τ′′
jk
zs,τ
j
, (13)
πs,τ
jk =
∑
τ′′≥τ
πs,τ,τ′′
jk =
fs,τ
jk
zs,τ
j
, and (14)
πs,τ
j =
zs,τ
j − fs,τ
jk
zs,τ
j
. (15)
The stochastic loading procedure will be explained in detail using the example in Figure 2
which is built based upon Figure 1. Not all nodes and links are shown for the sake of
clarity. We focus on the loading process at nodes q10, o15, a15, b20, and c25. The loading
process starts at node q10 where 10 passengers using strategy s1
and 5 passengers using
s3
are loaded onto access arc (q10, a15). Thus, fs1,10
qa = 10, fs3,10
qa = 5, πs1,10
qa = πs3,10
qa = 1,
and πs1,10
q = πs3,10
q = 0. At node o15, 30 passengers using strategy s2
are loaded onto
access arc (o15, b20) and we get fs2,15
ob = 30, πs1,15
ob = 1, and πs2,15
q = 0. At node a15, the 10
passengers using strategy s1
and the 5 passengers using s3
want to board the second run
of line 1 and access arc (a, b) at time 15 (P2,a(l1)(15) = 1, zs1,15
a2l1
= 10, and zs3,15
a2l1
= 5). The
time to reach node b depends on the probabilities associated with the random travel time
Tab. From Tables 2 and 4, we know that Pab(4) = 0.25, Pab(5) = 0.5, and Pab(6) = 0.25.
19

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
∞
]
[
]
[
]
[
Figure 2: An example of stochastic loading
Therefore, from equation (10), we obtain the following:
fs1,15,19
ab2l1
= Pab(4)zs1,15
a2l1
= 0.25(10) = 2.5,
fs3,15,19
ab2l1
= Pab(4)zs3,15
a2l1
= 0.25(5) = 1.25,
fs1,15,20
ab2l1
= Pab(5)zs1,15
a2l1
= 0.5(10) = 5,
fs3,15,20
ab2l1
= Pab(5)zs3,15
a2l1
= 0.5(5) = 2.5,
fs1,15,21
ab2l1
= Pab(6)zs1,15
a2l1
= 0.25(10) = 2.5,
fs3,15,21
ab2l1
= Pab(6)zs3,15
a2l1
= 0.25(5) = 1.25,
fs1,15
ab = fs1,15,19
ab2l1
+ fs1,15,20
ab2l1
+ fs1,15,21
ab2l1
= 10,
fs3,15
ab = fs3,15,19
ab2l1
+ fs3,15,20
ab2l1
+ fs3,15,21
ab2l1
= 5,
πs1,15
ab = πs3,15
ab = 1,
πs1,15
a = πs3,15
a = 0,
πs1,15,19
ab = πs3,15,19
ab = 0.25,
πs1,15,20
ab = πs3,15,20
ab = 0.5,
πs1,15,21
ab = πs3,15,21
ab = 0.25.
20

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
When the process undergoes node b20, there are two classes of passengers: W20,0,1
b =
{zs1,20
b2l1
= 5, zs3,20
b2l1
= 2.5} and W20,20
b = {zs2,20
b = 30}. The process starts by loading the
passengers belonging to W20,0,1
b , where 5 passengers using strategy s1
and 2.5 passengers
using s3
board the second run of line 1 and travel on arcs (b20, c23), (b20, c24), (b20, c25),
(b20, c26), and (b20, c27) with probabilities 0.175, 0.075, 0.5, 0.075, and 0.175, respectively
(see Table 2). Thus, we get fs1,20,25
bc2l1
= Pbc(5)zs1,20
b2l1
= 0.5(5) = 2.5, fs3,20,25
bc2l1
= Pbc(5)zs3,20
b2l1
=
0.5(2.5) = 1.25, fs1,20
bc2l1
=
26
∑
τ′′=23
Pbc(τ′′
− 20)fs1,20,τ′′
bc2l1
= 5, and fs3,20
bc2l1
=
26
∑
τ′′=23
Pbc(τ′′
−
20)fs3,20,τ′′
bc2l1
= 2.5. Then, the residual capacity, u20
bc2l1
, is updated (u20
bc2l1
= 20−7.5 = 12.5).
The next step is to load passengers belonging to the class W20,20
b . Among the 30 passen-
gers using strategy s2
and belonging to this class, 10 passengers travel on the second run
of line 3 and 12.5 passengers travel on the second run of line 1. The times to reach nodes
c and d depend on the probabilities associated with the variables Tbc and Tbd, respectively.
The remaining passengers 7.5 use wait arc (b20, b21).
Finally, at node c25, 1.25 passengers using s3
alight from line 1 to take egress arc
(c25, y30). Therefore, only passengers using strategies s1
and s2
continue on line 1 and
access arcs (c25, d28), (c25, d29), (c25, d30), (c25, d31), and (c25, d32) with probabilities 0.1225,
0.1275, 0.43, 0.2675, and 0.0525, respectively (see Table 2). Relevant arcs flows and
probabilities for this stochastic loading example are displayed in Table 6.
(q, a) (o, b) (a, b) (b, c) (b, d) (c, d)
fs,τ
jk fs1
,10
qa = 10 fs2
,15
ob = 30 fs1
,15
ab = 10 fs1
,20
bc = 5 fs2
,20
bd = 10 fs1
,25
cd = 2.5
fs3
,10
qa = 5 fs3
,15
ab = 5 fs3
,20
bc = 2.5 fs2
,25
cd = 6.25
fs2
,20
bc = 12.5
fs,τ,τ′′
jkml fs1
,15,20
ab = 5 fs1
,20,25
bc = 2.5 fs2
,20,30
bd = 4
fs3
,15,20
ab = 2.5 fs3
,20,25
bc = 1.25
fs2
,20,25
bc = 6.25
πs,τ
jk πs1
,10
qa = 1 πs2
,15
ob = 1 πs1
,15
ab = 1 πs1
,20
bc = 1 πs2
,20
bd = 0.33 πs1
,25
cd = 1
πs3
,10
qa = 1 πs3
,15
ab = 1 πs3
,20
bc = 1 πs2
,25
cd = 1
πs2
,20
bc = 0.42
πs,τ,τ′′
jk πs1
,10,15
qa = 1 πs2
,15,20
ob = 1 πs1
,15,20
ab = 0.5 πs1
,20,25
bc = 0.5 πs2
,20,30
bd = 0.13
πs3
,10,15
qa = 1 πs3
,15,20
ab = 0.5 πs3
,20,25
bc = 0.5
πs2
,20,25
bc = 0.21
πs,τ
j πs1
,10
q = 0 πs2
,15
o = 0 πs1
,15
a = 0 πs1
,20
b = 0 πs1
,20
b = 0 πs1
,25
c = 0
πs3
,10
q = 0 πs3
,15
a = 0 πs3
,20
b = 0 πs3
,20
b = 0 πs2
,25
c = 0
πs2
,20
b = 0.25 πs2
,20
b = 0.25
Table 6: Stochastic loading process at nodes q10, o15, a15, b20, and c25
21

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
4.3 Effective strategy cost
In our model, passengers are dealing with two types of randomness when deciding on the
strategy to travel from their origins to their destinations. The first type of randomness
is due to the possibility to fail to board a vehicle as a result of limited transit capacities.
At each node jτ , the node selected from the preference set Es,τ
j is random and depends
on the residual capacities of the transit vehicles passing though j at time τ. The second
type of randomness comes from the in-vehicle arc travel times, Tjk, that follow a discrete
distribution with the probabilities Pjk(t), a mean µjk and a variance σ2
jk. To take into
account of these two types of uncertainties, a mean variance cost function is used to
model the passengers’ averseness to both failure to board a vehicle and link travel time
variability.
At each node jτ , let Y s,τ
j be the random variable representing the node selected from
the preference set Es,τ
j and Cg
jY s,τ
j
the random cost associated with link (jτ , (Y s,τ
j )τ+Tjk
)
and group g:
Cg
jY s,τ
j
=
{
γtravelTjk + vτ
jk + eτ,g
jk + b
eτ
jk(fτ
jk) if Y s,τ
j = k ∈ Es,τ
j − {jτ+1};
γwait if Y s,τ
j = jτ+1.
Using a mean variance approach, the effective cost of a strategy s (according to a
strategy assignment vector X) can be determined as
ECs,τs
(q,r,g)(X) = E(Cs,τs
(q,r,g)(X)) + ηg
3V ar(Cs,τs
(q,r,g)(X)), (16)
where τs
is the starting time of strategy s.
For a given triplet (j, r, g), let
Cs,τ
(j,r,g)(X) be the cost for reaching node r from node jτ using strategy s.
ECs,τ
(j,r,g)(X) be the effective cost for reaching node r from node jτ using strategy s.
ECs,τ
(j,r,g)(X) = E(Cs,τ
(j,r,g)(X)) + ηg
3V ar(Cs,τ
(j,r,g)(X)).
The effective costs ECs,τ
(j,r,g)(X) are computed by scanning TE nodes in reverse T&C order
starting from destination r and applying Bellman’s equation.
Ending Conditions at node r:
i) Set E(Cs,τ
(r,r,g)) = 0, ∀τ ∈ Ω.
ii) Set V ar(Cs,τ
(r,r,g)) = 0, ∀τ ∈ Ω.
22

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Recursions at node jτ (j = jn(l), j ̸= q):
Bellman’s equation for the random cost Cs,τ
(j,r,g)(X) at node jτ is given as follows:
Cs,τ
(j,r,g)(X) =





