Transportation Research Part B 117 (2018) 37–62

Contents lists available at ScienceDirect

Transportation Research Part B

journal homepage: www.elsevier.com/locate/trb

Remote park-and-ride network equilibrium model and its

applications
Zhiyuan Liu a,∗, Xinyuan Chen b, Qiang Meng c, Inhi Kim b
a
Jiangsu Key Laboratory of Urban ITS, Jiangsu Province Collaborative Innovation Center of Modern Urban Traffic Technologies, School of
Transportation, Southeast University, Nanjing, 210096, China
b
Institute of Transport Studies, Department of Civil Engineering, Monash University, Clayton, Victoria 3800, Australia
c
Department of Civil and Environmental Engineering, National University of Singapore, Singapore 117576, Singapore

a r t i c l e i n f o a b s t r a c t

Article history: Existing park-and-ride (P&R) sites are usually located near a train/bus station where con-
Received 18 January 2018 struction and operation costs are considerably high. Thus, this paper proposes a new P&R
Revised 18 June 2018
service mode, “Remote P&R (RPR)”, where the car park locates in a suburban area with
Accepted 6 August 2018
lower land value. Dedicated express bus service is used to connect this site and a nearby
Available online 6 September 2018
train station. To quantitatively evaluate the impacts of RPR on the network flows, a com-
Keywords: bined modal split and traffic assignment model (CMSTA) is developed, where a cross-
Remote park-and-ride nested logit (CNL) model is adopted to cope with the mode similarity. The problem is
Cross-nested logit formulated as a convex programming model and solved by the Evans algorithm, and then
Combined modal split and traffic extended to asymmetric path-based cases, where a variational inequality (VI) model is
assignment built and solved by a self-adaptive gradient projection (SAGP) algorithm. Taking the CM-
Multimodal network design STA as the lower level and multimodal stochastic system optimum (MSSO) as the objec-
tive, we further develop a mathematical programming model with equilibrium constraints
(MPEC) for the optimal network design of RPR. Based on an origin-based reformulation
of the MPEC model, an exact solution method based on the nonlinear valid inequalities
(NVI) is applied. Numerical examples demonstrate that the RPR services can significantly
influence network users’ travel decisions, promote the usage of public transportation and
mitigate traffic congestion in the congested area of metropolitan cities.
© 2018 Elsevier Ltd. All rights reserved.

1. Introduction

Prompting the usage of public transport (bus, train and tram) has been regarded as a universal solution to the severe
congestion in urban areas. The public transport systems in e.g., many Australian cities, however, are less attractive compared
with other dense cities in Europe and Asia. The low passenger demand in the vast suburban areas gives rise to: (a) low bus
frequency; (b) tortuous bus itinerary; (c) much longer trip time; (d) lack of door-to-door services. Thus, for the majority of
commuters living away from the train lines, it is very inconvenient to take public transport, in contrast with driving. P&R
therefore becomes an important scheme to prompt the public transport usage in Australian cities. It allows the commuters
to drive from home, park their car in a dedicated P&R site, and then take public transport to city area or other town centers.

∗
Corresponding author.
E-mail address: zhiyuanl@seu.edu.cn (Z. Liu).

https://doi.org/10.1016/j.trb.2018.08.004
0191-2615/© 2018 Elsevier Ltd. All rights reserved.
38 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62

P&R can be categorized into bus-based P&R and rail-based P&R, based on different “ride” components. Most of the bus-
based P&R is commonly adopted in small cities in US and UK with no developed rail system, where a dedicated express
bus line is used for the P&R (Duncan, 2010 and Meek et al., 2008). In a trainpolitan area with an advanced rail system,
rail-based P&R is more superior to bus-based P&R, due to the faster speed, higher level of safety and comfortability of rail
system. Also, compared with bus-based P&R, the rail-based system is independent of traffic flows, thus its travel time is
more accurate and stable. The existing P&R services in most Australian cities are also rail-based. Hence, this paper focuses
on rail-based P&R network design problem.
Despite the convenience to the commuters, providing more parking spaces beside the train stations is usually not consid-
ered as transit-oriented-development (TOD). In addition, train stations are usually located in town centers, where the land
value is relatively high, thus largely increasing the construction cost of P&R services. Thus, in this paper, we propose and
focus on a new concept of P&R: site the P&R in undeveloped areas with low land value, and use dedicated express bus ser-
vices to connect the P&R site to the train stations. This sort of P&R scheme is termed as “Remote P&R” (RPR) or “rail-based
P&R with feeder bus services”. Compared with the existing P&R systems, RPR provides a more general service framework
where the car parks do not have to be close to a train station. With sufficient parking spaces and rapid feeder bus services,
the travel impedance would be further reduced which results in an essentially different travelling behavior of the network
users. Meanwhile, the redistribution of network flows on the network allows the transportation authorities to re-balance the
flow distribution, promote the usage of public transport, alleviate traffic congestion in urban areas and reduce other adverse
external effects of pure auto mode.
Previous studies revealed that the P&R scheme should be planned systemically, because it may also encourage existing
public transport riders to drive, due to the convenience provided to driving (Hamer, 2010). Hence, arbitrarily constructed
RPR sites could attract more public transport riders to drive, rather than the desired outcome of prompting more drivers to
ride. It could also increase the total driving distance per vehicle, and thus deteriorate the traffic congestion/pollution. Thus, a
systematic study is necessary to properly determine the optimal location and capacity of RPR sites subject to a given budget
limit, which is taken as the main objective of this paper.

1.1. Literature review

Most of the previous studies on P&R focus on the policy side. Bus-based P&R has been commonly implemented in many
UK cities (Parkhurst, 1995, 20 0 0 and Parkhurst and Richardson, 20 02). According to the study of Meek et al. (2008) on
the bus-based P&R in the UK, 50–71% of weekday P&R users reported the car as their previous mode of travel, suggesting
that P&R does assist in reducing car use. Cairns (1998) and Dijk and Montalvo (2011) summarized the policy frames of
P&R in European cities. Lam et al. (2001) studied a trial P&R scheme in Hong Kong and gave a guideline for the design of
P&R systems in Eastern Asia. Duncan and Christensen (2013) pointed out that the station area characteristics are significant
predictors of the light-rail based P&R in the US. P&R facilities tend to be built in suburban areas with cheaper land and
lower population density. Hamer (2010) conducted a survey in Melbourne to assess the impacts of P&R on the modal split,
which indicated that the proportion of former drivers could be higher than that recorded. Wiseman et al. (2012) investigated
a newly established Adelaide Entertainment Centre P&R facility located in the inner suburb of Hindmarsh, and concluded
that location of P&R station is an important factor which can influence the performance of P&R system. They suggested the
needs for more studies in Australian cities to better understand the P&R systems.
Quantitative studies on the modeling of P&R are relatively sparse. An initial work was carried out by
Fernandez et al. (1994), where a pseudo link was adopted to represent the transfer behavior of passengers in the P&R
car parks. This work was later extended to the case of asymmetric link travel time functions by Garcıá and Marıń (2005),
and to the logit-based stochastic model by Li et al. (2007) and Lam et al. (2007). Kitthamkesorn et al. (2016) considered
mode similarity and route similarity of the green travel modes, and further extended to the case of webit-based model
(Kitthamkesorn and Chen, 2017). Some other studies focus on the P&R schemes in a linear corridor, e.g., Wang et al. (2004),
Wang et al. (2004b), Liu et al. (2009), Wang and Du (2013) and Du and Wang (2014).
Liu and Meng (2014) provided a formulation for the bus-based P&R in a network with the congestion charge.
Wang et al. (2014) studied the P&R schemes in dynamic transport networks. Pineda et al. (2016) proposed an integrated
traffic-transit stochastic equilibrium model with park-and-ride facilities. Song et al. (2017) provided an integrated planning
scheme of P&R facilities and transit service.
Network analysis of P&R schemes endogenously involves the modal split as well as traffic assignment. The abovemen-
tioned studies on network analysis, however, have not well captured a unique feature of P&R: it has heavy overlaps with
other transport systems; the first-half of a P&R trip overlaps with the road network and the second-half coincides with the
transit network. In view of these overlaps, the cross-nested logit (CNL) model is more suitable for the network analysis of
P&R schemes. The CNL model was initially proposed by Vovsha (1997) for transport network modeling. However, this sem-
inal study cannot be used directly for the P&R network flow analysis since it does not involve traffic assignment. We thus
aim to provide a holistic study on the flow equilibrium of a transport network with RPR schemes. Based on the CNL model,
a combined modal split and traffic assignment (CMSTA) modeling framework is built. The CMSTA problem is first formu-
lated as an optimization model for the simplified link-based case and then modeled as a VI problem for the asymmetric
path-based case. The CMSTA model is then taken as equilibrium constraints for the optimal planning (siting and capacity
design) of the P&R scheme.
 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62 39

Fig. 1. A linear corridor with RPR facility.

To sum up, this study addresses a new concept for the trainpolitan P&R scheme, i.e., Remote P&R (RPR) with dedicated
feeder bus services to the train stations. The main objective of this study is to provide new methodologies for the assessment
and optimal design of the RPR scheme in the multimodal transport network. In terms of formulation, the traditional rail-
based P&R can be taken as a special case of the RPR scheme. The tasks and achievements of this study consist of (i) develop
an in-depth and systematic method for the assessment of RPR schemes, where a CNL-based CMSTA model is proposed;
(ii) build a method for the optimal network design (siting and capacity design) of the RPR scheme. As an objective of the
network design, a multimodal stochastic system optimum (MSSO) model is proposed. Furthermore, based on an origin-based
reformulation of the MPEC model, an exact solution algorithm is applied to address the optimal solution of the RPR network
design.
The remainder of this paper is organized as follows: in the next section, we present a new CMSTA model for the mul-
timodal network with RPR, which is solved by Evans algorithm along with an innovative shortest path searching approach
for the RPR travel mode. In Section 3, we extend the CMSTA to a more general case with asymmetric and non-additive
link travel time functions, where a VI formulation is provided and solved by a SAGP algorithm. Section 4 first presents a
formulation of MSSO. Then, taking the MSSO as an objective, an MPEC model is proposed for the optimal planning of RPR
scheme, which is solved by an exact solution method based on the NVI in Section 5. Numerical studies are conducted in
Section 6 to validate the proposed methodology. Finally, Section 7 concludes this study.

2. Equilibrium flow and network assessment

Given a particular RPR scheme, to analyze its impacts on the system, we first build the models for the equilibrium
flows. To better present the modeling framework, a simple linear corridor is first adopted in Section 2.1 as shown in Fig. 1,
where only one highway, one rail line, and one RPR site exist. Then, the model is extended to the case of general transport
networks in Section 2.2. The notations used are listed as follows:
Sets
N Set of nodes in the network.
A Set of links in the network.
W Set of OD pairs.
E Set of nests which can be specified by OD pairs, e.g., Eod .
M Set of modes which can specified by nest and OD pairs, e.g., Meod .
K Set of paths which can be specified by mode, and OD pairs, e.g., Km od .