γtravelTjk + vτ
jk
+eτ,g
jk + b
eτ
jk(fτ
jk) + C
s,τ+Tjk
(k,r,g) (X) if Y s,τ
j = k ∈ Es,τ
j − {jτ+1};
γwait + Cs,τ+1
(j,r,g)(X) if Y s,τ
j = jτ+1.
In the above expression, the first case represents the cost associated with on-board
passengers that consists of the travel cost of link (jτ , kτ+Tjk
) plus the cost for reaching
node r from node kτ+Tjk
, C
s,τ+Tjk
(k,r,g) (X). The travel cost of link (jτ , kτ+Tjk
) includes the
travel time, transit fare, the penalty eτ,g
jk as well as the penalty, b
eτ
jk(fτ
jk), for being in
a crowded vehicle. The second case represents the cost associated with waiting that
comprises the travel cost of link (jτ , jτ+1), γwait, and the cost for reaching node r from
node jτ+1, Cs,τ+1
(j,r,g)(X).
From the above formulation, the expected cost E(Cs,τ
(j,r,g)(X)) can be calculated as
E(Cs,τ
(j,r,g)(X)) =
∑
k∈Es,τ
j −{jτ+1}
πs,τ
jk (X)
∑
τ′′>τ
Pjk(τ′′
− τ)
[
γtravel(τ′′
− τ) + vτ
jk + eτ,g
jk
+b
eτ
jk(fτ
jk) + E(Cs,τ′′
(k,r,g)(X))
]
+πs,τ+1
j (X)
[
γwait + E(Cs,τ+1
(j,r,g)(X))
]
.
Using µjk =
∑
τ′′>τ
(τ′′
− τ)Pjk(τ′′
− τ) and setting
ϕs,τ,g
k = γtravelµjk + vτ
jk + eτ,g
jk + b
eτ
jk(fτ
jk) +
∑
τ′′>τ
Pjk(τ′′
− τ)E(Cs,τ′′
(k,r,g)(X)),
ϕs,τ,g
j = γwait + E(Cs,τ+1
(j,r,g)(X)),
the expected cost E(Cs,τ
(j,r,g)(X)) can be expressed as
E(Cs,τ
(j,r,g)(X)) =
∑
k∈Es,τ
j −{jτ+1}
πs,τ
jk (X)ϕs,τ,g
k + πs,τ+1
j (X)ϕs,τ,g
j .
For the variance strategy cost, we use the formula for variance decomposition, V ar(X1) =
E(V ar(X1|X2)) + V ar(E(X1|X2)), where X1 = Cs,τ
(j,r,g)(X) and X2 = Y s,τ
j . Therefore, we
23

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
obtain
V ar(Cs,τ
(j,r,g)(X)) =
∑
k∈Es,τ
j −{jτ+1}
πs,τ
jk (X)ψs,τ,g
k + πs,τ+1
j (X)ψs,τ,g
j
+
∑
k∈Es,τ
j −{jτ+1}
πs,τ
jk (X)(ϕs,τ,g
k )2
+ πs,τ+1
j (X)(ϕs,τ,g
j )2
−
( ∑
k∈Es,τ
j −{jτ+1}
πs,τ
jk (X)ϕs,τ,g
k + πs,τ+1
j (X)ϕs,τ,g
j
)2
,
where ψs,τ,g
j = V ar(γwait + Cs,τ+1
(j,r,g)(X)) = V ar(Cs,τ+1
(j,r,g)(X)) and
ψs,τ,g
k = V ar(γtravelTjk + vτ
jk + eτ,g
jk + b
eτ
jk(fτ
jk) + Cs,τ+Tjk
(k,r,g) (X))
= γ2
travelV ar(Tjk) + V ar(Cs,τ+Tjk
(k,r,g) (X)) + 2γtravelCov(Tjk, Cs,τ′′
(k,r,g)(X))
= γ2
travelσ2
jk +
∑
τ′′>τ
Pjk(τ′′
− τ)
(
V ar(Cs,τ′′
(k,r,g)(X)) + 2γtravelCov(Tjk, Cs,τ′′
(k,r,g)(X))
)
.
Cov(Tjk, Cs,τ′′
(k,r,g)(X)) =
∑
n′∈N′
s
Cov(Tjn(l),jn+1(l), Tjn′ (l),jn′+1(l)),
where Ns
is the set of nodes included in the user-preference sets of strategy s and
N′
s = {n′
: n + 1 ≤ n′
≤ Nl − 1, jn′+1(l) ∈ Ns
}.
Determining the effective cost of strategy s:
ECs,τs
(q,r,g)(X) = E(Cs,τs
(q,r,g)(X)) + ηg
3V ar(Cs,τs
(q,r,g)(X)),
= γtraveltqj + E(Cs,τ
(j,r,g)(X)) + ηg
3V ar(Cs,τ
(j,r,g)(X)),
where Es,τs
q = {jτ }, tqj is the walking time of access arc (qτs , jτ ) and E(Cs,τ
(j,r,g)(X)) and
V ar(Cs,τ
(j,r,g)(X)) are available from previous recursions.
5 User equilibrium
A strategic assignment vector X∗
is in a user equilibrium if no passenger has any incentive
to change his or her strategy based on effective strategy costs. X∗
is in a user equilibrium
if and only if X∗
solves the following variational inequality (denoted as VI[EC(X), X]):
EC(X∗
)T
(X − X∗
) ≥ 0, ∀X ∈ X, (17)
where EC(X) is a vector of the effective strategy costs associated with X and X is the
set of all feasible SA vectors:
X = {X :
∑
s∈S(q,r)
xs,τs
(q,r,g) = dg
(q,r), ∀(q, r, g)}. (18)
24

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
This is an extension of the strategy-based equilibrium conditions to stochastic networks,
where we replace an expected strategy cost function by an effective strategy cost function
obtained through a mean-variance approach.
5.1 Computation of an optimal strategy
In finding a strategic equilibrium solution, we need to compute, for each triplet (q, r, g),
an optimal strategy s∗
(q,r,g) with the least effective cost given (or in response to) the current
strategy assignment X:
ECs∗,τs∗
(q,r,g) (X) = min
s∈S(q,r),τs
ECs,τs
(q,r,g)(X)
= min
s∈S(q,r),τs
E(Cs,τs
(q,r,g)(X)) + ηg
3V ar(Cs,τs
(q,r,g)(X)).
As in previous work with fixed timetables, the construction of the optimal strategy s∗
is based on dynamic programming and uses the information (strategic flows in the classes
Wτ,0
j and Wτ,τ′
j (τ′
≤ τ)) generated by the stochastic loading process.
Since the computation of the effective cost ECs∗,τs∗
(q,r,g) (X) involves the arc probabili-
ties associated with the optimal (unknown) strategy being constructed, these probabili-
ties have to be computed in reverse T&C order. The resulting procedure resembles the
stochastic loading process described in Section 4.2 with the small difference that the flow
corresponding to the optimal strategy being computed is set to zero. This micro-loading
phase (loading of zero or virtual flow) faces the same challenge occurred in the determin-
istic case. Indeed, since stochastic loading is performed in reverse T&C order, one might
be unaware of the priority status of the virtual flow at loading times. To make up for
this, we consider two situations:
i) The virtual (zero) flow arrives at node jτ with continuance priority and the micro-
loading is performed over the set Wτ,0
j ∪{s∗
} yielding the effective cost ECs∗,τ,0
(j,r,g)(X) =
E(Cs∗,τ,0
(j,r,g)(X)) + ηg
3V ar(Cs∗,τ,0
(j,r,g)(X)), where Cs∗,τ,0
(j,r,g)(X) is the cost for reaching desti-
nation r from node jτ assuming that the passengers using the optimal strategy, s∗
,
arrive at node jτ with continuance priority.
ii) The virtual (zero) flow arrives at node j at time τ′
= 1, 2, . . . , τ and tries to
board transit line l ar node j = jn(l). The micro-loading is then performed
over the sets Wτ,0
j , Wτ,1
j , · · · Wτ,τ′−1
j and Wτ,τ′
j ∪ {s∗
} yielding the effective cost
ECs∗,τ,τ′
(j,r,g) (X) = E(Cs∗,τ,τ′
(j,r,g) (X))+ηg
3V ar(Cs∗,τ,τ′
(j,r,g) (X)), where Cs∗,τ,τ′
(j,r,g) (X) is the cost for
reaching destination r from node j assuming that the passengers using the optimal
strategy, s∗
, arrive at node j at time τ′
, where τ′
= 1, 2, . . . , τ.
25

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
The user-preference set Es∗,τ
j and the effective costs ECs∗,τ,τ′
(j,r,g) (τ′
= 0, 1, . . . , τ) are
computed by scanning TE nodes in reverse T&C order and applying Bellman’s generalized
recursion.
Ending Conditions:
For the current destination r associated with s∗
,
i) Set Es∗,τ
r = ∅, ∀τ ∈ Ω.
ii) Set E(Cs∗,τ,τ′
(r,r,g) ) = 0, ∀τ ∈ Ω, and τ′
= 0, . . . , τ.
iii) Set V ar(Cs∗,τ,τ′
(r,r,g) ) = 0, ∀τ ∈ Ω, and τ′
= 0, . . . , τ.
For the destination b
r ̸= r not covered by s∗
,
i) Set Es∗,τ
b
r = ∅, ∀τ ∈ Ω.
ii) Set E(Cs∗,τ,τ′
(b
r,r,g) ) = ∞, ∀τ ∈ Ω, and τ′
= 0, . . . , τ.
iii) Set V ar(Cs∗,τ,τ′
(b
r,r,g) ) = ∞, ∀τ ∈ Ω, and τ′
= 0, . . . , τ.
Recursions at node jτ (j = jn(l), j ̸= q):
To compute the user-preference set Es∗,τ
j and the effective costs at node jτ , we
first determine Bellman’s equations for the expected and variance costs at node jτ .
Following section 3.3, the equation for the expected cost is:
E(Cs∗,τ,τ′
(j,r,g) (X)) =
∑
k∈Es∗,τ
j −{jτ+1}
πs∗,τ
jk (X)ϕs∗,τ,g
k + πs∗,τ+1
j (X)ϕs∗,τ,τ′,g
j ,
where ϕs∗,τ,τ′,g
j = γwait + E(Cs∗,τ+1,τ′
(j,r,g) (X)) and
ϕs∗,τ,g
k = γtravelµjk + vτ
jk + eτ,g
jk + b
ejk(fτ
jk)
+
∑
τ′′≥τ
Pjk(τ′′
− τ)