Parameters
δa,k
em,od
A binary coefficient which equals 1 if path k ∈ Kmod uses link a ∈ A, and 0, otherwise.

δa m A binary coefficient which equals 1 if mode m ∈ M uses link a ∈ A, and 0, otherwise.

q̄ Total travel demand which can be specified by OD pairs, e.g., q̄od .
ta ( va ) Travel time function on link a ∈ A, which is monotonically increasing and continuously differentiable.
ck Travel time on path k which can be specified by mode, nest, and OD pairs, e.g., ckem,od .
ϕm Travel impedance of mode m which can be specified by nest and OD pairs, e.g., ϕem
od .

θ Root dispersion parameter of CNL model which can be specified by OD pairs, e.g., θ od .
μe Dispersion parameter of nest e which can be specified by OD pairs, e.g., μod
e .
αem Degree of membership of mode m in nest e which can be specified by OD pairs, e.g., αem od .

Decision variables
fk Flow on path k which can be specified by mode, nest, and OD pairs, e.g., f kem,od .
va Flow on link a ∈ A.
qm Demand of travel mode m which can be specified by nest and OD pairs, e.g., qod em .
zai A binary decision variable which equals 1 if the capacity of RPR link equals i and 0, otherwise.
40 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62

Fig. 2. Tree structure of the CNL in the linear corridor.

2.1. The case of a linear corridor

As shown in Fig. 1, considering a linear corridor G(N, A), with only one rail line (links 3 and 4) and another highway
(links 1 and 2) from suburb o to the urban city area d. One RPR site is constructed at np , which is adjacent to the highway.
One dedicated express bus service (link 5) is provided to connect the RPR site np and railway station nr . RPR users first
drive from home, park their cars at np , take the express bus service, and then take train service at the station nr .
For this linear corridor and also the whole network in Section 2.2, the set of links A consists of three sub-sets: set of
auto links Aa ; set of rail links Ar ; and set of RPR links Ap ; A = Aa ∪Ar ∪Ap . A RPR link always connects an auto link and a
rail link, which represents the transfer behavior. The total trip impedance on a RPR link includes the time spent on parking,
waiting/boarding the dedicated bus as well as the bus in-vehicle travel time, which is clearly flow-dependent. The link travel
time ta , a ∈ A is a continuously differentiable and monotonically increasing function of the link flow va . The path travel time

ck is defined as ck = a ta δa,k , where δa,k = 1 if path k uses link a ∈ A, and 0 otherwise.
In this section, we further make the following two assumptions.

Assumption 1. The link travel time functions are separable, i.e., ta is only a function of its own flow va .

Assumption 2. The path travel time ck is additive to its links, i.e. ck = a ta δa,k , where δa,k = 1 if path k uses link a ∈ A, and 0
otherwise.

Clearly, for any RPR route, its first half overlaps with the highway and second half coincides with the rail line. Therefore,
the multinomial logit (MNL) or nest logit (NL) model are not able to depict such sort of cross-attribution. Thus, the CNL
structure in Fig. 2 is suitable to analyze the modal split in such a network.
Let e denote a nest of the CNL model, and m denote a travel mode, and all the nests and modes are grouped into set E
and M, respectively. All the travel modes attached to nest e are grouped into Me . We can see that in Fig. 2, E: ={e1 = road,
e2 = rail}, M: ={m1 = auto, m2 = train, m3 = RPR}, Me1 = {m1 , m3 } and Me2 = {m2 , m3 }. In the linear corridor, each mode m in
nest e (em) has only one path, thus the travel time of this path is taken as the travel impedance of mode em.
The probability of choosing any mode m equals (Kitthamkesorn et al., 2016 and Papola, 2004):
   μe −1
 1 / μe θ  1 / μe θ
(αem ) exp V (αei ) exp V
e∈E
μe em i∈Me
μe ei
pm =   μl (1)
  θ
(αls )1/μl exp V
l∈E s∈Ml
μl ls

where μe ∈ [0, 1] is the nest dispersion parameter and θ ∈ [0, +∞) is the root dispersion parameter. α em ∈ [0, 1] denotes
the degree of membership of mode m in nest e. Vem is the deterministic component of the utility of mode m in nest e,
which can be taken as the negative path travel time.

2.1.1. Convex programming model

Let qm denotes the demand on travel mode m. Since each mode only has one available path, qm is also its path flow, and
we have:

qm = q̄ pm (2)
where q̄ is the total travel demand. Let qem be the demand of mode m in nest e, and we have:

qm = qem (3)
e∈E
 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62 41

The flow on link a, denoted by va equals:


va = qm δam (4)
m∈M

where δam = 1 if mode m uses link a, and δam = 0, otherwise.

The following model is then built for the modal split problem of the linear corridor.
[MP–MS]
min Z (q ) = Z1 + Z2 + Z3
  va   μe qem  1 − μe  
= ta (x )dx + qem ln + qem ln qem (5)
a∈A
0 e∈E m∈Me
θ (αem ) e e∈E θ
1 / μ
m∈Me m∈Me

subject to

qm = q̄ (6)
m∈M

qem = qm , ∀m ∈ Me (7)
e∈E


va = qm δam , ∀a ∈ A (8)
m

qem ≥ 0, ∀e ∈ E, m ∈ M (9)
where q is the vector of modal demands, q: =(qem ,m ∈ Me ,e ∈ E). Z1 is the Beckmann’s transformation to reflect the con-
gestion effect; Z2 and Z3 are, respectively, the conditional and marginal entropy terms corresponding to the CNL two-level
probability structure (Kitthamkesorn et al., 2016); When μe = 1, the mode choice probability function collapses from CNL to
the MNL. As θ approaches infinity, mode choice probability becomes a deterministic pattern (user equilibrium). The follow-
ing two propositions are provided, rigorous proofs of which are in the Appendixs A.1 and A.2.

Proposition 1. The [MP–MS] gives the CNL-based modal split solution (1).

Proposition 2. The [MP–MS] is a convex program and its optimal solution is unique.

2.2. The case of whole transport network

We now proceed to the case of a real transport network, which is strongly connected and more representative. Between
any OD pair (o, d), there is more than one route, and the origin o could be in the catchment of more than one RPR site. If o
is near to any train station, the users also have direct access to the rail network. In order to assess a RPR plan in a transport
network, the network equilibrium flows should be evaluated. The travel plan of a network user consists of mode choice
and route choice. Thus, a CMSTA model should be proposed for the network equilibrium problem, which provides behavior
richness, quantitative evaluation of different RPR schemes as well as computational tractability (Kitthamkesorn et al., 2016).

2.2.1. Mathematical model

Given a strongly connected multimodal transport network G = (N, A), where the sets of nodes N and links A are similarly
defined as in Section 2.1. Let W denote the set of OD pairs, q̄od denote the total travel demand for (o, d) ∈ W, Eod denote the
set of nests between OD pair (o, d) ∈ W, and Meod denote the set of modes in nest e between OD pair (o, d). If the auto, rail
and RPR services are all available between OD pair (o, d), i.e., Eod : ={e1 = road, e2 = rail} and Mod : ={m1 = auto, m2 = train,
m3 = RPR}, we can get Meod1 := {m1 , m3 } and Meod2 := {m2 , m3 }. Let Km
od denote the set of paths under m between OD pair (o,

d); fkem,od denote the flow on path k ∈ Km

od for mode m in nest e; qod denote the demand of mode m in nest e between OD
em
pair (o, d).
Then, we have the following flow conservation equations:
 
em = q̄ ,
qod od
∀od ∈ W (10)
e∈Eod m∈Meod


fkem,od = qod
em , ∀m ∈ Meod , e ∈ Eod , od ∈ W (11)
k∈Km
od

em ≥ 0,
qod ∀m ∈ Meod , e ∈ Eod , od ∈ W (12)

fkem,od ≥ 0, ∀k ∈ Kmod , m ∈ Meod , e ∈ Eod , od ∈ W (13)

42 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62

Fig. 3. Tree structure for the CMSTA.

   
va = fkem,od δa,k
em,od
, ∀a ∈ A (14)
od∈W e∈Eod m∈Meod k∈Km
od

where Eqs. (10) and (11) define demand/flow conservation conditions. Conditions (12) and (13) are the non-negativities.
Constraints (14) define the link-path flow relationship. Eqs. (10–14) define the feasible set  of (v, f, q), where v: =(va ,a
∈ A) is the vector of link flows; f := ( fkem,od , k ∈ Km
od ,m ∈ M od , e ∈ E , od ∈ W ) is the vector of path flows; and with a little
e od
abuse of the notation q := (qod od
em , m ∈ Me , e ∈ Eod ,od ∈ W) is the vector of demands.
For the CMSTA, similar to the case of the linear corridor in Section 2.1, a CNL framework should be adopted for the modal
split component, and the users’ route choice behavior is depicted by the deterministic user equilibrium (DUE) principle.
Thus, Fig. 3 gives the model structure for the CMSTA. In such a framework, the optimal flow solution should fulfill the DUE
conditions:

fkem,od ckem,od − ϕem

od
= 0, ∀k ∈ Kmod , m ∈ Meod , e ∈ Eod , od ∈ W (15)

fkem,od ≥ 0, ckem,od − ϕem

od
≥ 0, ∀k ∈ Kmod , m ∈ Meod , e ∈ Eod , od ∈ W (16)

where ϕem od is the Lagrange multiplier associated with the constraints (11), which represents the travel impedance of using

mode em. At equilibrium, the travel time ckem,od of a used path k ∈ Km od equals ϕ od . All commuters under mode em have
em
identical path travel time, and no traveler can unilaterally reduce his/her travel time by changing routes. As per the DUE
conditions (15) and (16), ϕem od equals the travel time on the shortest path k̄ ∈ K od , which means that for any two nests e, l ∈
m
Eod , ϕem
od equals ϕ od . In this regard, we denote ϕ od as the travel impedance of mode m and thus ϕ od = ϕ od , e ∈ E .
lm m m em od
The optimal demand solution should meanwhile fulfill the following CNL condition:

   μode −1
 1/μod θod od  θod od
od 1/μe
od
αem
od e
exp Vem α exp V
e∈Eod
μod
e i∈Meod
ei
μ od ei
e
qod
m = q̄
od
⎡ ⎤ od (17)
  μl
  1/μod θ
⎣ αlsod l
exp od
Vlsod ⎦
l∈Eod s∈Mlod
μod
l

where μode ∈ [0, 1] is the nest dispersion parameter, θ od ∈ [0, +∞) is the root dispersion parameter and αem denotes the
od
od
degree of membership of mode m in nest e between OD pair (o, d). Vem is the deterministic component of the utility of
mode m in nest e, which can be taken as the negative minimum travel time i.e., Vem od = −ϕ od .
em
The following mathematical programming model is proposed.
 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62 43

Fig. 4. An illustrative example.