E(Cs∗,τ′′,0
(k,r,g) (X)), if (j, k) and (k, kτ′′
1 ) belong
to same transit line,
E(Cs∗,τ′′,τ′′
(k,r,g) (X)), otherwise,
kτ′′
1 is the first element in the preference set Es∗,τ′′
k .
For the variance cost, Bellman’s equation can be expressed as:
V ar(Cs∗,τ
(j,r,g)(X)) =
∑
k∈Es∗,τ
j −{jτ+1}
πs∗,τ
jk (X)ψs∗,τ,g
k + πs∗,τ+1
j (X)ψs∗,τ,τ′,g
j
+
∑
k∈Es∗,τ
j −{jτ+1}
πs∗,τ
jk (X)(ϕs∗,τ,g
k )2
+ πs∗,τ+1
j (X)(ϕs∗,τ,τ′,g
j )2
−
( ∑
k∈Es∗,τ
j −{jτ+1}
πs∗,τ
jk (X)ϕs∗,τ,g
k + πs∗,τ+1
j (X)ϕs∗,τ,τ′,g
j
)2
,
26

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
where ψs∗,τ,τ′,g
j = V ar(Cs∗,τ+1,τ′
(j,r,g) (X)) and
ψs∗,τ,g
k = γ2
travelσ2
jk
+
∑
τ′′≥τ
Pjk(τ′′
− τ)











V ar(Cs∗,τ′′,0
(k,r,g) (X))
+2γtravelC, if (j, k) and (k, kτ′′
1 ) belong
to same transit line,
V ar(Cs∗,τ′′,τ′′
(k,r,g) (X)), otherwise,
where C =
∑
n′∈N′
s
Cov(Tjn(l),jn+1(l), Tjn′ (l),jn′+1(l)), Ns∗
is the set of nodes included in
the user-preference sets of strategy s∗
and
N′
s∗ = {n′
: n + 1 ≤ n′
≤ Nl − 1, jn′+1(l) ∈ Ns∗
}.
After determining Bellman’s equations for the expected and variance costs at node
jτ , we must calculate the arc probabilities πs∗,τ
jk (X) and πs∗,τ
j (X). As mentioned
before, to make up for the unawareness of the priority status at the current time τ,
we consider two cases.
i) With continuance priority:
To consider this case, we should have at least one transit line l ∈ Lj such that
j = jn(l) and 1 < n < Nl. In this case, the virtual passenger using strategy
s∗
is added to the first class Wτ,0,1
j and has continuance priority to access arc
(j, j1
) where j1
is the first element of the set Es∗,τ
j . Node j1
= Es∗,τ
j (1) is
determined as follows:
j1
= arg min
l∈Lj:1<n<Nl
{ϕs∗,τ,g
jn+1(l) + ηg
3ψs∗,τ,g
jn+1(l)}.
After determining the first element of Es∗,τ
j , the effective cost, ECs∗,τ,0
(j,r,g)(X) is
calculated as follows:
ECs∗,τ,0
(j,r,g)(X) = πs∗,τ
jj1 (X)(ϕs∗,τ,g
j1 + ηg
3ψs∗,τ,g
j1 ) = ϕs∗,τ,g
j1 + ηg
3ψs∗,τ,g
j1 .
ii) Without continuance priority:
In this case, the virtual passenger using strategy s∗
can arrive at node j at
time τ′
, where τ′
= 1, . . . , τ. For each τ′
= 1, . . . , τ, we load passengers over
the sets Wτ,0,1
j , Wτ,1
j , · · · Wτ,τ′−1
j and the virtual passenger, zs∗,τ
j , is added to
the class Wτ,τ′
j . Then, the effective cost ECs∗,τ,τ′
(j,r,g) (X) is computed using the
27

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
recursion:
ECs∗,τ,τ′
(j,r,g) (X) = E(Cs∗,τ,τ′
(j,r,g) (X)) + ηg
3V ar(Cs∗,τ,τ′
(j,r,g) (X))
= πs∗,τ
j (X)
[
ϕs∗,τ,τ′,g
j + ηg
3
(
ψs∗,τ,τ′,g
j + (ϕs∗,τ,τ′,g
j )2
)]
+
∑
k∈Es∗,τ
j −{jτ+1}
πs∗,τ
jk
[
ϕs∗,τ,g
k + ηg
3
(
ψs∗,τ,g
k + (ϕs∗,τ,g
k )2
)]
−ηg
3
(
πs∗,τ
j ϕs∗,τ,τ′,g
j +
∑
k∈Es∗,τ
j −{jτ+1}
πs∗,τ
jk ϕs∗,τ,g
k
)2
,
where the optimal preference set Es∗,τ
j is the solution of the following combi-
natorial problem:
Es∗,τ
j = arg min
Es,τ
j ⊆I+(jτ )
{
πs,τ
j (X)
[
ϕs,τ,τ′,g
j + ηg
3
(
ψs,τ,τ′,g
j + (ϕs,τ,τ′,g
j )2
)]
+
∑
k∈Es,τ
j −{jτ+1}
πs,τ
jk
[
ϕs,τ,g
k + ηg
3
(
ψs,τ,g
k + (ϕs,τ,g
k )2
)]
−ηg
3
(
πs,τ
j ϕs,τ,τ′,g
j +
∑
k∈Es,τ
j −{jτ+1}
πs,τ
jk ϕs,τ,g
k
)2
}
.
Determining the minimum effective strategy cost for (q, r, g):
i) The user-preference set Es∗,τ
q : For each τ ∈ Ω, compute
j∗
τ∗
c
= arg min
jτc ∈I+(qτ )
{γtraveltqj + ECs∗,τc,τc
(j,r,g) (X)},
where tqj is the walking time of access arc (qτ , jτc ), τc = τ+tqj and ECs∗,τc,τc
(j,r,g) (X)
is the effective cost for reaching destination r from node jτc assuming that the
passengers using the optimal strategy, s∗
, arrive at node jτc at time τc.
Then, set Es∗,τ
q = {j∗
τ∗
c
} for all τ ∈ Ω.
ii) The effective strategy cost ECs∗,τ,τ
(q,r,g)(X): For each τ ∈ Ω,
ECs∗,τ,τ
(q,r,g)(X) = γtravelTqj∗ + EC
s∗,τ∗
c ,τ∗
c
(j∗,r,g) (X).
iii) The optimal starting time τs∗
:
τs∗
= arg min
τ∈Ω
{ECs∗,τ,τ
(q,r,g)(X)}.
iv) The effective cost for the optimal s∗
is determined as follows:
ECs∗,τs∗
(q,r,g) (X) = ECs∗,τs∗
,τs∗
(q,r,g) (X).
28

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
5.2 Solution algorithm
As in Hamdouch and Lawphongpanich (2008) and Hamdouch et al. (2011), we use the
method of successive averages (MSA) that generates strategies one at time by solving a
dynamic program. The convergence condition of the MSA was stated in Theorem 3 in
Cantarella (1997). This theorem states that if the existence and uniqueness conditions
mentioned in theorems 1 and 2 (including continuity and strictly monotonicity of cost
function) hold and the link cost-flow functions have a symmetric continuous Jacobian
with respect to link flows over the feasible solution set, then the MSA converges to the
equilibrium link flow vector. Because the cost function EC may fail to meet the symmetric
continuous Jacobian condition or strictly monotone condition, the convergence of the
iterates towards an equilibrium solution is not guaranteed. In this context, the method
must be viewed as a heuristic procedure.
The proposed algorithm first assumes that the TE network is not loaded with passen-
gers (i.e., zs,τ
j = 0 for all nodes within the TE network). With the empty TE network,
the corresponding optimal strategy for each OD pair (s∗
[0]) is computed by the optimal
strategy method described in Section 5.1 and is set to be the initial strategy set S[0]
for
network loading. Also, the initial strategic flow X[0]
is set to be the travel demand of the
corresponding OD pair and β is set to be zero. Then, the strategic flow X[0]
is loaded
using the stochastic loading process described in Section 4.2 for getting the corresponding
flow of passengers within the TE network at iteration β, zs,τ
j (X[β]
), and the effective cost
of the strategic flow, EC(X[β]
) is computed using the procedure illustrated in Section
4.3. Based on the current flow of passengers, an updated optimal strategy s∗
[β] can be
found and the strategic assignment vector for this step, Y [β]
with y
s∗[β]
(q,r,g) = dg
(q,r) and
ys
(q,r,g) = 0, ∀s ̸= s∗
[β], can be determined. With the current strategic assignment vector,
the convergence of the algorithm is checked by the following relative gap function (see
e.g., Hamdouch and Lawphongpanich, 2008):
g(x) =
EC(X[β]
)T
(X[β]
− Y [β]
)
EC(X[β])T X[β]
(19)
If the value of the above gap function is less than some predetermined tolerance, the
algorithm stops with X[β]
and S[β]
as the optimal strategic flow vector and strategy set,
respectively. Otherwise, X[β]
and S[β]
are updated for the next MSA step by the following
equation:
S[β+1]
= S[β]
∪ s∗
[β] (20)
X[β+1]
=
1
(β + 1)
(βX[β]
+ Y [β]
), β = 0, 1, 2, · · · (21)
29

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
The value of β is also increased by one. The updated strategy set and strategic flow are
then inputed to the dynamic loading process for getting the updated flow of passengers.
A flowchart of the proposed algorithm is shown in Figure 3.

Set forall
nodesintheTEnetwork.
Updatethestrategicflow
(X[E+1]
)byusingequation
(21)andstrategyset(S[E+1]
)
byusingequation(20).Set
E=E+1.

X[E]
,S[E]

Converge?
No
StopwithX[E]
and
S[E]
asthesolution
Yes
Findinganinitialstrategy
set andaninitial
strategicflowX[0]
.Set
.
Stochasticloadingand
effectivestrategycosts
(Section4)
X[0]
,S[0]
Findingtheoptimal
strategies, ,and
strategicassignment
vector,Y[E]
.
Figure 3: Flowchart of the proposed solution algorithm
At each iteration of the MSA, the stochastic loading process and the optimal strategy
computation are performed within the TE network. For each of these two processes, the
loading step is performed at most |I+
(jτ )| times for each TE node jτ , where I+
(jτ ) is
the set of successor nodes for jτ . It follows that the loading process is executed at most
∑
τ∈Ω
∑
j∈N |I+
(jτ )| = |Ω| × |A| times and that the total running time of the solution
algorithm (MSA) is polynomial.
30

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
6 Numerical examples
To illustrate our approach, we consider passengers travelling in the morning peak within
time period between 7:00 and 8:00 am. The network setting is the same as before. There
are four OD pairs, namely (q-y), (q-r), (o-y), and (o-r). The desired arrival intervals
for the passengers departing at q and o are [25, 35] and [30, 40], respectively. The travel
demands for these four OD pairs are [40, 35, 10, 20]. The transit fares for arcs (a, b), (b, c),
(c, d), (a, c), and (b, d) are 0.25, 0.50, 0.75, 1.00, and 0.50, respectively. The capacities for
lines 1, 2, and 3 are 20, 30, and 10, respectively. In addition, the penalties for early and
late arrivals are both taken as 0.1 per minute for all OD pairs, while the parameter for
the discomfort penalty is 0.2. Based on this setting, five examples were conducted and
are depicted in the following subsections. All the results were obtained by a laptop with a
Core i7-3770 CPU @ 3.4GHz and a 32GB RAM. The memory required for each example
is 115MB RAM.
6.1 Effects of value of travel time variance on the number of
utilized strategies and convergence
To show the effect of the value of travel time variance, η3 is set to be the same for all OD
pairs and varied from 0.0 to 1.0. Table 7 displays the relative gap value of the method
of successive averages (MSA) as well as the number of utilized strategies in selected
iterations. It can be seen that the algorithm successfully achieved a relative gap of 0.1%
or 0.001 for all three values of η3. By comparing the three cases, it is noticed that a larger
value of η3 requires more iterations to converge. Moreover, the value of the relative gap
fluctuated more significantly when η3 = 1.0. Probably, this is because the effective cost
function is not necessarily monotonic.
From the column for the number of utilized strategies, it can be seen that the algorithm
generated new strategies in early few iterations, and the number of utilized strategies
increased. Later, the algorithm updated the strategy set by removing unused strategies
at later iterations and hence the number of utilized strategies dropped to the minimum
and the number remained unchanged finally.
31

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Iteration number Number of utilized strategies Relative gap (%)
η3 = 0.0 η3= 0.5 η3 = 1.0 η3 = 0.0 η3 = 0.5 η3 = 1.0
1 8 8 8 7.81 5.05 7.24
2 9 9 8 8.38 8.94 13.53
3 10 11 10 4.35 3.59 2.92
4 10 11 10 5.21 3.81 6.91
5 10 11 12 2.97 2.94 2.34
10 9 10 11 1.83 1.81 1.98
20 8 10 12 0.81 1.34 1.24
35 7 10 10 0.48 0.45 8.26
41 6 9 9 0.08 1.36 1.00
200 7 9 0.34 0.28
292 7 9 0.10 0.18
323 9 0.10
Table 7: Iterates of MSA: utilized strategies and relative gap
6.2 Effects of the value of travel time variance on flow distribu-
tions
Tables 8, 9, and 10 show how strategies are utilized by passengers with various OD pairs
and under different values of η3, where all utilized strategies are depicted in Appendix B.
Take Table 10 as an example, when η3 = 1.0. Only two strategies, namely s2
(1.0) and
s7
(1.0) are adopted, where s2
(1.0) and s7
(1.0), respectively, denote the second and the
seventh strategies when η3 = 1.0. There are 40 passengers for OD pair (q-y). 30.08 pas-
sengers use strategy s2
(1.0) and 9.92 passengers use strategy s7
(1.0). The other strategies
are not used by these passengers.
For the same OD, the algorithm can generate different strategies that have the same
preference set at some TE nodes but have different preference sets for at least one TE node.
Take for example strategies s8(1.0) and s9(1.0) in Table 15, Appendix B. At nodes q0, c14-
c16 and d17-d23, the two strategies have the same preference sets. However, the preference
set at node a5 is [bl1 , cl2 , a6] for strategy s8(1.0) while it is [cl2 , bl1 , a6] for strategy s9(1.0).
Although the total number of utilized strategies (after considering all OD pairs) is
increasing with the increase of η3 when passengers are more risk aversive, it is not the
case for some OD pairs. For example, the number of utilized strategies of OD pairs (o-y)
and (q-y) remain unchanged despite the change in η3, while the number increases for the
32

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
OD Strategies
s1
(0.0) s2
(0.0) s3
(0.0) s4
(0.0) s5
(0.0) s6
(0.0)
q-y 19.52 20.48
q-r 35.00
o-y 10.00
o-r 10.16 9.84
Table 8: Strategy utilization: η3 = 0.0
OD Strategies
s1
(0.5) s2
(0.5) s3
(0.5) s4
(0.5) s5
(0.5) s6
(0.5) s7
(0.5)
q-y 12.85 27.15
q-r 14.73 20.27
o-y 10.00
o-r 14.50 5.50
OD Strategies
s1
(1.0) s2
(1.0) s3
(1.0) s4
(1.0) s5
(1.0) s6
(1.0) s7
(1.0) s8
(1.0) s9
(1.0)
q-y 30.08 9.92
q-r 20.20 13.45 1.35
o-y 10.00
o-r 3.83 13.21 2.96
33

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
q-y q-r o-y o-r
η3 = 0.0 6.05 8.55 3.20 6.19
η3 = 0.5 6.23 9.93 3.42 6.53
η3 = 1.0 6.47 10.42 3.56 6.65
Table 11: Effective costs under different values of η3
other OD pairs. This indicates that the effect of varying η3 on the number of utilized
strategies is different for various OD pairs. Moreover, the number of possible strategies
is different for various OD pairs. For example, for OD pair (o-y), there is only one line
connecting origin o and destination y. Comparing with OD pair (q-r), where there are
two lines at the first boarding node, the total number of possible strategies for OD pair
(o-y) is comparably smaller. Such issue is related to the design of transit networks and
implies that the risk aversive passengers can experience higher travel cost because of
limited choices of strategies (see Section 6.3).
6.3 Effects of the value of travel time variance on effective costs
Table 11 presents the optimal effective costs for all OD pairs when adjusting η3. As
expected, the effective cost increases with the increase of η3, since the value of variance
increases. However, the increment varies significantly for different OD pairs. For example,
for OD pair (q-r), the effective cost grows by 1.38 when η3 increases from 0.0 to 0.5, while
it only increases by 0.49 when η3 increases from 0.5 to 1.0. This is because passengers
utilize more strategies (which can be verified from the previous tables) to minimize the
effective cost when η3 is larger. In contrast, for OD pair (q-y), the increment of effective
cost is 0.18 when η3 increases from 0.0 to 0.5 while it is 0.24 when η3 increases from 0.5
to 1.0, since the number of utilized strategies remains unchanged when η3 increases.
6.4 Effects of the value of travel time variance and capacity on
departure and arrival times
Figures 4-9 demonstrate that the departure (and arrival) patterns are completely different
under various values of η3. Except for passengers of OD pair (o-y) constantly using one
strategy, other passengers either advance or postpone their departure times to switch to
a line with a lower variance when the value of travel time variance increases. For OD
pair (q-y), some depart at 7:10 to take the second run of line 1, when η3 = 0.0. When
34

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
7:00 7:05 7:10 7:15 7:20 7:25 7:30
0
5
10
15
20
25
30
35
40
45
50
Departure time of passengers
Number
of
passengers
OD(q,y)
OD(q,r)
OD(o,y)
OD(o,r)
Figure 4: Departure time of passengers (η3 = 0)
7:20 7:25 7:30 7:35 7:40 7:45
0
5
10
15
20
25
Arrival time of passengers
Number
of
passengers
OD(q,y)
OD(q,r)
OD(o,y)
OD(o,r)
Figure 5: Arrival time of passengers (η3 = 0)
considering variance, those passengers switch to a late departure strategy, for which they
depart at 7:15 but board another transit line (the second run of line 2). This is because
line 2 has a lower variance comparing with line 1. In contrast, some passengers of OD
pair (o-r) tend to depart earlier when η3 increases to 1.0, although such strategy induces
more early arrival penalty, implying that passengers are more willing to have a higher
arrival penalty to counteract the effect of the variance of travel time when they are more
risk aversive.
With the adjustment of departure time, the arrival time changes accordingly. More
importantly, when the level of risk aversion increases, more passengers arrive on time by
adjusting their departure times or selecting a different strategy to reach their individual
destination.
35