[MP–CMSTA]

min Z = Z1 + Z2 + Z3
( v , f , q )∈ 
  va    μod qod
= ta (x )dx + e
qod
em ln
em

a∈A
0
od∈W e∈Eod m∈Meod
θod od 1/μe
αem
od

  1 − μod  
e
+ qod ln qod (18)
od∈W e∈Eod
θod em em
m∈Meod m∈Meod

where Z1 , Z2 and Z3 are similarly defined as in [MP–MS]. The following two propositions are then proposed, rigorous proofs
of which are in the Appendixs A.3 and A.4.

Proposition 3. The optimal solution of the [MP–CMSTA] gives the CNL modal split solution (17) and DUE route choice solution
(15) and (16).
Proposition 4. The [MP–CMSTA] is a convex program and its optimal modal demand and link flow solutions are unique.

2.2.2. A two-stage shortest path searching approach for RPR travel mode
To build solution methods for the [MP–CMSTA] model in a transport network with RPR schemes, it is crucial to first
solve the shortest path problems. This is because loading the demands to the shortest paths (known as all-or-nothing) is
usually an important subroutine for the gradient-based or projection-based solution algorithms. In the multimodal network
with RPR, there are three shortest path problems for the auto, rail and RPR models respectively. The shortest path problems
for auto and rail modes are quite straightforward, which can be directly solved by any existing label-based algorithms in
the corresponding sub-network. However, it is difficult for the RPR mode, because if we simply cope with the multimodal
network G, it cannot be guaranteed that the shortest path between OD pair (o, d) uses a PRP path. This sub-section thus
provides an approach for the shortest-path problem of RPR mode.
In theory, we can enumerate the shortest paths from origin o to each available RPR site, as well as the shortest paths
from each RPR sites to the destination d, add up the two components for each RPR site and eventually find the shortest
one. However, for a medium or large-scale network, such sort of brute-force enumeration could be very tedious. Hence, to
efficiently find the shortest path of RPR mode, we propose a two-stage shortest path searching approach, which is capable
to obtain the shortest RPR path within 2 rounds of searching.
We use an illustrative example in Fig. 4 to further elaborate the problem and approach. We can see that the each RPR
link connects an auto link in set Aa and a rail link Ar . Let h(p) and t(p) denote the head node and tail node of a RPR link, p
∈ Ap , respectively. Herein, t(p) represents a RPR site/car-park, and h(p) is the destination railway station of this RPR service.
In the first stage, in the auto sub-network (remove the rail sub-network), we execute the label-correcting shortest path
algorithm (see, e.g., Sheffi, 1985) to simultaneously obtain the shortest paths from origin node 1 to each RPR site. Then,
44 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62

using k̄(1,t ( p)) to denote the itinerary of the shortest path from origin to the RPR site t(p), and d(1,t(p)) to denote the travel
time on this path.
In the second stage, in the rail sub-network (remove the auto sub-network), adding a dummy link between the origin
and each RPR site, which gives a modified network. The dummy link represents the shortest path k̄(1,t ( p)) and the travel
time on the corresponding dummy link is d(1,t(p)) . Then, executing the second round shortest path searching in the modified
network, and it eventually gives the RPR shortest path.

2.2.3. Evans algorithm

The [MP–CMSTA] is a convex programming problem with linear constraints, thus it can be efficiently solved by the Evans
algorithm (Evans, 1976 and Lam and Huang, 1992), which is essentially a partial linearization method (Patriksson, 2015).
The partial linearized sub-problem is defined by a linear approximation of the objective function (18), by taking a first-
order approximation of the first term (Beckmann’s transformation). The detailed steps are provided as follows:

Step 0: Initiation

(1) Set the iteration counter: n = 0.

(2) Based on the free flow travel time ta0 , get the shortest path k̄ of mode m in nest e, and it gives the minimal travel
od (0 )
 (0 ) em,od
impedance mode m in nest e: ϕem = cem,od (0 ) = ta δ .
k̄ a,k̄
a∈A
(3) Calculate the initial demand of mode m in nest e.
⎛ ⎞
  μode −1
⎜ od 1 /μ od
   1 /μ θod od (0)
od ⎟
exp − μθodod ϕem
od (0 )
⎜ αem e
αeiod e
ϕ exp − ⎟
⎜ e μ od ei ⎟
od ⎜ ⎟
i∈Meod e
od (0 )
qem = q̄ ⎜ ⎡ ⎤ od ⎟
⎜   μl ⎟
⎜   θ ⎟
⎝ ⎣ αlsod
1/μod
l
exp − od ϕls ( ) ⎦
od 0 ⎠
l∈Eod s∈Mlod
μod
l

(4) Perform the all or nothing (AON) assignment to obtain the initial link flow va(0 ) .

Step 1: Descent Direction

(1) Increase the iteration counter: n = n + 1.

(2) Update the link travel time ta(n ) = ta (va(n−1) ), and the shortest route travel time of mode m in nest e: ϕem
od (n )
=
em,od (n )
 (n−1) em,od
c
k̄
= ta δ where k̄ is the shortest path of mode m in nest e between OD pair (o, d).
ak̄
a∈A
(3) Calculate the auxiliary demand of mode m in nest e.
⎛ ⎞
  μode −1
⎜ od 1    1
θod od (n) ⎟
exp − μθodod ϕem
od (n )
⎜ αem μod
e αeiod μod
e exp −ϕei ⎟
⎜ e μod ⎟
od ⎜ ⎟
i∈Meod e
od (n )
q˜em = q̄ ⎜ ⎡ ⎤ od ⎟
⎜   μl ⎟
⎜   θ ⎟
⎝ ⎠
1
⎣ αlsod μod
e exp − od ϕlsod (n) ⎦
l∈Eod s∈Mlod
μod
l

(n )
(4) Perform the AON assignment to assign q˜od em to the shortest path of each mode m in nest e to obtain the auxiliary
(n )
link flow v˜ a .
(n ) (n−1)
(5) Set the descent direction as: (q˜od
em − qod
em ) and (v˜ a(n) − va(n−1) ).

Step 2: Line search

Get an appropriate step size α (n ) , based on e.g. golden section method.
Step 3: Update
(n ) (n−1) (n ) (n−1)
Update: qodem = qod
em + α (n ) (q˜od
em − qod
em ) and va(n) = va(n−1) + α (n−1) (v˜ a(n) − va(n−1) ).
Step 4: Convergence test
If the convergence criterion holds, terminate; otherwise, go to Step 1.

For the convex model [MP–CMSTA], the commonly used Frank-Wolfe algorithm (Sheffi, 1985) is also suitable. However,
Lam and Huang (1992) claimed that Evans algorithm is more robust and converges consistently faster than Frank-Wolfe
algorithm by computational experiments.
 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62 45

3. A general case with asymmetric and non-additive travel time functions

In this section, we further extend the CMSTA model proposed in Section 2.2.1 by relaxing the two Assumptions 1 and
2 to more general cases, which are summarized in Assumptions 3 and 4 below.

Assumption 3. The link travel time functions are asymmetric, i.e., ta is not only affected by its own flow va , but also flows on
some other links, which is expressed as ta = ta (v).

Assumption 4. The path travel time ckem is non-additive to its links, i.e. ckem = ta δa,k
em .
a

Considering realistic situations, e.g., amount of emission, nonlinear valuation of travel times, and travel time uncertainty
(Yang and Huang, 2005), it is therefore necessary to use path-specific travel time variables.
The rationale of the above two extensions is well established and recognized in the literature. Assumption 3 reflects
the interactions between the flows in different links (García and Marín, 2005), for example, a bi-directional road segment
with no median. For Assumption 4, as pointed out by Chen et al. (2012), the additive road cost structure is inadequate for
modeling various transportation policies, environmental and management issues, e.g., the nonlinear transit fare structures,
toll pricing, emission fees, etc. To cope with the non-additivity issue, path-based models and solution methods are needed.
Thus, for the addressed topic in this paper, the two assumptions/extensions lead to an asymmetric path-based CMSTA.

3.1. Variational inequality formulation

It’s well recognized that optimization model is not suitable for the asymmetric cases, and VI models should be alterna-
tively used (Nagurney, 1993). Thus, we propose a VI model for the CMSTA as follows:
[VI–CMSTA]
   
ckem,od (f∗ , q∗ ) fkem,od − fkem,od∗ +
od∈W e∈Eod m∈Meod k∈Kem
od

    , ∀ ( f, q ) ∈  (19)
μod qod∗ 1−μod
θod ln
e em
1/μod
+ θ e ln qod∗
em em − qem ≥ 0
qod od∗
od∈W e∈Eod m∈Meod (αem
od
) e od
m∈Meod

μod qod
where fkem,od∗ and qod∗em is the optimal solution of the problem. We use ψem to denote the term θ
od e
ln em
+
od
od od )1/μe
(αem
1−μod

θ
e
ln em . Let F (x ) = [ck
qod em,od
, ψem
od ]T and x = [ f em,od , qod ]T . The VI model (19) can be simplified to a standard form:
k em
od
m∈Meod

F(x∗ )T (x − x∗ ) ≥ 0, ∀x ∈  (20)

The KKT conditions show that solution of the [VI–CMSTA] fulfills the CNL condition and the DUE conditions in Eqs. (15)–
(17). Detailed descriptions of the KKT conditions of VI can be found in Facchinei and Pang (2003) and Zhou et al. (2009),
which is not further provided here due to the space limit. The VI model addresses a general case for the separable and
additive CMSTA problem in Section 2.2, which is summarized in Proposition 5. Rigorous proof of Proposition 5 is in the
Appendix A.5.

Proposition 5. If the link travel time function is separable and the path travel cost is additive, i.e., ckem,od = ta δa,k
em,od
, model
a∈A
[VI–CMSTA] is equivalent to [MP-CMSTA].

3.2. Solution method for the VI-CMSTA

Projection type algorithms have been thoroughly studied and proven to be efficient in solving traffic equilibrium prob-
lems (Chen et al., 2013; Chen et al., 2012 and Zhou et al., 2009). Amongst them, the gradient projection (GP) is commonly
adopted due to its concise form and good performance, for which a recent survey can be found in Ryu et al. (2017). Given
an initial solution x(0) ∈ , this projection method iteratively executes the following function:
 
x(k+1) = P x(k ) − α k+1 F xk (21)

where P [·] denotes the orthogonal projection map on set  and α k + 1 is a judiciously chosen positive step size. If the
feasible set, , is a non-negative orthant, the projection operator becomes much straightforward: for any vector y ∈ n , we
have P [y] = (x1 , x2 , · · · xn )T , where xi = max(0, yi ), i = 1, 2, · · · n .
However, due to the existence of flow/demand conservation conditions (10) and (12), the feasible set of CMSTA is not a
non-negative orthant. Therefore, in order to efficiently solve the projection operation, we modified the VI model to eliminate
the flow/demand conservation equations, which is introduced in Section 3.2.1.
46 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62

3.2.1. Reformulate the VI model

We first build a non-negative orthant  , based on the non-negative constraints only:
 
 = em |Eqs (12 )∼ Eqs (13 )
fkem,od , qod (22)

Let k̄od
em denote the shortest path for mode m in nest e between OD pair (o, d), and its path flow can be expressed as:

f em,od = qod
k̄ em − fkem,od ∀m ∈ Meod , e ∈ Eod , od ∈ W
k∈Km od (23)
k=k̄od
em

Let ēm̄od denote the mode with minimal travel impedance between OD pair (o, d). The demand of mode ēm̄od can be
expressed as:
 
qod
ēm̄
= q̄od − qod
em ∀od ∈ W
e∈Eod m∈Meod (24)
em=ēm̄od

Replacing f em,od and qod

ēm̄
in Eq. (19) by Eqs. (23) and (24), respectively, gives:
k̄
   
ckem,od∗ fkem,od − fkem,od∗
od∈W e∈Eod m∈Meod k∈Km od

k=k̄od
em ⎛⎛ ⎞ ⎛ ⎞⎞
   ⎜⎜  ⎟ ⎜  ⎟⎟
+ cem,od∗ ⎝⎝qod
em − fkem,od ⎠ − ⎝qod
em − fkem,od∗ ⎠⎠
k̄
od∈W e∈Eod m∈Meod k∈Km od
k∈Km od

  
k=k̄od
em k=k̄od
em ∀ ( f, q ) ∈  (25)
+ ψem
od∗
em −
qod qod∗
em
od∈W e∈Eod m∈Meod
em=ēm̄od ⎛⎛ ⎞ ⎛ ⎞⎞
  ⎜⎜ od   ⎟ ⎜   ⎟⎟
+ ψēod∗
m̄ ⎝⎝
q̄ − em ⎠ − ⎝q̄ −
qod od
em ⎠⎠ ≥ 0
qod∗
od∈W e∈Eod e∈Eod m∈Meod e∈Eod m∈Meod
em=ēm̄od em=ēm̄od

Rearranging Eq. (25) gives a new VI formulation that only has non-negative constraints:
   
 
ckem,od∗ − cem,od∗ fkem,od − fkem,od∗ +
k̄
od∈W e∈Eod m∈Meod k∈Km od

 
k=k̄od
em ∀ ( f, q ) ∈  (26)
ψem
od∗
− ψēod∗
m̄ em − qem ≥ 0
qod od∗
od∈W e∈Eod m∈Meod
em=ēm̄od

With the above transformation, the projection operation in GP algorithm is performed on the non-negative orthant which
is efficient to solve.

3.2.2. Projection algorithm for the model

In this section, we adopt a self-adaptive gradient projection (SAGP) algorithm to solve the [VI-CMSTA]. The SAGP al-
gorithm was originally proposed by He et al. (2002), which was later adopted to solve the asymmetric traffic assignment
problems (Han and Sun, 2004), non-additive problems (Chen et al., 2012), and combined travel demand models (Zhou et al.,
2009). The SAGP can automatically adjust the step size based on the information obtained from the last iteration. Thus, it
avoids the “trial-and-error” step in the Goldstein–Levitin–Polyak (GLP) projection algorithm (Chen et al., 2012). It enhances
the efficiency and robustness of the algorithm, without solving the convex quadratic sub-problems in the GLP projection
algorithm.

Step 0: Initiation

(1) Set parameters of SAGP within the given range: δ ∈ (0, 1), μ ∈ [0.5, 1], ε > 0, α max > 0, α (0) > 0, γ (0) = α (0) .
(2) Set the iteration counter: n = 0.
(0 )
(3) Set the initial solution which should be a feasible solution: qod
em and fkem,od (0 ) .

Step 1: Self-adaptive scaling procedure

(n )
Find the largest α (n + 1) that satisfies the given stop criterion (27). Let α (n+1) = μl γ (n ) , as the parameter μ and
γ are given, the task of finding the largest α
( n ) ( n + 1) becomes finding the smallest non-negative integer l(n) . Set non-
negative integer l(n) = 0 and performing the following procedures, if the given criterion is satisfied, terminate and go
to Step 2; otherwise, let l(n) = l(n) + 1, and perform the following procedures again.
 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62 47

(n+1)
(1) Perform the projection operation to obtain a new solution qod,
em and fkod,em,(n+1) .
od, (n ) od, (n ) (n )
• Let Fem = ψem − ψēod,
m̄
+ ckem,od,(n ) − cem,od,(n ) , perform the projection operation for qod,
em
(n )
.
k̄
 
q˜em ( ) 0, qem ( ) − α (n+1) F ( )
od, n+1
, ∀m ∈ Meod , e ∈ Eod , od ∈ W, em = ēm̄od,(n )
od, n od, n
= max em

(n+1)
• Update q˜od,
ēm̄
according to the demand conservation constraint.
 
q˜ēm̄ (
od, n+1 )
q˜em (
od, n+1 )
= q̄od −
e∈Eod m∈Meod
em=ēm̄od (n )

• Let Fkem,od,(n ) = ckem,od,(n ) − cem,od,(n ) , perform the projection operation for fkod,em(n ) .
k̄
 
od,em,(n+1 ) od,em,(n ) od,em,(n )
f˜k = max 0, fk − α (n+1) Fk
(n )
∀k ∈ Kmod , e ∈ Eod , m ∈ Meod , od ∈ W, k = k̄od,
em

• Update f˜em,od (n+1) according to the flow conversion equation.

k̄
em,od (n+1 )

= q˜em( ) −
od n+1 em,od (n+1 )
f˜ f˜k
k̄
k∈Km
od

k=k̄em ( )
od, n

(2) Stop test:

Let x(n ) = ( fkod,em,(n ) , qod,
em
(n ) T
) , x˜ (n+1) = ( f˜kod,em,(n+1) , q˜od,
em
(n+1) T
) , F(x(n) ) be the mapping of x(n) , and compute
x̄(n+1) = max[0, x(n ) − α (n ) F(x(n ) )]. If the following inequality is satisfied:
2 2
α (n+1) F x(n) − F x˜ (n+1) 
T
(2 − δ )α (n+1) x(n) − x˜ (n+1) F x(n) − F ! x˜ (n+1) +
(α (n+1) ) −(α (n) )   (27)
2 2

≥ max 0, x(n) − x̄(n+1) 2

( α (n ) )
2

go the Step 2. Otherwise, let l(n) = l(n) + 1, update step size α (n + 1) and perform Step 1 again.

Step 2: Selection of γ (n + 1)
If
2 2
α (n+1) F x(n) − F x˜ (n+1) 
T
0.5α (n+1) x(n ) − x˜ (n+1) F x(n ) − F x˜ (n+1) +
!
(α (n+1) ) −(α (n) )  
2 2

≥ max 0, x(n) − x̄(n+1) 2

( )
2
α (n )

is satisfied, γ (n + 1) = max {α (n + 1) /μ, α max }.

Step 3: Convergence test
If the convergence criterion is satisfied, terminate; otherwise, let iteration counter: n = n + 1, go to Step 1.

Compared with Evans algorithm, the SAGP algorithm has a faster convergence rate, and better fitness for the more general
asymmetric and non-additive cases. The Evans algorithm is well-known for its simplicity and low memory requirements. It
is implemented in numerous commercial software packages. The presentation of these algorithms provides flexibility for
practical implementations in terms of different contexts and engineering requirements.
4. Optimal network design for RPR schemes

The equilibrium model proposed in Section 2 can be used to assess any given RPR scheme. This section then proceeds to
discuss about the optimal design of RPR schemes.
The objective of the existing network design problems is usually taken as the system optimum (SO). However, in the
context of CMSTA model, the concept of SO should be extended to the multimodal situations, i.e., multimodal system opti-
mum (MSO). Also, considering that the CNL framework in this paper is a stochastic discrete choice model, the MSO should
be further extended to the stochastic case (Yang, 1999), which leads to the multimodal stochastic system optimum (MSSO).
Thus, in this section, we first propose a formulation for the MSSO.

4.1. Multimodal stochastic system optimum


In the uni-modal road network, the SO is to minimize total travel time, i.e., a∈A va ta (va ), while the stochastic system
optimum (SSO) is to minimize the following expected total travel time (ETTT)

ET T T (f ) = qod Sod cod (28)
od
48 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62

where f and cod represent the vector of path flows fkod and travel time ckod ,k ∈ Kod , respectively. Sod (cod ):=E[ min {ckod }] is
k∈K od
known as the satisfaction function (Sheffi, 1985).
In the case of logit-based SUE, the satisfaction function Sod (cod ) can be expressed as (Yang, 1999):

1  1
Sod cod = − ln exp −θ ckod = ckod + ln fkod (29)
θ θ
k∈K od

and the logit based ETTT is expressed as (Yang, 1999)

[Logit-based ETTT]
1   
ET T Tlogit (f ) = fkod ln fkod + ta (va )va (30)
θ od∈W k∈K od a∈A

Inspired by Yang (1999), in the context of CMSTA, we extend the traditional SSO to the case of multimodal transport
equilibrium problems, i.e. MSSO. First, considering the linear corridor discussed in Section 2.1.1, the satisfaction function for
the multimodal network equals:
μod
  1/μod
  e

Sod ϕ od = − θ1od ln αem

od
exp − μθodod ϕem
od
e

e
e∈Eod m∈Meod (31)
μod pod 1−μod  od
= ϕem
od
+ θ e ln em
od + θod ln
e
pem
od 1/μe
od
(αem ) m∈Meod

where ϕ od is the vector of modal travel impedance, ϕ od := (ϕem

od , e ∈ E , m ∈ M od ); pod denote the chosen probability of
od e em
mode m in nest e between OD pair (o, d), we have:
μod pod 1 − μod 
pod
em S
od
em ϕem +
ϕ od = pod od e
pod
em ln
em
+ e
pod
em ln pod
θod αem
od 1 / μ od
e θ od
em
m∈Meod

and
    μod pod 1 − μod 
pod
em S
od
ϕ od = em ϕem +
pod od e
pod
em ln
em
+ e
pod
em ln pod
e∈Eod m∈Meod e∈Eod m∈Meod
θod αem
od 1 / μ od
e θ od
em
m∈Meod

which gives
  μod pod 1 − μod 
Sod ϕ od = em ϕem +
pod od e
pod ln em
+ e
pod
em ln pod
e∈Eod m∈Meod
θod em αem
od 1 / μ od
e θ od
em
m∈Me
od

Therefore, multimodal ETTT for the linear corridor, denoted as ETTTCNL , equals:
  
ET T TCNL (q ) = em ϕem
qod od

od∈W e∈Eod m∈Meod

   μod qod 1 − μod 

e em e
+ qod ln + qod
em ln qod
od∈W e∈Eod m∈Meod
θod em αem
od 1 / μ od
e θ od
em
(32)
m∈Me
od

 μod pod 1 − μod 

e e
− pod ln + pod ln pod
θod αem
od 1 / μ od
e θ od
od∈W m∈Me
od

where the last term in Eq. (32) is fixed, and thus can be eliminated. We have:
  
ET T TCNL (q ) = em ϕem
qod od

od∈W e∈Eod m∈Meod

   μod qod 1 − μod 

e em e
+ qod
em ln + qod
em ln qod (33)
od∈W e∈Eod m∈Meod
θod od 1/μe
αem
od
θ od
em
m∈Meod

We proceed to cope with the whole network by extending Eq. (33). The multimodal ETTT in the context of CMSTA is
given as
   
MET T T (f, q ) = fkem,od ckod,em
od∈W e∈Eod m∈Meod k∈Km
od

   μod qod   1 − μod  

e em e
+ qod
em ln + qod
em ln qod (34)
od∈W e∈Eod m∈Meod
θod αem
od 1 / μ od
e
od e∈Eod
θ od
em
m∈Me
od m∈Me
od
 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62 49

which is equivalent to:


MET T T (v, q ) = ta va
a∈A
   μod qod   1 − μod  
e em e
+ qod
em ln + qod
em ln qod (35)
od∈W e∈Eod m∈Meod
θod αem
od 1 / μ od
e
od e∈Eod
θ od
em
m∈Meod m∈Meod

The MSSO model minimizes the METTT, over the set (v, q) ∈ . Note that the above MSSO models are built in the CNL
framework (see Fig. 3) and thus could be termed as a CNL-based MSSO. It could be easily extended to the other types of
discrete-choice model framework. Based on the KKT conditions, it is straightforward to see that the optimal solution of
these models can fulfill the conditions summarized in Proposition 6, proof of which is not further included.