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
7:00 7:05 7:10 7:15 7:20 7:25 7:30
0
5
10
15
20
25
30
35
40
45
50
Number
of
passengers
OD(q,y)
OD(q,r)
OD(o,y)
OD(o,r)
Figure 6: Departure time of passengers (η3 = 0.5)
7:20 7:25 7:30 7:35 7:40 7:45 7:50
0
5
10
15
20
25
Number
of
passengers
OD(q,y)
OD(q,r)
OD(o,y)
OD(o,r)
Figure 7: Arrival time of passengers (η3 = 0.5)
To investigate the effects of early/late arrival penalties on the departure time choices,
we set η3 = 0.5 and both penalties for early and late arrivals are reduced by half. The
departure and arrival patterns are plotted in Figures 10 and 11.
Figures 6-7 (base case) and Figures 10-11 (case with reduced early/late arrival penal-
ties) illustrate that the penalties have different effects on different OD pairs. For example,
the departure and arrival times of the passengers of OD pair (o, y) are not affected, imply-
ing that their choices are irrespective of the arrival penalty values. This is because these
passengers can arrive at their destination within the desired arrival interval in the base
case, where the arrival penalties are high. Therefore, they can use their original strate-
gies, despite the reduction in the arrival penalties. For the other OD pairs, especially OD
pairs (q, r) and (q, y), it is interesting to notice that these passengers choose to depart
36

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
7:00 7:05 7:10 7:15 7:20 7:25 7:30
0
5
10
15
20
25
30
35
40
45
50
Number
of
passengers
OD(q,y)
OD(q,r)
OD(o,y)
OD(o,r)
Figure 8: Departure time of passengers (η3 = 1)
7:20 7:25 7:30 7:35 7:40 7:45
0
5
10
15
20
25
Number
of
passengers
OD(q,y)
OD(q,r)
OD(o,y)
OD(o,r)
Figure 9: Arrival time of passengers (η3 = 1)
earlier when the value of early/late arrival penalty is reduced by half. By investigat-
ing the arrival patterns, such departure time choice can be explained by the passengers’
trade-off between the penalties and congestion effect. In the base case, more than 50
passengers arrive within 7:33-7:37, while in the case with reduced penalties, only around
30 passengers arrive within that time interval, implying that passengers incur a higher
congestion cost in the base case. Therefore, when the values of early/arrival penalty is
reduced, passengers select the departure time that allow them to board a less congested
line. The trade-off between the congestion cost and arrival penalties can also be used to
explain why the passengers of OD pair (q, r) select the early run instead of postponing
their departure. The reason is that in the early run, they are the only passengers on the
transit line who arrive within the time window 7:20-7:26. If they postpone their departure
37

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
7:00 7:05 7:10 7:15 7:20 7:25 7:30
0
5
10
15
20
25
30
35
40
45
50
Number
of
passengers
OD(q,y)
OD(q,r)
OD(o,y)
OD(o,r)
Figure 10: Departure time of passengers: η3 = 0.5 and reduced early/late arrival penalties
7:20 7:25 7:30 7:35 7:40 7:45
0
5
10
15
20
25
Number
of
passengers
OD(q,y)
OD(q,r)
OD(o,y)
OD(o,r)
Figure 11: Arrival time of passengers: η3 = 0.5 and reduced early/late arrival penalties
times, they must bear a higher congestion cost due to the boarding of the passengers of
other OD pairs.
To illustrate the effect of capacity on departure and arrival times, the capacities of all
the lines are doubled. The departure and arrival patterns are plotted in Figures 12 and
13. For OD pair (q, r), all the passengers depart at 7:10 when the capacity is doubled. By
doing so, the total early arrival penalty of these passengers is reduced, because all these
passengers arrive within the desired time interval as shown in Figure 13 . For OD pair
(o, r), the passengers that depart at 7:15 switch from the second run to the first run of
line 3, because such choice reduces their congestion cost. More importantly, it is worth
mentioning the reason that these passengers do not depart at 7:05 before the capacity
improvement. This is because some of the passengers of OD pair (q, r) that take line 1
38

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
7:00 7:05 7:10 7:15 7:20 7:25 7:30
0
5
10
15
20
25
30
35
40
45
50
Number
of
passengers
OD(q,y)
OD(q,r)
OD(o,y)
OD(o,r)
Figure 12: Departure time of passengers: η3 = 0.5 and double capacity
7:25 7:30 7:35 7:40 7:45
0
5
10
15
20
25
Number
of
passengers
OD(q,y)
OD(q,r)
OD(o,y)
OD(o,r)
Figure 13: Arrival time of passengers: η3 = 0.5 and double capacity
arrive at node b at an earlier time interval than the passengers between OD pair (o, r) and
wait for line 3. Hence, these passengers have the priority to board line 3. Consequently,
there is no residual capacity for the passengers of OD pair (o, r).
For the other two OD pairs, the increment in the capacity does not affect their depar-
ture choices. On one hand, this is because the demand (i.e., OD pair (o, y)) is low and can
be accommodated before the capacity improvement. On the other hand, the stop that
they board is the first stop of the transit line; thus boarding priority can be guaranteed.
Moreover, it is observed that when the capacities are doubled, the first run of line 1 is
not used. This implies that the operator can cancel certain runs of transit services by
increasing vehicle capacity, resulting in a lower operational cost.
39

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
6.5 Effects of the value of travel time variance on equilibrium
arc flows
To further illustrate the properties of the proposed model, three strategies adopted by
the passengers of OD pair (q-r) are plotted in Figures 14, 15, and 16. Only the arcs with
positive passenger flows are displayed and the number inside square blankets beside a
selected arc denotes the passenger flow of OD pair (q-r) using a certain strategy. Due to
the space limitation, node r (which can be easily reached from node d) is not shown in
those figures.
0
Line 1, capacity = 20
Walking, capacity = ∞
[35]
[5]
[5]
[10]
[3.75]
[3.75]
[7.5]
[1.0]
[0.875]
[0.459]
Waiting arc, capacity = ∞
Figure 14: Equilibrium strategic flow of strategy s1
(0.0)
Figure 14 displays strategy s1
(0.0) when η3 = 0.0. All travellers depart from q at
time interval 0 (i.e., 7:00 am) and arrive at stop a at time interval 5 via the walking arc.
Afterwards, 20 passengers board the first run of line 1 and 15 passengers board line 2,
respectively. It is worth to mention that, at node a15, all 35 passengers have an identical
40

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
optimal preference set, which is [bl1 , cl2 , a6]. However, due to the capacity constraint of
line 1, only 20 passengers can board line 1 (i.e., their first choice) while 15 passengers use
their second choice. In addition, it is found that the uncertainty of travel time affects the
passengers’ travel choice. Take node b as an example, where the arrival time depends on
the in-vehicle travel time distribution of arc (a-b) of line 1. From Table 1, the probability
of arriving at node b10 is 0.5. In such case, passengers alight from line 1 and transfer to
the first run of line 3. In contrast, if passengers arrive at b9 or b11 with a probability of
0.25, they select to use line 1 continuously. This is because passengers arrive at b11 after
the departure time of line 3, while the expected travel time from b9 using line 3 is longer
after considering the additional waiting time.
[13.45]
0
[3.375]
[3.375] [0.59]
[0.59]
[0.0723]
[0.031]
(1.0)
Figures 15 and 16 display strategies s8
(1.0) and s9
(1.0) utilized by OD pair (q-r)
when η3 = 1.0. When the passengers consider the effect of variance, only 14.8 passengers
depart at node q0, and others postpone their departure time. Those 14.8 passengers are
41

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
divided into two groups. 13.45 passengers use strategy s8
(1.0), while 1.35 passengers
use strategy s9
(1.0). At node a5, the passengers utilizing strategy s8
(1.0) select line 1
as their first choice, and the passengers utilizing strategy s9
(1.0) select line 2 as their
first choice. Comparing with the case of η3 = 0.0, where all passengers have the same
preference set [bl1 , cl2 , a6], the passengers have two different preference sets in the case of
η3 = 1.0, namely strategy s8
(1.0) ([bl1 , cl2 , a6]) and strategy s9
(1.0) ([cl2 , bl1 , a6]). These
two strategies are different in terms of the order of arriving nodes (and the lines used) in
their user-preference sets. Strategy s8
(1.0) involves no transfer throughout the journey.
Strategy s9
(1.0) involves one transfer at node c and two transit lines. The implication is
that there is a tradeoff between the variance of in-vehicle travel time and the variance of
waiting time.
[1.35]
0
[0.3375]
[0.3375]
[0.675]
[0.0413]
[0.0177]
(1.0)
The first run of line 1 is fully occupied when the variance effect is ignored by the
passengers. (The number of passengers in the first run is equal to the vehicle capacity of
20.) However, when the passengers consider the effect of variance, the first run of line 1 is
not fully occupied. Only 13.45 passengers board line 1. This implies that the route choice
behavior affects the level of service inside transit vehicles. In addition, it is observed that
at node b10, the choice of using line 3 is removed from the optimal strategy set unlike to
the case when η3 = 0.0. Passengers prefer to continuously stay on line 1, implying that
passengers may avoid transfer when they are more risk aversive. This may occur because
the waiting time uncertainty can be high at the next stop or the in-vehicle travel time of
42

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Nb. of utilized strategies
q-y q-r o-y o-r
η3 = 0.5, φ = 0.3 2 2 1 2
η3 = 0.5, φ = 0.6 2 2 1 3
η3 = 1.0, φ = 0.3 2 3 1 3
η3 = 1.0, φ = 0.6 2 2 1 3
Table 12: Effects of φ on the number of utilized strategies
the next service is high, and hence it is better to stay in transit vehicles so as to arrive on
time. Hence, it can be concluded that passengers can adjust their choices of next transit
stations, in addition to just their adjusting departure times, to counteract the effect of
travel time variance. Consequently, the resultant flow pattern is significantly different.
6.6 Effects of coefficient φ on the number of utilized strategies
Table 12 illustrates the effects of coefficient φ on the number of utilized strategies for each
OD pair. In general, a larger value of φ indicates that the mean and variance of a link
cost are affected more by its previous arc. Consequently, by increasing the value of φ , the
strategy cost as well as the number of utilized strategies can be changed. For OD pairs q-y
and o-y, the numbers of utilized strategies are unaffected by the value of φ, because the
transit lines used in these strategies only traverse one arc. For the other two OD pairs, the
effects of φ also depend on the value of ηg
3. Surprisingly, all possible trends for the number
of utilized strategies are observed including increasing (i.e., OD pair o-r, when ηg
3 = 0.5
), remaining stable (i.e., OD pair o-r, when ηg
3 = 1.0 ) or decreasing (i.e., OD pair q-r,
when ηg
3 = 1.0). These observations imply that the value of φ can induce various effects
on the number of utilized strategies; thus it is important to have an accurate estimation
of the value of φ, which is left for future study.
7 Considerations in real life applications
To apply the proposed methodology to real-life applications, three issues are required
to consider: computational resource requirement, convergence conditions, and computa-
tional time. Because some variables in the proposed model are indexed by at least transit
station, time period, strategy, and transit line, the size of the matrices (computer stor-
age) grows exponentially when the TE network becomes larger or more transit lines are
43