Proposition 6. Optimal solution of the MSSO model gives the CNL-based modal split solution (17) and DUE route choice solution
(15) and (16), in terms of the marginal travel time function, t˜a (va ) = ta (va ) + va dtda (vva ) , ∀a ∈ A.
a

4.2. MPEC model for the optimal network design of RPR

Let y = (ya ,a ∈ AP ) denote the capacity of potential RPR sites. For any ya , it only takes integer values and all the feasible
values are grouped in set Ia . If ya =0, it implies that RPR site a will not be built. For each feasible capacity value of ya , we
further introduce a binary vairable zai , ∀i ∈ Ia . Herein, zai =1 means ya takes the value i ∈ Ia . Clearly,
zai ={0, 1}, ∀a ∈ AP , i ∈ Ia (36)


zai =1, ∀a ∈ AP (37)
i∈Ia

ya = i × zai , ∀a ∈ AP (38)
i∈Ia

Define the vector z := {zai , a ∈ AP , i ∈ Ia } and Eqs. (36)–(38) define the feasible set z of z. Taking minimizing METTT in
Eq. (35) as the objective function, the optimal RPR network design problem is formulated as the following MPEC model:
[NDP–RPR]
  
min MET T T (z, v, q) = ta (va )va + tai (va )va zai
z ∈ z
a∈A\A P a∈AP i∈Ia \{0}

   μod qod
e em
+ qod
em ln
od∈W e∈Eod m∈Meod
θod od 1/μe
αem
od

  1 − μod  
e
+ qod ln qod (39)
od∈W e∈Eod
θod em em
m∈Meod m∈Meod

subject to

τa y a ≤ B (40)
a∈AP

va ≤ bigM · zai , ∀a ∈ AP (41)
i∈Ia \{0}

  va    va
(v(z ), q(z )) = arg min ta (x )dx + zai tai (x )dx
( v , q )∈  a ∈ A \ A P 0 a∈AP i∈Ia \{0}
0

   μod qod
e em
+ qod ln
od∈W e∈Eod m∈Meod
θod em od 1/μe
αem
od

  1 − μod  
e
+ qod
em ln qod (42)
od∈W e∈Eod
θod em
m∈Meod m∈Meod

Constraint (40) reflects the budget limit, where τ a is the unit construction cost for each link a ∈ Ap . Constraints (41) en-
sure that the RPR link flows can take possible value only if its capacity is greater than zero. Constraints (42) guarantee that
for any RPR network plan z, the equilibrium flow and demand pattern (v(z), q(z)) should fulfill the CMSTA model, see Model
(18). For any feasible solution( v, q ) ∈ , we use Ob jCMSTA (z, v, q) to denote the function value in the right-hand-side of (42),
namely:
50 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62

    va    va
Ob jCMSTA z, v, q = ta (x )dx + zai tai (x )dx
0 0
a∈A\A P a∈AP i∈Ia \{0}
od
   μod od qem
e
+ qem ln
od∈W e∈Eod m∈Meod
θod od 1/μe
αem
od

  1 − μod  od  od
e
+ qem ln qem (43)
od∈W e∈Eod
θod
m∈Meod m∈Meod

Thus, constraints (42) is equivalent to:

   
Ob jCMSTA (z, v, q) ≤ Ob jCMSTA z, v, q , ∀ v, q ∈  (44)

5. Solution method

5.1. An exact algorithm for the optimal RPR network design problem

The proposed [NDP-RPR] is, per se, a NP-hard problem. The involved CNL model and multimodal contexts have further
increased the complexity and challenges of solving this problem. The [NDP-RPR] is clearly a mixed integer nonlinear pro-
gram. The concept of valid inequalities is used in the existing literature for mixed integer nonlinear problems in general
and also for transport network design (e.g., Bertsimas and Weismantel, 2005; Lee and Leyffer, 2012 and Wang et al., 2013),
and we apply a similar approach to solve the addressed problem in this paper. The algorithm introduced below can be
regarded as an nonlinear valid inequalities (NVI) method. For the [NDP-RPR] model, we can easily relax the equilibrium
constraints (44), and then get a lower bound of the optimal solution. We use z̄ to denote such a lower bound solution, and
it’s straightforward to solve the corresponding CMSTA equilibrium link flows and modal demands (v̄, q̄ ) w.r.t. z̄. Then, a cut
can be added back to the relaxed model: Ob jCMSTA (z, v, q) ≤ Ob jCMSTA (z̄, v̄, q̄ ). Note that feasible solution (z̄, v̄, q̄ ) also gives
an upper bound for the optimal solution of Model [NDP-RPR]. Detailed procedure of this NVI method are summarized as
follows:
¯ z(n ) contains all the generated solution z, and inilitize it 
Step 0: Set n = 0. Let set  ¯ z(1 ) =∅. Define set 
¯ (n ) to contain all
( )
the generated link flows and modal demands, and inilitize it  = ∅. We define the upper bound UB(n) = +∞, and
¯ 1

lower bound LB(n) = 0.

Step 1: Set n = n + 1, and solve the relaxed [NDP-RPR] problem below:
[RNDP]
min MET T T (z, v, q) (45)
z ∈  z , ( v , q )∈ 

subject to
  
1 − z̄ai zai + z̄ai 1 − zai ¯ z(n )
≥ 1, ∀z̄ ∈  (46)
a∈AP i∈Ia
   
¯ (n )
Ob jCMSTA (z, v, q) ≤ Ob jCMSTA z, v, q , ∀ v, q ∈  (47)

Let (z̄(n ) , v̄(n ) , q̄(n ) ) denote the optimal solution of [RNDP] and its corresponding objective function value is a new lower
bound, represented by LB(n) .
Step 2: Based on the new RPR scheme z̄(n ) , solve the CMSTA problem, which gives the equilibrium link flows and modal
(n ) (n )
demands (v̄CMSTA , q̄CMSTA ). The corresponding objective function value of this CMSTA is an auxiliary UB (n) . If UB (n) ≤
UB(n) , set UB(n) = UB (n) .
Step 3: If LB(n) ≥ UB(n) (1 − ε ), stop and output z̄(n ) . Otherwise, set  ¯ z(n+1) =
¯ z(n ) ∪ {z̄(n ) } and 
¯ (n+1) =
¯ (n ) ∪ {(v̄(n ) , q̄(n ) )} ∪
(n ) (n )
{(v̄CMSTA , q̄CMSTA )}. Go to Step 1.
Note that the tolerance ε in Step 3 is a trival positive value, i.e., ε > 0. In the NVI method, constraints (46) exclude one
solution each time. Constraints (47) are the added cuts to tighten the feasible set. In Step 1, the [RNDP] needs to be solved,
which is also a challenging issue, and we provide a method in Section 5.2 for the [RNDP], which is taken as a subroutine of
the whole algorithm.

5.2. Origin-based formulation of [RNDP]

We point out that the Model [RNDP] is highly challenging to be directly solved, because the feasible set is defined with
the path flows fkem,od and the number of paths is infinite in a real transport network. Thus, we proceed to proprose an
origin-based reformulation of [RNDP], to avoid the paths.
 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62 51

For each feasible capacity value i ∈ Ia of RPR link a ∈ AP , we introduce a link flow via . Clearly, via > 0 only if zai = 1. For
each link a ∈ A, let t(a) ∈ N represent its tail node and h(a) ∈ N represent the head node. Let O denote the set of origins in
the network, O⊆N; Wo denote the set of OD pairs originating from origin o; Eo denote the set of nests and Meo denote the
set of travel modes in nest e w.r.t. origin o; vem,o
a is the flow on link a with mode m in nest e w.r.t. origin o. Therefore, an
origin-based reformulatin of [RNDP] is given as:
[ORNDP]
  
min MET T T (z, v, q) = ta (va )va + tai via via
z ∈ z
a∈A\A P a∈AP i∈Ia \{0}

   μod qod
e em
+ qod
em ln
od∈W e∈Eod m∈Meod
θod od 1/μe
αem
od

  1 − μod  
e
+ qod ln qod (48)
od∈W e∈Eod
θod em em
m∈Meod m∈Meod

subject to
⎧  od
⎨ qem , j = o
  (o,d )∈W
o
vem,o − vem,o = , ∀ j ∈ N, m ∈ Meo, e ∈ Eo, o ∈ O (49)
a∈A
a
a∈A
a
⎩ 0, j = o, j = d
t (a )= j h(a )= j −qem , j = d
od

  
va = vem,o
a , ∀a ∈ A (50)
o∈O e∈Eo m∈Meo

vem,o
a ≥ 0, ∀m ∈ Meo, e ∈ Eo, o ∈ O, a ∈ A (51)

v0a = 0, ∀a ∈ AP (52)


va = via , ∀a ∈ AP (53)
i∈Ia \{0}

via ≤ bigM · zai , ∀a ∈ AP , i ∈ Ia \{0} (54)

and constraints (10), (12), (46) and (47).
Constraints (49)–(53) are the flow conservation equations. Constraints (54) ensure that via > 0 only if zai = 1, and bigM
here is a larger positive value. With a little abuse of notaiton, we still use (v, q) hereafter for the origin-based link flows and
modal demands;  to denote their feasible set. It is easy to prove that the objective function (48) is convex with via , and the
model [ORNDP] only has linear constraints. Thus, [ORNDP] is a convex mixed-integer nonlinear programming problem. Some
techniques are further used in Section 5.3 to linearize the terms in this model, in order to use the existing mixed-integer
linear programming solvers, e.g., Gurobi.