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
modelled. In such cases, additional effort should be made on the effective allocation of
computer storage as most of the matrices are sparse matrices. A good data structure (for
example, using vectors and pointers) can be developed to reduce the computation storage.
Moreover, when solving for optimal solutions, it is not necessary to explicitly construct
and maintain the whole TE network in computer memory. Time-expanded nodes and
arcs can be generated as needed when solving the problem, e.g., to find a strategy with
the least effective cost.
The convergence condition of the solution algorithm is another important issue for
the applicability of our proposed methodology. The proposed MSA requires that the cost
function EC satisfies the symmetric continuous Jacobian condition or strictly monotone
condition for convergence. However, this condition may not be satisfied, especially for
realistic, large transit networks. Canteralla (1997) proposed the cost averaging algorithm.
Compared with the MSA, the cost averaging algorithm is a method of successive cost av-
erages instead of successive flow averages. To ensure convergence, this algorithm does not
rely on that the cost function EC satisfies the symmetric continuous Jacobian condition
or strictly monotone condition for convergence. Instead, the algorithm only requires some
milder assumptions for convergence (see Theorem 4 in his paper). Some assumptions are
used to ensure that a link flow solution exists to the problem and is unique (see Theorem
2 in his paper).
Computational time is also a crucial issue for the applicability of our methodology
in real transit network applications. Compared to the other transit assignment models,
the proposed model is more suitable to adopt parallel computing for reducing computa-
tional times. It is because the loading process and the computation of optimal strategy
are performed on a node basis. Thus, for each of these processes, they can be started
simultaneously from different nodes given that the specific criteria (reverse TC order
for stochastic loading and optimal strategy computation) are satisfied. Moreover, the
convergence speed of the MSA may be slow even for solving medium-size network transit
assignment problems, because of the step size used. The self-regulated averaging method
proposed by Liu et al. (2009) was shown to converge to the equilibrium solution faster
than the MSA. The self-regulated averaging method adjusts the step size to speed up the
convergence and has been applied to solve other traffic assignment problems (e.g., Szeto
et al., 2011a; Long et al., 2014). The convergence requirements are basically the same as
those for the MSA. This algorithm can be one of the candidate solution methods for real-
life applications. Furthermore, the cost averaging version of the self-regulated averaging
method proposed by Long et al. (2014) can be another choice. It has the advantages of
both the cost averaging method and the self-regulated averaging method.
44

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
8 Conclusion
In this paper, we propose a new schedule-based transit assignment model in which pas-
sengers adopt strategies to travel from their origins to their destinations. While this
strategy concept has been successfully used in previous transit assignment studies with
fixed timetables, the new proposed model captures explicitly the stochastic nature of the
transit schedules and in-vehicle travel times due to road conditions, incidents, or adverse
weather. No such analytical schedule-based model has been developed in the literature to
consider both travel strategies and supply uncertainties. When loading passengers on a
first-come-first-serve basis, the model takes into account the transit capacities explicitly.
Using a mean-variance approach, the equilibrium conditions for this schedule-based tran-
sit assignment problem are stated as a variational inequality involving a vector-valued
function of effective strategy costs. To find an equilibrium solution, we adopt the method
of successive averages in which the optimal strategy of each iteration is generated by
solving a dynamic program.
Numerical studies are included to illustrate the effect of supply uncertainties, vehicle
capacity and early/late arrival penalty parameters on travel strategies and/or departure
times of passengers. In particular, we show that
i) When the value of travel time variance increases, people may decide to leave later.
ii) Increasing/reducing vehicle capacity may have no effect on departure time choice.
iii) Early/late arrival penalties may have no effect on departure time choice.
iv) Passengers may make a tradeoff between the variance of in-vehicle travel time and
the variance of waiting time.
v) Passengers can adjust their choices of next transit stations, in addition to just
adjusting their departure times, to counteract the effect of travel time variance.
vi) For the same OD, the algorithm can generate different strategies that have the same
preference set at some TE nodes but have different preference sets for at least one
TE node.
vii) The number of utilized strategies for an OD pair does not necessary increase with
the value of travel time variance.
This study opens up many future research directions. One direction is to extend our
model to consider stochastic user equilibrium (SUE) and different nonlinear and non-
additive fare structures. Given that a fixed point formulation can easily cope with these
45

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
types of fare structures and SUE (see Cantarella, 1997) simultaneously, a resultant fixed
point formulation may be developed in the future. Moreover, the resultant formulation can
be solved by the cost averaging algorithm proposed by Cantarella (1997) or its extension
such as the cost averaging version of self-regulated averaging method proposed by Long
et al. (2014). Other future research of this study include the consideration of demand
uncertainties (Ng et al., 2011), the extra considerations of other dynamics such as the
year-to-year dynamic (e.g., Szeto and Lo, 2008; Lo and Szeto, 2009) and the day-to-
day dynamic (Watling and Cantarella, 2013), and the development of efficient solution
algorithms (e.g., Long et al., 2010; Szeto and Wu, 2011) for the large scale implementation
of the proposed model for transit assignment and vehicle scheduling.
Acknowledgements
The work described in this paper was partially supported by Grants from the United
Arab Emirates University - National Research Foundation (UAEU-NRF-58), the Central
Policy Unit of the Government of the Hong Kong Special Administrative Region and
the Research Grants Council of the Hong Kong Special Administrative Region, China
(HKU7026-PPR-12), the Hong Kong University Research Committee (201011159026), the
National Natural Science Foundation of China (71271183), and a Research Postgraduate
Studentship from the University of Hong Kong.
46

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Appendix A
i) We first show that σ2
jk = φσ2
ij + (1 − φ)V ar(Yjk) + φ(1 − φ)
(
µij − E(Yjk)
)2
σ2
jk =
∑
t
t2
P(Tjk = t) −
( ∑
t
tP(Tjk = t)
)2
=
∑
t
t2
(
φP(Tjk = t) + (1 − φ)P(Yjk = t)
)
−
( ∑
t
t
(
φP(Tjk = t) + (1 − φ)P(Yjk = t)
))2
= φ
∑
t
t2
P(Tij = t) + (1 − φ)
∑
t
t2
P(Yjk = t) − φ2
( ∑
t
tP(Tij = t)
)2
−(1 − φ)2
( ∑
t
tP(Yjk = t)
)2
− 2φ(1 − φ)
∑
t
tP(Tij = t)
∑
t
tP(Yjk = t)
= φE(T2
ij) − φ2
(E(Tij))2
+ (1 − φ)E(Y 2
jk) − (1 − φ)2
(E(Tjk))2
−2φ(1 − φ)E(Tij)E(Yjk)
= φ
(
E(T2
ij) − (E(Tij))2
)
+ (1 − φ)
(
E(Y 2
jk) − (E(Yjk))2
)
+φ(1 − φ)(E(Tij))2
+ φ(1 − φ)(E(Yjk))2
− 2φ(1 − φ)E(Tij)E(Yjk)
= φσ2
ij + (1 − φ)V ar(Yjk) + φ(1 − φ)
(
(E(Tij))2
+ (E(Yjk))2
− 2E(Tij)E(Yjk)
)
= φσ2
ij + (1 − φ)V ar(Yjk) + φ(1 − φ)
(
E(Tij) − E(Yjk)
)2
= φσ2
ij + (1 − φ)V ar(Yjk) + φ(1 − φ)
(
µij − E(Yjk)
)2
.
ii) For each 1 ≤ n ≤ Nl − 2, we will show by induction on n′
, 1 ≤ n′
≤ Nl − 1 − n that:
Cov(Tin(l)in+1(l), Tin+n′ (l)in+n′+1(l)) = φn′
σ2
in(l)in+1(l).
– n′
= 1
Cov(Tin(l)in+1(l), Tin+1(l)in+2(l))
= Cov(Tin(l)in+1(l), φTin(l)in+1(l) + (1 − φ)Yin+1(l)in+2(l))
= Cov(Tin(l)in+1(l), φTin(l)in+1(l)) (since Tjk and Yjk are independent)
= E(φT2
in(l)in+1(l)) − E(Tin(l)in+1(l))E(φTin(l)in+1(l))
= φ
(
E(T2
in(l)in+1(l)) − (E(Tin(l)in+1(l)))2
)
= φ σ2
in(l)in+1(l).
– Assume Cov(Tin(l)in+1(l), Tin+n′ (l)in+n′+1(l)) = φn′
σ2
in(l)in+1(l)
47

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
– n′
+ 1
Cov(Tin(l)in+1(l), Tin+n′+1(l)in+n′+2(l))
= Cov(Tin(l)in+1(l), φTin+n′ (l)in+n′+1(l) + (1 − φ)Yin+n′+1(l)in+n′+2(l))
= Cov(Tin(l)in+1(l), φTin+n′ (l)in+n′+1(l))
= φ Cov(Tin(l)in+1(l), Tin+n′ (l)in+n′+1(l))
= φ φn′
σ2
in(l)in+1(l)
= φn′+1
σ2
in(l)in+1(l).
48