5.3. A combined tangent line and tangent plane approximation algorithm

The nonlinear terms in model [ORNDP] are approximated/linearized by the tangent lines and tangent planes introduced
in this section, which forms a combined tangent line and tangent plane approximation algorithm for the [ORNDP]. Note that
this algorithm is incorporated in the NVI method in Section 5.1 as a subroutine for the RPR network design problem.
e (q ), and [ORNDP] is rewritten as follows:
For the sake of presentation, we define two new functions h(v) and god
    
min MET T T (z, v, q) = ta (va )va + tai via via + e (q )
god (55)
z ∈  z ( v , q )∈ 
a∈A\ AP a∈AP i∈Ia \{0} od∈W e∈Eod

 μod qod 1 − μod  

e (q ) :=
e em e
god qod
em ln + qod ln qod (56)
θod od 1/μe
αem
od
θod em em
m∈Meod m∈Meod m∈Meod

  va    via
h(v ):= ta (x )dx + tai (x )dx (57)
0 0
a∈A\ AP a∈AP i∈Ia \{0}

  
1 − z̄ai zai + z̄ai 1 − zai ≥ 2, ∀z̄ ∈ 
¯z (58)
a∈AP i∈Ia
52 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62

         
h (v ) + e (q ) ≤ h v +
god god
e q , ∀ v, q ∈ 
¯ (59)
od∈W e∈Eod od∈W e∈Eod

We can see that functions god e (q ) and h(v) defined in (56) and (57) are also convex ones w.r.t. (v, q).
Firstly, for the linearization of two convex terms ta (va )va and tai (via )via in the objective (55), some approximation tangent
lines are built based on the link flow solution in  ¯ , see Eq. (47). We introduce some intermediate variables a ,∀a ∈ A\AP
and ai , ∀a ∈ AP , i ∈ Ia \{0}, and it gives:
    
min MET T T (z, v, q) = a+
i
a + e (q )
god (60)
z ∈  z ( v , q )∈ 
a∈A\A P a∈AP i∈Ia \{0} od∈W e∈Eod

subject to
        2
a ≥ ta v a + ta v a v a · va − ta v a v a , ∀a ∈ A\AP , v a ∈ 
¯ (61)

%     &  
i
a ≥ tai via + tai via via · via − tai via vi,a 2 , ∀a ∈ AP , i ∈ Ia \ {0}, vai ∈  (62)

and constraints (56)–(59).

Similarily, h(v) in Eq. (59) could be also relaxed as linear constraints. We introduce intermediate decision variables a ,∀a
∈ A\AP and ia , ∀a ∈ AP , i ∈ Ia \{0}, and thus:
            
a+
i
a + e (q ) ≤ h v +
god god
e q , ∀ v, q ∈ 
¯ (63)
a∈A\A P a∈AP i∈Ia \{0} od∈W e∈Eod od∈W e∈Eod

   va  
a ≥ ta v a va + ta (x )dx − ta v a v a , ∀a ∈ A\AP , v a ∈ 
¯ (64)
0

 i  va
i
 i i i
i
a ≥ tai
v a via + tai (x )dx − tai v a v a , ∀a ∈ AP , v a ∈ 
¯ (65)
0

As per the proof for Proposition 4, the Hessian matrix of god e (q ) is positive definite, which means ge (q ) is a convex
od

function. As god
e ( q ) is a multi-variant convex function, some tangent planes are used to relax/approximate this term. We
introduce intermediate decision variables eod , ∀e ∈ Eod , od ∈ W . Then, [ORNDP] can be reformulated as:
    
min MET T T (z, v, q) = a+
i
a + eod (66)
z ∈  z ( v , q )∈ 
a∈A\A P a∈AP i∈Ia \{0} od∈W e∈Eod

subject to:
 
   ∂ god
e q  od

eod ≥ god em − q em , ∀ q ∈ 
q + qod ¯ (67)
e
m∈Meod
∂ qod
em

            
a+
i
a + eod ≤ h v + god
e q , ∀ v, q ∈ 
¯ (68)
a∈A\A P a∈AP i∈Ia \{0} od∈W e∈Eod od∈W e∈Eod

and constraints (58), (61), (62), (64) and (65).

The above model is evidently a mixed linear integer program, which can be efficiently solved by some commercial
solvers.

6. Numerical example

In this section, we provide two numerical examples to validate the proposed models and methods. The first example (a
linear corridor) focuses on the network equilibrium, while the second one discusses the optimal RPR scheme design.
4
The BPR type function is adopted for the link travel time i.e., ta = ta0 (1 + 0.15(va /Ca ) ) for road links, and ta =
4
ta (1 + 0.1(va /Ca ) ) for rail links. As to any RPR link, it represents the whole transfer behavior of a user, thus the travel
0

impedance ta , a ∈ AP on such a link should inlcude a flow-dependent travel time as well as a fixed waiting time for the
4
express buses; e.g., ta (va , ya ) = 1/ fa + ta0 (1 + 0.1(va /ya ) ), a ∈ AP , where fa represents the frequency of express bus services
on a, and 1/fa is the average waiting time (Spiess and Florian, 1989). Herein, ta0 is the free flow travel time for each link. The
CNL dispersion parameters are set as θ o = 0.5 (θ o ≥ 0), μe = 0.5 (0 ≤ μe ≤ 1).
 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62 53

Table 1
Attributes of the linear corridor.

Link type Auto Rail RPR (Dummy link)

Link no. 1 2 3 4 5
Free flow travel time 6 8 10 6 3
Capacity 120 75 100 100 100

Fig. 5. Four choice structure.

6.1. A linear corridor

The linear corridor shown in Fig. 1 is used, the data for which is given in Table 1. Its RPR route consists of: link 1→link
5→link 4. The OD demand is 150. Four scenarios are considered as a sensitivity test of the CMSTA model. Scenario 1 consid-
ers various nest structures and its effect on mode share. Scenario 2 further excavates the feature of CNL model by varying
the degree of membership. Scenario 3 investigates the combined effect of CNL dispersion parameter θ e . Scenario 4 tests the
effect of congestion under different demand level.

6.1.1. Effect of the mode similarity

We consider three types of discrete choice structure: MNL, NL, and CNL, as shown in Fig. 5. For the NL model, two cases
are proposed to represent different mode similarity configurations. Compared with MNL, NL partially captures the mode
similarity by using a two-level nested tree structure (Kitthamkesorn et al., 2016), where correlated modes are nested.
The equilibrium flows of each choice structure are calculated, which is shown in Fig. 6(a) and (b). The evident difference
between each mode reflects the significance of choosing a suitable choice structure. From Fig. 6(c), we know that NL mea-
sures mode similarity by the dispersion parameter on the nest level, μe ∈ [0, 1]. As μe reduces, the similarity increases.
Therefore, the train mode in Case 1 or auto in Case 2 (doesn’t share the upper nest with other modes) obtains higher
demand, compared with the correlated modes (auto and RPR in Case 1 or train and RPR in Case 2).
As shown in Fig. 6(c), CNL extends NL by introducing the degree of membership α em ∈ [0, 1] where each mode may
belong to multiple nests. In each nest, compared with the mode belonging to a single nest (auto/train), the case of multiple
nests would be assigned a smaller α em . To sum up, the effect of mode similarity is non-negligible. Compared with correlated
modes, the independent mode tends to have a larger amount of demand.

6.1.2. Effect of the degree of membership

We proceed to examine the degree of membership α em . As the degree of membership of road nest α e1 for RPR mode
varies from 0 to 1, the corresponding mode share of each travel mode in the CNL model is given in Fig. 7. We can see that
the mode share (demand) is highly affected by the changes in α e1 . As α e1 increases from 0 to 1, RPR has higher (lower)
degree of membership in road (rail) nest, thus when α e1 equals zero, the model becomes NL (Case 2) and when α e1 equals
1 the model becomes NL (Case 1). The two dotted lines indicate the demand of RPR from two sources: the road nest e1 and
rail nest e2 , respectively. The sum of these two terms equals the total RPR demand. These two lines show that when α e1
increases from 0 to 1, the demand stems from road nest increases while the rail nest decreases.
Fig. 8 shows the equilibrium modal travel impedance ϕm od produced by the CNL model, which strengthens the phe-

nomenon that less travel time does not always attract more travel demand (which is quite different with the conventional
model, e.g., MNL). Even though the travel impedance of RPR is less than the train, smaller α em impedes its utility and results
54 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62

Fig. 6. Mode share of each travel mode corresponding to all four models.

Fig. 7. Effect of the degree of membership.

into lower mode share. The vertical axis on the right-hand-side of Fig. 8 indicates the changes in the expected perceived
travel impedance (EPI), i.e., satisfaction Sod (ϕ od ), of each model: MNL, NL and CNL. The EPI of NL is obtained when α e1
equals 1 (NL Case 1) or 0 (NL Case 2). Compared with the fixed EPI produced by MNL, CNL produces a variant and larger
EPI, which further implies the significance of considering the mode similarity. The highest EPI-CNL is achieved when α e1
takes a value of about 0.4.

6.1.3. Effect of the CNL dispersion parameters

This section examines the combined effect of CNL dispersion parameters (μe ). The road nest parameter μe1 and rail
nest parameter μe2 varies from 0 to 1, respectively. When μe = 1, the CNL structure of nest e would collapse to the MNL
structure. When μe = 0, the CNL choice structure becomes a deterministic choice pattern; the alternative(s) with smallest
travel cost takes all the demand while no user takes the other modes.
As shown in Fig. 9(a), RPR has no users from road nest when μe1 = 0. As μe1 increase, RPR mode becomes more inde-
pendent under nest e1 (road) and shares an increasing number of network users from e2 . Similar conclusion can be drawn
from Fig. 9(b).
 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62 55

Fig. 8. Equilibrium modal travel impedance and EPI.

Fig. 9. RPR demand with different CNL dispersion parameters.

6.1.4. Effect of the congestion

We further investigate the congestion effect under different demand levels. The OD demand varies from 50 to 300 with
an increment of 50. As the OD demand increases, the MNL, NL and CNL models show similar variation tendency as shown
in Fig. 10. In addition, the difference in the mode share of each model becomes smaller as the OD demand increases, which
reflects that the modal similarity has little impacts in such cases.

6.2. Sioux-Falls network

Another numerical example built from the Sioux-Falls network (Fig. 11) is then used for the optimal design of RPR
facilities. It has 28 nodes and 76 links. The area highlighted by the red ellipse is the CBD district. It has 4 rail lines between
the suburb areas and city loop. Detailed attributions of this network are given in Appendix B.1.
We focus on the travel demands in commuting hours, where majority of demands originate from the suburbs and go
into business districts. There are totally 40 OD pairs, and the OD demand data are provided in Appendix B.2. Four sites
(nodes 4, 6, 15, 19) outside the city center are candidate RPR car parks. The feasible capacity of each RPR site (represented
by the dummy links) is assumed to be (0, 200, 250, 300, 350, 400, 450, 500). Under this setting, the number of feasible
56 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62

Fig. 10. Mode share of each mode under different OD demand levels.