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Appendix B
Strategy: s1
(0.0)
q0 [a5] a5 [bl1
, cl2
, a6] b9 [cl1
, b10] c12 [dl1
, c13] d15 [r20]
b10 [dl3
, cl1
, b11] c13 [dl1
, c14] d16 [r21]
b11 [cl1
, b12] c14 [dl1
, c15] d17 [r22]
c15 [dl1
, c16] d18 [r23]
c16 [dl1
, c17] d19 [r24]
c17 [dl1
, c18] d20 [r25]
c18 [dl1
, c19] d21 [r26]
d22 [r27]
d23 [r28]
d24 [r29]
d25 [r30]
Strategy: s2
(0.0)
q20 [a25] a25 [bl1
, a26] b29 [cl1
, b30] c32 [y37]
b30 [cl1
, b31] c33 [y38]
b31 [cl1
, b32] c34 [y39]
c35 [y40]
c36 [y41]
c37 [y42]
c38 [y43]
Strategy: s3
(0.0)
o25 [b30] b30 [cl1
, b31] c33 [y38]
c34 [y39]
c35 [y40]
c36 [y41]
c37 [y42]
Strategy: s4
(0.0)
o25 [b30] b30 [dl3
, cl1
, b31] c33 [dl1
, c34] d36 [r41]
c34 [dl1
, c35] d37 [r42]
c35 [dl1
, c36] d38 [r43]
c36 [dl1
, c37] d39 [r44]
c37 [dl1
, c38] d40 [r45]
d41 [r46]
d42 [r47]
d43 [r48]
d44 [r49]
Strategy: s5
(0.0)
q10 [a15] a15 [bl1
, a16] b19 [cl1
, b20] c22 [y27]
a16 [a17] b20 [cl1
, b21] c23 [y28]
a17 [a18] b21 [cl1
, b22] c24 [y29]
a18 [a19] c25 [y30]
a19 [a20] c26 [y31]
a20 [cl2
, a21] c27 [y32]
c28 [y33]
c29 [y34]
c30 [y35]
c31 [y36]
Strategy: s6
(0.0)
o15 [b20] b20 [dl3
, cl1
, b21] d28 [r33]
d29 [r34]
d30 [r35]
d31 [r36]
Table 13: Utilized strategies (η3 = 0.0)
49

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Strategy: s1
(0.5)
q0 [a5] a5 [bl1
, cl2
, a6] b9 [cl1
, b10] c12 [dl1
, c13] d15 [r20]
b10 [dl3
, cl1
, b11] c13 [dl1
, c14] d16 [r21]
b11 [cl1
, b12] c14 [dl1
, c15] d17 [r22]
c15 [dl1
, c16] d18 [r23]
c16 [dl1
, c17] d19 [r24]
c17 [dl1
, c18] d20 [r25]
c18 [dl1
, c19] d21 [r26]
d22 [r27]
d23 [r28]
d24 [r29]
d25 [r30]
Strategy: s2
(0.5)
q20 [a25] a25 [bl1
, a26] b29 [cl1
, b30] c32 [y37]
b30 [cl1
, b31] c33 [y38]
b31 [cl1
, b32] c34 [y39]
c35 [y40]
c36 [y41]
c37 [y42]
c38 [y43]
Strategy: s3
(0.5)
q10 [a15] a15 [bl1
, a16] b19 [cl1
, b20] c22 [dl1
, c23] d25 [r30]
a16 [a17] b20 [dl3
, cl1
, b21] c23 [dl1
, c24] d26 [r31]
a17 [a18] b21 [cl1
, b22] c24 [dl1
, c25] d27 [r32]
a18 [a19] c25 [dl1
, c26] d28 [r33]
a19 [a20] c26 [dl1
, c27] d29 [r34]
a20 [cl2
, a21] c27 [dl1
, c28] d30 [r35]
c28 [dl1
, c29] d31 [r36]
c29 [c30] d32 [r37]
c30 [c31] d33 [r38]
c31 [c32] d34 [r39]
c32 [dl1
, c33] d35 [r40]
d36 [r41]
d37 [r42]
d38 [r43]
d39 [r44]
Strategy: s4
(0.5)
o25 [b30] b30 [cl1
, b31] c33 [y38]
b31 [cl1
, b32] c34 [y39]
c35 [y40]
c36 [y41]
c37 [y42]
c38 [y43]
Strategy: s5
(0.5)
o25 [b30] b30 [dl3
, cl1
, b31] c33 [dl1
, c34] d36 [r41]
b31 [cl1
, b32] c34 [dl1
, c35] d37 [r42]
c35 [dl1
, c36] d38 [r43]
c36 [dl1
, c37] d39 [r44]
c37 [dl1
, c38] d40 [r45]
c38 [dl1
, c39] d41 [r46]
d42 [r47]
d43 [r48]
d44 [r49]
d45 [r50]
Strategy: s6
(0.5)
q15 [a20] a20 [cl2
, a21] c29 [y34]
c30 [y35]
c31 [y36]
Strategy: s7
(0.5)
o15 [b20] b20 [dl3
, cl1
, b21] c23 [dl1
, c24] d26 [r31]
c24 [dl1
, c25] d27 [r32]
c25 [dl1
, c26] d28 [r33]
c26 [dl1
, c27] d29 [r34]
c27 [dl1
, c28] d30 [r35]
d31 [r36]
d32 [r37]
d33 [r38]
d34 [r39]
Table 14: Utilized strategies (η3 = 0.5)
50

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Strategy: s1
(1.0)
o5 [b10] b10 [dl3
, cl1
, b11] c13 [dl1
, c14] d16 [r21]
c14 [dl1
, c15] d17 [r22]
c15 [dl1
, c16] d18 [r23]
c16 [dl1
, c17] d19 [r24]
c17 [dl1
, c18] d20 [r25]
d21 [r26]
d22 [r27]
d23 [r28]
d24 [r29]
Strategy: s2
(1.0)
q15 [a20] a20 [cl2
, a21] b29 [cl1
, b30] c33 [y38]
a21 [a22] b30 [cl1
, b31] c34 [y39]
a22 [a23] b31 [cl1
, b32] c35 [y40]
a23 [a24] c36 [y41]
a24 [a25] c37 [y42]
a25 [bl1
, a26] c38 [y43]
Strategy: s3
(1.0)
q10 [a15] a15 [bl1
, a16] b19 [cl1
, b20] c22 [dl1
, c23] d25 [r30]
a16 [a17] b20 [dl3
, cl1
, b21] c23 [dl1
, c24] d26 [r31]
a17 [a18] b21 [cl1
, b22] c24 [dl1
, c25] d27 [r32]
a18 [a19] c25 [dl1
, c26] d28 [r33]
a19 [a20] c26 [dl1
, c27] d29 [r34]
a20 [cl2
, a21] c27 [dl1
, c28] d30 [r35]
c28 [dl1
, c29] d31 [r36]
c29 [c30] d32 [r37]
c30 [c31] d33 [r38]
c31 [c32] d34 [r39]
c32 [dl1
, c33] d35 [r40]
d36 [r41]
d37 [r42]
d38 [r43]
d39 [r44]
Strategy: s4
(1.0)
o25 [b30] b30 [cl1
, b31] c33 [y38]
c34 [y39]
c35 [y40]
c36 [y41]
c37 [y42]
Strategy: s5
(1.0)
o25 [b30] b30 [dl3
, cl1
, b31] c33 [dl1
, c34] d36 [r41]
c34 [dl1
, c35] d37 [r42]
c35 [dl1
, c36] d38 [r43]
c36 [dl1
, c37] d39 [r44]
c37 [dl1
, c38] d40 [r45]
d41 [r46]
d42 [r47]
d43 [r48]
d44 [r49]
Table 15: Utilized strategies (η3 = 1.0, first five strategies)
51

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Strategy: s6
(1.0)
o15 [b20] b20 [dl3
, cl1
, b21] c23 [dl1
, c24] d26 [r31]
c24 [dl1
, c25] d27 [r32]
c25 [dl1
, c26] d28 [r33]
c26 [dl1
, c27] d29 [r34]
c27 [dl1
, c28] d30 [r35]
d31 [r36]
d32 [r37]
d33 [r38]
d34 [r39]
Strategy: s7
(1.0)
q20 [a25] a25 [bl1
, a26] b29 [cl1
, b30] c32 [y37]
b30 [cl1
, b31] c33 [y38]
b31 [cl1
, b32] c34 [y39]
c35 [y40]
c36 [y41]
c37 [y42]
c38 [y43]
Strategy: s8
(1.0)
q0 [a5] a5 [bl1
, cl2
, a6] b9 [cl1
, b10] c12 [dl1
, c13] d15 [r20]
b10 [dl3
, cl1
, b11] c13 [dl1
, c14] d16 [r21]
b11 [cl1
, b12] c14 [dl1
, c15] d17 [r22]
c15 [dl1
, c16] d18 [r23]
c16 [dl1
, c17] d19 [r24]
c17 [dl1
, c18] d20 [r25]
c18 [dl1
, c19] d21 [r26]
d22 [r27]
d23 [r28]
d24 [r29]
d25 [r30]
Strategy: s9
(1.0)
q0 [a5] a5 [cl2
, bl1
, a6] c14 [dl1
, c15] d17 [r22]
c15 [dl1
, c16] d18 [r23]
c16 [dl1
, c17] d19 [r24]
d20 [r24]
d21 [r26]
d22 [r27]
d23 [r28]
Table 16: Utilized strategies (η3 = 1.0, last four strategies)
52