Fig. 11. Sioux-Falls network.

 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62 57

Table 2
Optimal RPR facility location and capacity design.

Total budget Candidate RPR nodes METTT

4 6 15 19

10 0 0 250 Do not build 300 450 208,956.1

1200 350 Do not build 350 500 208,410.8
1400 450 Do not build 450 500 208,030.7
1600 450 200 450 500 207,862.8
1800 500 300 500 500 207,716.9
20 0 0 500 500 500 500 207,716.2

Fig. 12. Convergence trend of the lower bound and upper bound values.

RPR network pattern is 84 = 4096. The frequency of the dedicated shuttle buses are assumed to be 4, 7, 5, 5 (per hour) for
the transfer links 4, 6, 15, 19, respectively. Other settings are similar to those of the small network above, which are not
repeated here.
The NVI method is then used to solve the optimal RPR network plan. The problem is coded in Python calling Gurobi
running on a 2.2 GHz Dual Core laptop with 8 GB of RAM. The optimal solution in terms of different budget levels are
provided by Table 2. We can see that when budget increases, the system optimum METTT can be improved, until the budget
reaches 1800. It implies that 1800 is a suitable budget level to be provided.
The convergence trend of the NVI method is provided in Fig. 12 when the budget is 1800. The lower bound evidently
increases from 203,824.6 to 205,734.26 within 8 iterations, where a 0.01-optimal solution is obtained. Considering that there
are totally 4068 feasible solutions, the algorithm can efficiently obtain a 0.01-optimal solution after only exploring 0.19% of
the feasible set. Performance of the NVI method is further tested for 102 iterations until converge. We can see that a suitable
ε can largely reduce the workload while only slightly compromising the accuracy.

7. Conclusion

In this paper, we proposed a new P&R service scheme, termed as RPR. RPR users first drive their car to a RPR site, park
their car, take a dedicated express bus line to a train station, and then take a train to their destination. A CNL-based CMSTA
model was proposed to describe the route choice and mode choice behavior of network users. Convex programming and VI
models were provided, respectively, to consider the symmetric and asymmetric situations. Further, two solution algorithms
based on Evans algorithm and SAGP algorithm were developed. To optimize the RPR network plan, taking the CMSTA model
as a constraint, we proposed a MPEC model with the objective of minimizing METTT. An exact solution algorithm was
employed based on the NVI to solve this problem. Numerical examples showed that the proposed RPR scheme is effective
in congestion mitigation. For future research, the network capacity, vulnerability and flexible analysis for RPR services could
be considered, as well as the corresponding solution algorithms.

Acknowledgment

This study is supported by the General Projects (No. 71771050), Youth Projects (No. 71501038), and Key Projects (No.
51638004) of the National Natural Science Foundation of China.
58 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62

Appendix

A.1. Proof of Proposition 1

The Lagrangian of the [MP–MS] with respect to the equality constraint (6) and (7) can be formulated as:

 
L = Z + φ q̄ − qem (69)
e∈E m∈Me

where ϕ is the Lagrange multiplier associated with the demand conservation constraint (6). At the stationary point of the
Lagrangian, the following conditions should be satisfied by calculating the partial derivative of L w.r.t. demand qem :
∂L μe qem 1 − μe 
= ϕem + ln + ln qem − φ = 0 (70)
∂ qem θ α
( em )1 / μe θ m∈Me

rearranging Eq. (70) gives:

1−μe
μe  
 1 θ
qem qem = (αem ) μe exp ( φ − ϕ em ) (71)
m∈Me
μe
1
μe    
 θ  1 θ
qem = exp φ (αem ) μe exp − ϕem (72)
m∈Me
μe m∈Me
μe

  μe
  1 θ
qem = exp (θ φ ) (αem ) μe exp − ϕ , and (73)
m∈Me m∈Me
μe em

  μe
    1 θ
qem = q̄ = exp (θ φ ) (αem ) μe exp − ϕ (74)
e∈E m∈Me e∈E m∈Me
μe em

Dividing Eq. (71) by Eq. (72) gives the conditional probability of choosing nest e:
1
θ
qem (αem ) μe exp − μe ϕem
 = pm|e =   (75)
qem  1 θ
(αes ) e exp − ϕes
μ
s∈Me
s∈Me
μe
Dividing Eq. (73) by Eq. (74) gives the marginal probability:
  μe
  θ 1
qem (αem ) exp − ϕem μe

m∈Me
μe
m∈Me
 = pe =   μl (76)
qem   1 θ
l∈E s∈Me (αls ) exp − ϕls
μl

l∈E s∈M
μl
l

Hence, these marginal and conditional probabilities fulfill the CNL conditions.

A.2. Proof of Proposition 2

It is sufficient to prove that objective function (5) of [MP–MS] is strictly convex in the vicinity of demand qem and the
feasible region is convex. The convexity of feasible region is assured by the linear constraints (6)–(8). The non-negative
constraints (9) do not alter this characteristic. The Hessian matrix of the objective (5) w.r.t. the demand qem can be defined
as:
⎧
⎪ ∂t μ 1−μ
⎨ a δael + e +  e > 0; e = l, m = s
∂ 2Z ∂ 2 ( Z1 + Z2 + Z3 ) ∂ v a θ qem θ qem
= = (77)
∂ qem ∂ qls ∂ qem ∂ qls ⎪
⎩ m∈M e
0; e = l, m = s
which implies the positive definiteness of the matrix. Therefore, the solution of the proposed model is unique.
 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62 59

A.3. Proof of Proposition 3

The Lagrangian of the [MP–CMSTA] w.r.t. the equality constraints (10) and (12) can be formulated as:

      
L=Z+ ϕem
od
qod
em − fkem,od + φod q̄od − qod
em (78)
od∈W e∈E m∈Me k∈Kod
em od∈W e∈Eod m∈Meod

where ϕ od is the Lagrange multiplier associated with the modal demand conservation constraints (10) and ϕem
od is the La-

grange multiplier associated with the flow conservation constraints (12).

(1) At the stationary point of the Lagrangian, the following conditions should hold w.r.t. the path flow f kem,od :

∂L 
= ta δa,k
em,od
− λod em,od
em = ck − ϕem
od
=0 (79)
∂ fkem,od a∈A

rearranging Eq. (79):

ckem,od =ϕem
od
(80)

which implies the travel time ckem,od

of a used path k equals ϕem
od .
All commuters under mode em between OD pair (o, d)
have identical path travel time, and no traveler can unilaterally reduce his/her travel time by changing routes. Thus, the
user equilibrium condition is obtained.
(2) Now we consider the mode choice:
∂L μod qod 1 − μod 
em + ϕem − φod = 0
e em e
= od ln + ln qod od
(81)
∂ qem θ
od
αem
od 1 / μ e θ od
m∈Me

From Eq. (80), ϕem

od equals the shortest path travel time of mode m in nest e between OD pair (o, d). Rearranging

Eq. (81) gives:

1−μod
e
μeod
 
 1
θod
qod qod = αod μod
e exp φod − ϕem
od
(82)
em
m∈Me
em em
μod
e

1
μod
   
 e
θ od  1
θ od od
qod = exp φod α od μod
e exp − od ϕem (83)
m∈Me
em
μod
e m∈Me
em
μe

  μod
e
  1
θ
qod = exp (θod φod ) α od μod
e exp − od ϕem
od
, (84)
m∈Me
em
m∈Me
em
μod
e

and
⎛ ⎞
  μod
e
    1
θ
em = exp (θod φod )
qod ⎝ αem
od μod
e exp − od ϕem
od ⎠ (85)
e∈E m∈Me e∈E m∈Me
μod
e

Dividing Eq. (83) by Eq. (82) gives the marginal probability of choosing nest e:
1  
αem
od μod
e exp − μθodod ϕem
od
qod
em = p m |e =
e
  (86)
qod  1
θ
em
α od μod
e exp − od ϕesod
s∈Me
s∈Me
es
μod
e

Dividing Eq. (85) by Eq. (84) gives the conditional probability:

  μod
e
  1
θ
qod α od μod
e exp − od ϕem
od

m∈Me
em
m∈Me
em
μod
e
 = pe = ⎛ ⎞ (87)
qod   μod
l
em   1
θ
l∈E s∈Me ⎝ αlsod μod
l exp − od ϕlsod ⎠
l∈E s∈Ml
μod
l

Hence, these marginal and conditional probabilities fulfill the CNL conditions.
60 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62

A.4. Proof of Proposition 4

It is sufficient to prove that objective function (18) of [MP-CMSTA] is strictly convex and that the feasible region is convex.
The convexity of the feasible region is assured by linear equality constraints (10) and (12). The non-negative constraints
(11) and (13) do not alter this characteristic.
The Hessian matrix of Z1 w.r.t. the link flow can be defined as:
∂ 2 Z1 ∂ ta
= ∂ va ≥ 0 a=b
(88)
∂ va ∂ vb 0 a = b
which implies the positive definite matrix. The Hessian matrix of Z2 + Z3 w.r.t. the modal demand variables can be defined
as:
⎧ μod 1−μod
⎨θod efemod + e >0 e = l, m = s
∂ ( Z2 + Z3 )
2
θod od
fem
= (89)
∂ qod
em ∂ qls
od
⎩ m∈M e
0 e = l, m = s
which also implies the positive definiteness of the matrix. Therefore, the [MP–CMSTA] has a unique solution with respect to
both link flows and modal demands.

A.5. Proof of Proposition 5

The objective function (18) of [MP–CMSTA] is continuously differentiable. The feasible region composed of linear equality
constraints (10) and (12), and the non-negative constraints (11) and (13) is closed and convex. The first-order partial deriva-
tion of the objective function (18) w.r.t. the route flow variables fkem,od and modal demand variables qodem can be expressed
as:
  va
∂ ta (x )dx
∂Z a∈A
0 
= = ta δa,k
em,od
= ckem,od (90)
∂ fkem,od ∂ fkem,od a∈A

∂Z μod qod∗ 1 − μod 

em = ψem
e em e
= ln + ln qod∗ od
(91)
∂ qod
em θ od αem
od 1 / μ od
e θ od
m∈Me
od

where ckem,od and ψem od is the mapping of route flow and modal demand, formulated in [VI–CMSTA]. Therefore,

Proposition 1.2 of Nagurney (1993) indicates the [VI–CMSTA] is equivalent to [MP–CMSTA].