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
References
Bartholdi III, J.J., Eisenstein, D.D., 2012. A self-coördinating bus route to resist bus
bunching. Transportation Research Part B 46 (4), 481-491.
Benezech, V., Coulombel, N., 2013. The value of service reliability. Transportation
Research Part B 58, 1-15.
Biswas, A., Song, P., 2009. Discrete-valued ARMA processes. Statistical and Probability
Letters (79), 1884-1889.
Börjesson, M., Eliasson, J., Franklin, J.P., 2012. Valuations of travel time variability in
scheduling versus mean-variance models. Transportation Research Part B 46 (7),
855-873.
Brockwell, P.J., Davis, R.A., 1991. Time Series: Theory and Methods. Springer-Verlag,
New York.
Cantarella, G.E., 1997. A general fixed-point approach to multimode multi-user equilib-
rium assignment with elastic demand. Transportation Science 31 (2), 107-128.
Cascetta, E., Nuzzolo, A., Russo, F., Vitetta, A., 1996. A modified logit route choice
model overcoming path overlapping problems: specification and some calibration
results for interurban networks, J. B. Lesort (ed), Transportation and Traffic Theory,
Pergamon, New York, 697-711.
Cepeda, M., Cominetti, R., Florian, M., 2006. A frequency-based assignment model for
congested transit networks with strict capacity constraints: characterization and
computation of equilibrium. Transportation Research Part B 40 (6), 437-459.
Cortés, C.E., Jara-Moroni, P., Moreno, E., Pineda, C., 2013. Stochastic transit equilib-
rium. Transportation Research Part B 51, 29-44.
De Cea, J., Fernandez, E., 1993. Transit assignment for congested public transport
system: An equilibrium model. Transportation Science 27 (2), 133-147.
Hamdouch, Y., Marcotte, P., Nguyen, S., 2004. A strategic model for dynamic traffic
assignment. Networks and Spatial Economics 4 (3), 291-315.
Hamdouch, Y., Lawphongpanich, S., 2008. Schedule-based transit assignment model
with travel strategies and capacity constraints. Transportation Research Part B 42
(7-8), 663-684.
53

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Hamdouch, Y., Ho, H.W., Sumalee, A., Wang, G., 2011. Schedule-based transit assign-
ment model with vehicle capacity and seat availability. Transportation Research
Part B 45 (10), 1805-1830.
Jenelius, E., 2012. The value of travel time variability with trip chains, flexible scheduling
and correlated travel times. Transportation Research Part B 46 (6), 762-780.
Kurauchi, F., Bell, M.G.H., Schmöcker, J.-D., 2003. Capacity constrained transit as-
signment with common lines. Journal of Mathematical Modelling and Algorithms
2 (4), 309-327.
Lam, W.H.K., Gao, Z.Y., Chan, K.S., Yang, H., 1999. A stochastic user equilibrium
assignment model for congested transit networks. Transportation Research Part B
33 (5), 351-368.
Lam, W.H.K., Zhou, J., Sheng, Z.H., 2002. A capacity restraint transit assignment with
elastic line frequency. Transportation Research Part B 36 (10), 919-938.
Li, Z.C., Lam, W.H.K., Sumalee, A., 2008. Modeling impacts of transit operator fleet
size under various market regimes with uncertainty in network. Transportation
Research Record 2063, 18-27.
Li, Z.C., Lam, W.H.K., Wong, S.C., 2009. The optimal transit fare structure under
different market regimes with uncertainty in the network. Networks and Spatial
Economics 9 (2), 191-216.
Liu, H., He, X., He, B., 2009. Method of successive weighted averages (MSWA) and
self-regulated averaging schemes for solving stochastic user equilibrium problem.
Networks and Spatial Economics 9 (4), 485-503.
Lo, H.K., Yip, C.W., Wan, K.H., 2003. Modeling transfer and non-linear fare structure
in multi-modal network. Transportation Research Part B 37 (2), 149-170.
Lo, H.K., Szeto, W.Y., 2009. Time-dependent transport network design under cost-
recovery. Transportation Research Part B, 43 (1), 142-158.
Long, J.C., Gao, Z.Y., Zhang, H.Z., Szeto, W.Y., 2010. A turning restriction design
problem in urban road networks. European Journal of Operational Research 206
(3), 569-578.
54

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Long, J., Szeto, W.Y., Huang, H.J., 2014. A bi-objective turning restriction design
problem in urban road networks. European Journal of Operational Research 237
(2), 426-439
Marcotte, P., Nquven, S., Schoeb, A., 2004. A strategic flow model of traffic assignment
in static capacitated networks. Operations Research 52 (2), 191-212.
Moller-Pedersen, J., 1999. Assignment model of timetable based systems (TPSCHED-
ULE). Proceedings of 27th European Transportation Forum, Seminar F, Cambridge,
England, 159-168.
Nagurney, A. 1993. Network Economics: A Variational Inequality Approach. Kluwer
Academic Publishers. Norwell, Massachusetts, USA.
Nielsen, O.A., 2000. A stochastic transit assignment model considering differences in
passengers utility functions. Transportation Research Part B 34 (5), 377-402.
Nielsen, O.A., Frederiksen, R.D., 2006. Optimisation of timetable-based, stochastic
transit assignment models based on MSA. Annals of Operations Research 144 (1),
263-285.
Ng, M.W., Szeto, W.Y., Waller, S.T., 2011. Distribution-free travel time reliability
assessment with probability inequalities. Transportation Research Part B 45 (6),
852-866.
Nguyen, S., Pallottino, S., Malucelli, F., 2001. A modeling framework for passenger
assignment on a transport network with timetables. Transportation Science 35 (3),
238-249.
Nuzzolo, A., Russo, F., Crisalli, U., 2001. A doubly dynamic schedule-based assignment
model for transit networks. Transportation Science 35 (3), 268-285.
Nuzzolo, A., Crisalli, U., Rosati, L., 2012. A schedule-based assignment model with
explicit capacity constraints for congested transit networks. Transportation Resarch
Part C 20 (1), 16-33.
Poon, M.H., Wong, S.C., Tong, C.O., 2004. A dynamic schedule-based model for con-
gested transit networks. Transportation Research Part B 38 (4), 343-368.
Schmöcker, J.D., Bell, M.G.H., Kurauchi, F., Shimamoto, H., 2009. Frequency-based
assignment with consideration of seat availability. 11th International Conference on
Advanced Systems for Public Transport, Hong Kong.
55

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Schmöcker, J.D., Fonzone, A., Shimamoto, H., Kurauchi, F., Bell, M.G.H., 2011. Frequency-
based transit assignment considering seat capacities. Transportation Research Part
B 45 (2), 392-408.
Shao, H., Lam, W.H.K., Tam, M.L., 2006. A reliability-based stochastic traffic assign-
ment model for network with multiple user classes under uncertainty in demand.
Networks and Spatial Economics 6 (3-4), 173-204.
Spiess, H., Florian, M., 1989. Optimal strategies: A new assignment model for transit
networks. Transportation Research Part B 23 (2), 83-102.
Sumalee, A., Tan, Z.J., Lam, W.H.K., 2009. Dynamic stochastic transit assignment with
explicit seat allocation model. Transportation Research Part B 43 (8-9), 895-912.
Sumalee A., Uchida, K., Lam, W.H.K., 2011. Stochastic multi-modal transport net-
work under demand uncertainties and adverse weather condition. Transportation
Research Part C 19 (2), 338-350.
Szeto, W.Y., Lo, H.K., 2008. Time-dependent transport network improvement and
tolling strategies. Transportation Research Part A 42 (2), 376-391.
Szeto, W.Y., Wu, Y.Z., 2011. A simultaneous bus route design and frequency setting
problem for Tin Shui Wai, Hong Kong. European Journal of Operational Research
209 (2), 141-155.
Szeto, W.Y., Jiang, Y., Sumalee, A., 2011a. A cell-based model for multi-class doubly
stochastic dynamic traffic assignment. Computer-Aided Civil and Infrastructure
Engineering, 26 (8), 595-611.
Szeto, W.Y., Solayappan, M., Jiang, Y., 2011b. Reliability-based transit assignment
for congested stochastic transit networks. Computer-Aided Civil and Infrastructure
Engineering 26 (4), 311-326.
Szeto, W.Y., Jiang, Y., Wong, K.I., Solayappan, M., 2013. Reliability-based stochastic
transit assignment with capacity constraints: Formulation and solution method.
Transportation Research Part C 35, 286-304.
Szeto, W.Y., Jiang, Y., 2014. Transit assignment: approach-based formulation, extra-
gradient method and paradox. Transportation Research Part B 62, 51-76.
56

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
Tong, C.O., Wong, S.C., 1998. A stochastic transit assignment model using a dynamic
schedule-based network. Transportation Research Part B 33 (2), 107-121.
Trozzi, V., Gentile, G., Bell, M.G.H., Kaparias, I., 2013. Dynamic user equilibrium in
public transport networks with passenger congestion and hyperpaths. Transporta-
tion Research Part B 57, 266-285.
Watling, D.P., Cantarella, G.E., 2013. Modelling sources of variation in transportation
systems: theoretical foundations of day-to-day dynamic models. Transportmetrica
B: Transport Dynamics 1 (1), 3-32.
Wilson, N.H.M., Nuzzolo, A., 2004. Scheduled-based dynamic transit modeling: Theory
and applications. Springer, New York.
Wu, J.H., Florian, M., Marcotte, P., 1994. Transit equilibrium assignment: A model
and solution algorithms. Transportation Science 28 (3), 193-203.
Yang, L., Lam, W.H.K., 2006. Probit-type reliability-based transit network assignment.
Transportation Research Record 1977, 154-163.
Zhang, Y., Lam, W.H.K., Sumalee, A., Lo, H.K., Tong, C.O., 2010. The multi-class
schedule-based transit assignment model under network uncertainties. Public Trans-
port 2 (1), 69-86.
57

A New Schedule-Based Transit Assignment Model With Travel Strategies And Supply Uncertainties

More Related Content

Similar to A New Schedule-Based Transit Assignment Model With Travel Strategies And Supply Uncertainties (20)

More from Susan Campos (20)

Recently uploaded (20)

A New Schedule-Based Transit Assignment Model With Travel Strategies And Supply Uncertainties