B.1. Link attributes of Sioux-Falls network

T H FFT C Type T H FFT C Type T H FFT C Type

1 2 10 600 Auto 15 14 2 900 Auto 27 26 2 1100 Train

1 3 3 900 Auto 15 19 2 600 Auto 28 27 7 1100 Train
2 1 10 600 Auto 16 8 3 300 Auto 29 28 9 1100 Train
2 6 3 900 Auto 16 10 3 300 Auto 30 29 3 1100 Train
3 4 4 1200 Auto 17 10 7 300 Auto 31 26 2 1300 Train
3 12 6 1200 Auto 17 16 3 600 Auto 31 40 2 1300 Train
4 5 5 900 Auto 18 7 4 600 Auto 32 31 2 1100 Train
4 11 7 600 Auto 18 16 5 600 Auto 33 32 12 1100 Train
5 6 3 900 Auto 19 15 2 600 Auto 34 33 4 1100 Train
5 9 5 900 Auto 19 17 4 1200 Auto 35 34 4 1100 Train
6 5 3 900 Auto 20 18 11 600 Auto 36 35 2 1100 Train
6 8 6 900 Auto 20 19 5 600 Auto 37 26 2 1300 Train
7 8 5 300 Auto 20 21 2 600 Auto 37 40 2 1300 Train
8 9 3 300 Auto 20 22 2 300 Auto 38 37 2 1100 Train
8 16 3 300 Auto 21 20 2 600 Auto 39 38 12 1100 Train
9 8 3 300 Auto 21 22 4 600 Auto 40 31 2 1300 Train
9 10 3 300 Auto 21 24 2 600 Auto 40 37 2 1300 Train
10 9 3 300 Auto 22 15 4 900 Auto 41 40 2 1100 Train
10 16 3 300 Auto 22 23 2 600 Auto 42 41 7 1100 Train
11 10 6 600 Auto 23 14 4 600 Auto 43 42 8 1100 Train
12 11 5 600 Auto 23 22 2 600 Auto 4 27 2 RPR
13 12 10 900 Auto 24 13 2 600 Auto 6 38 2 RPR
13 24 2 600 Auto 24 21 2 600 Auto 19 41 2 RPR
14 11 5 300 Auto 24 23 4 600 Auto 15 32 2 RPR
14 15 2 1200 Auto 26 31 2 1300 Train
15 10 8 900 Auto 26 37 2 1300 Train
∗
T is Tail node; H denotes Head node; C is Capacity; FFT is Free Flow Travel Time.
 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62 61

B.2. OD travel demand of Sioux-Falls network

OD_ID Origin Destination Demand OD_ID Origin Destination Demand

1 1 8 260 21 23 8 190
2 1 9 220 22 23 9 210
3 1 10 240 23 23 10 210
4 1 16 260 24 23 16 230
5 3 8 230 25 14 8 190
6 3 9 210 26 14 9 160
7 3 10 250 27 14 10 180
8 3 16 220 28 14 16 200
9 2 8 280 29 20 8 220
10 2 9 230 30 20 9 200
11 2 10 220 31 20 10 180
12 2 16 220 32 20 16 210
13 13 8 200 33 17 8 100
14 13 9 190 34 17 9 130
15 13 10 210 35 17 10 110
16 13 16 190 36 17 16 120
17 24 8 180 37 5 8 110
18 24 9 210 38 5 9 110
19 24 10 200 39 5 10 100
20 24 16 160 40 5 16 130

References

Bertsimas, D., Weismantel, R., 2005. Optimization over Integers. Dynamic Ideas.
Cairns, M.R., 1998. The development of park and ride in Scotland. J. Transp. Geogr. 6 (4), 295–307.
Chen, A., Xu, X., Ryu, S., Zhou, Z., 2013. A self-adaptive Armijo stepsize strategy with application to traffic assignment models and algorithms. Transp. A:
Transp. Sci. 9 (8), 695–712.
Chen, A., Zhou, Z., Xu, X., 2012. A self-adaptive gradient projection algorithm for the nonadditive traffic equilibrium problem. Comput. Oper. Res. 39 (2),
127–138.
Dijk, M., Montalvo, C., 2011. Policy frames of park-and-ride in Europe. J. Transp. Geogr. 19 (6), 1106–1119.
Du, B., Wang, D.Z.W., 2014. Continuum modeling of park-and-ride services considering travel time reliability and heterogeneous commuters - A linear
complementarity system approach. Transp. Res. Part E: Logist. Transp. Rev. 71 (0), 58–81.
Duncan, M., 2010. To park or to develop: trade-off in rail transit passenger demand. J. Plan. Educ. Res. 30 (2), 162–181.
Duncan, M., Christensen, R.K., 2013. An analysis of park-and-ride provision at light rail stations across the US. Transp. Policy 25 (1), 148–157.
Evans, S.P., 1976. Derivation and analysis of some models for combining trip distribution and assignment. Transp. Res. 10 (1), 37–57.
Facchinei, F., Pang, J.S., 2003. Finite-Dimensional Variational Inequalities and Complementarity Problems. Springer.
Fernandez, E., De Cea, J., Florian, M., Cabrera, E., 1994. Network equilibrium models with combined modes. Transp. Sci. 28 (3), 182–192.
García, R., Marín, A., 2005. Network equilibrium with combined modes: models and solution algorithms. Transp. Res. Part B: Methodol. 39 (3), 223–254.
Hamer, P., 2010. Analysing the effectiveness of park and ride as a generator of public transport mode shift. Road Transp. Res. 19 (1), 51–61.
Han, D., Sun, W., 2004. A new modified Goldstein–Levitin–Polyak projection method for variational inequality problems. Comput. Math. Appl. 47 (12),
1817–1825.
He, B.S., Yang, H., Meng, Q., Han, D.R., 2002. Modified Goldstein–Levitin–Polyak projection method for asymmetric strongly monotone variational inequali-
ties. J. Optim. Theory Appl. 112 (1), 129–143.
Kitthamkesorn, S., Chen, A., 2017. Alternate Weibit-based model for assessing green transport systems with combined mode and route travel choices. Transp.
Res. Part B: Methodol. 103, 291–310.
Kitthamkesorn, S., Chen, A., Xu, X., Ryu, S., 2016. Modeling mode and route similarities in network equilibrium problem with go-green modes. Netw. Spat.
Econ. 16 (1), 33–60.
Lam, W.H.K., Holyoak, N.M., Lo, H.P., 2001. How park-and-ride schemes can be successful in Eastern Asia. J. Urban Plan. Dev. 127 (2), 63–78.
Lam, W.H.K., Huang, H.J., 1992. A combined trip distribution and assignment model for multiple user classes. Transp. Res. Part B: Methodol. 26 (4), 275–287.
Lam, W.H.K., Li, Z., Wong, S.C., Zhu, D., 2007. Modeling an elastic-demand bimodal transport network with park-and-ride trips. Tsinghua Sci. Technol. 12
(2), 158–166.
Lee, J., Leyffer, S., 2012. Mixed Integer Nonlinear Programming. Springer-Verlag, New York.
Li, Z., Lam, W., Wong, S., Zhu, D.L., Huang, H.J., 2007. Modeling park-and-ride services in a multimodal transport network with elastic demand. Transp. Res.
Rec. J. Transp. Res. Board 1994, 101–109.
Liu, T.L., Huang, H.J., Yang, H., Zhang, X., 2009. Continuum modeling of park-and-ride services in a linear monocentric city with deterministic mode choice.
Transp. Res. Part B: Methodol. 43 (6), 692–707.
Liu, Z., Meng, Q., 2014. Bus-based park-and-ride system: a stochastic model on multimodal network with congestion pricing schemes. Int. J. Syst. Sci. 45
(5), 994–1006.
Meek, S., Ison, S., Enoch, M., 2008. Role of Bus-Based park and ride in the UK: a temporal and evaluative review. Transp. Rev. 28 (6), 781–803.
Nagurney, A., 1993. Network Economics: A Variational Inequality Approach. Springer, Netherlands.
Papola, A., 2004. Some developments on the cross-nested logit model. Transp. Res. Part B: Methodol. 38 (9), 833–851.
Parkhurst, G., 1995. Park and ride: Could it lead to an increase in car traffic? Transp. Policy 2 (1), 15–23.
Parkhurst, G., 20 0 0. Influence of bus-based park and ride facilities on users’ car traffic. Transp. Policy 7 (2), 159–172.
Parkhurst, G., Richardson, J., 2002. Modal integration of bus and car in UK local transport policy: the case for strategic environmental assessment. J. Transp.
Geogr. 10 (3), 195–206.
Patriksson, M., 2015. The Traffic Assignment problem: Models and Methods. Courier Dover Publications.
Pineda, C., Cortés, C.E., Jara-Moroni, P., Moreno, E., 2016. Integrated traffic-transit stochastic equilibrium model with park-and-ride facilities. Transp. Res.
Part C: Emerg. Technol. 71, 86–107.
Ryu, S., Chen, A., Choi, K., 2017. Solving the combined modal split and traffic assignment problem with two types of transit impedance function. Eur. J.
Oper. Res. 257 (3), 870–880.
Sheffi, Y., 1985. Urban Transportation networks: Equilibrium analysis with Mathematical Programming Methods. Prentice Hall.
Song, Z., He, Y., Zhang, L., 2017. Integrated planning of park-and-ride facilities and transit service. Transp. Res. Part C: Emerg. Technol. 74, 182–195.
62 Z. Liu et al. / Transportation Research Part B 117 (2018) 37–62

Spiess, H., Florian, M., 1989. Optimal strategies: A new assignment model for transit networks. Transp. Res. Part B: Methodol. 23 (2), 83–102.
Vovsha, P., 1997. Application of cross-nested logit model to mode choice in Tel Aviv, Israel, Metropolitan area. Transp. Res. Rec. J. Transp. Res. Board 1607,
6–15.
Wang, D., Du, B., 2013. Reliability-based modeling of park-and-ride service on linear travel corridor. Transp. Res. Rec. J. Transp. Res. Board 2333, 16–26.
Wang, H., Meng, Q., Zhang, X., 2014. Park-and-ride network equilibrium with heterogeneous commuters and parking space constraint. Transp. Res. Rec. J.
Transp. Res. Board 2466, 87–97.
Wang, J.Y.T., Yang, H., Lindsey, R., 2004. Locating and pricing park-and-ride facilities in a linear monocentric city with deterministic mode choice. Transp.
Res. Part B: Methodol. 38 (8), 709–731.
Wang, S., Meng, Q., Yang, H., 2013. Global optimization methods for the discrete network design problem. Transp. Res. Part B: Methodol. 50, 42–60.
Wiseman, N., Bonham, J., Mackintosh, M., Straschko, O., Xu, H., 2012. Park and ride: an Adelaide case study. Road Transp. Res.: A J. Aust. N. Z. Res. Pract. 21
(1), 39–52.
Yang, H., 1999. System optimum, stochastic user equilibrium, and optimal link tolls. Transp. Sci. 33 (4), 354–360.
Yang, H., Huang, H., 2005. Mathematical and Economic Theory of Road Pricing. Elsevier.
Zhou, Z., Chen, A., Wong, S.C., 2009. Alternative formulations of a combined trip generation, trip distribution, modal split, and trip assignment model. Eur.
J. Oper. Res. 198 (1), 129–138.