Professional Documents
Culture Documents
Abstract: In previous studies, almost all urban network design models/methods made attempts to optimize transportation system perfor-
mance, such as minimizing road traffic congestion, by determining optimal road capacity expansion schemes, with the assumption of inde-
structible road infrastructure. However, road infrastructure definitely deteriorates over time, and needs periodic maintenance and repairs. This
paper proposes a bilevel programming model for the urban freight transportation network design problem (UFTNDP) by taking into account
the heterogeneity of road classifications and road damages from the perspective of road maintenance. The model aims to improve network
performance as much as possible by determining an optimal joint scheme of road capacities and classifications. At the upper level, road
planning aims to minimize total system cost by determining a joint optimal scheme of road capacities and road classifications. The lower level
is a traffic assignment problem that characterizes network users’ path choice behavior. We then propose an improved system optimum (SO)
relaxation–based method to solve the optimization model. Numerical examples validated the developed model and tested the computational
efficiency of the proposed algorithm. Numerical results revealed that (1) the improved solution algorithm performs better, because it saves
97.3% of computational time compared with the original SO relaxation–based method; (2) the algorithm has the potential to be extended to
solving a general discrete network design problem; and (3) a simultaneous optimization of road classification and capacity works better in
improving system performance of an urban freight transportation network. DOI: 10.1061/JTEPBS.0000446. © 2020 American Society of
Civil Engineers.
Author keywords: Discrete network design problem; Bilevel programming model; Road classification; User equilibrium; Improved
system optimum (SO) relaxation–based algorithm.
Introduction studies have proposed NDPs to drive users’ route choices and
achieve better network performance.
Urban transportation plays an important role in promoting the eco- Although there are a vast number of NDP publications in the
nomic development of and social activities in cities. People often literature, almost all previous studies assumed a unified road qual-
regard road congestion as one of the most serious problems in ur- ity for all roads in the transportation network and overlooked the
ban transportation, and one that can be alleviated by building new heterogeneity of road damage maintenance costs. The differences
links and upgrading or enhancing existing roads. Such road capac- in road classifications and their actual conditions and qualities is
ity expansion is defined as a network design problem (NDP) in the apparent in practice. Bumpy Roads Ahead (2018) reported that
literature. It usually is formulated as a bilevel programming (BLP) about one-third of major urban roads in the US were poorly paved.
model, in which the upper-level planner aims to improve transpor- Road damage and maintenance costs are very important for road-
tation system performance by implementing optimal road capacity network planning and operation. For example, road damage creates
expansion schemes subject to investment constraints, and the lower- extra travel costs for road users. Martin and Thoresen (2015) re-
level users strive to optimize their routing decisions so as to minimize vealed that moving on damaged roads is estimated to cost US mo-
respective travel costs. According to different traffic assignment torists $80 billion in additional annual operating costs and vehicle
principles, lower-level problems can be differentiated further as user- repairs. Quite a few studies and reports also have pointed out that
equilibrium (UE) problems, stochastic user-equilibrium (SUE) prob- road maintenance costs cannot be negligible. Dueker and Fischer
lems, and so forth. Based on traffic assignment principles, many (2003) indicated that road investment was about 1.6% of GDP dur-
1
ing 1963–1984 and that road maintenance costs accounted for a
Lecturer, School of Economics and Management, Wuyi Univ., large part of road expenses during this period. Bumpy Roads Ahead
Guangdong 529020, China. Email: yeyipeng@foxmail.com
2 (2018) found that the backlog in America’s road and highway reha-
Senior Research Fellow, Dept. of Civil and Environmental Engi-
neering, National Univ. of Singapore, Singapore 117576, Singapore bilitation was $419.5 billion and that the annual investment in road
(corresponding author). ORCID: https://orcid.org/0000-0002-7150-5247. maintenance should be increased to $61 billion to improve the con-
Email: hwang191901@gmail.com dition of America’s roads and highways.
3
Professor, School of Economics and Management, Tongji Univ., Road damage is caused mainly by heavy goods vehicles (HGVs)
Shanghai 200092, China. Email: cexzhang@tongji.edu.cn and highly depends on their loads (Dodoo and Thorpe 2005). The
4
Associate Professor, College of Civil and Transportation Engineering, severity of road damage is related to vehicle loads and road clas-
Hohai Univ., Nanjing 210098, China. ORCID: https://orcid.org/0000-0001 sifications. The relationship between vehicle loads and road dam-
-9692-3771. Email: lirui2012@hhu.edu.cn
age has been investigated. Highway Research Board (1962) found
Note. This manuscript was submitted on April 3, 2020; approved on
June 8, 2020; published online on August 14, 2020. Discussion period open that road damage cost is proportional to the fourth power of the
until January 14, 2021; separate discussions must be submitted for indivi- total weight of the vehicle. However, road classification also is an
dual papers. This paper is part of the Journal of Transportation Engineer- important factor, because marginal road damage cost by equivalent
ing, Part A: Systems, © ASCE, ISSN 2473-2907. standard axles and fixed damage cost manifest very differently
cars and freight trucks coexist on the transportation network. This NDPs, such as the branch-and-bound algorithm and the cutting
investigated (1) how to formulate a DNDP that incorporates road plane algorithm. For example, Leblanc (1975) developed a bilevel
classification and road damage, and also takes into account the programming model for the DNDP and proposed a branch-and-
routing behavior of the two types of network users; and (2) how bound algorithm to solve it. Farvaresh and Sepehri (2013) inves-
to design a tailored solution algorithm so as to efficiently solve tigated a similar problem to Leblanc (1975) and proposed a tighter
the proposed DNDP model. lower bound (LB) to improve the solution algorithm. Wang et al.
(2013) proposed a system optimum (SO) relaxation–based method
to solve the DNDP, in which a feasible solution that satisfies the SO
Literature Review principle is regarded as a lower bound of UE solutions, and suc-
The urban transportation network design problem is concerned cessive cutting planes may be added to increase this lower bound.
mainly with expanding existing links or building new ones so The main merit of this method is that it ensures a globally optimal
as to optimize network performance while accounting for trav- solution.
elers’ route choice behavior. If the capacity of candidate links Another branch of exact solution algorithms seeks to reformu-
is regarded as a continuous variable, then the NDP can be catego- late the NDP as an equivalent mixed-integer linear program (MILP)
rized further as a continuous network design problem (CNDP) via linearization approximation. After that, the equivalent MILP
(e.g., Liu and Wang 2015). In turn, if discrete variables of link can be solved by state-of-the-art commercial solvers (e.g., CPLEX
capacities, e.g., lane numbers, are considered, the problem falls and GUROBI) or advanced solution techniques. For instance,
into the category of the DNDP (e.g., Wang et al. 2013). In addi- Wang and Lo (2010) developed a linearized formulation to approxi-
tion, the mixed network design problem (MNDP) (e.g., Luathep mate a CNDP that can be solved by CPLEX. This method is widely
et al. 2011) takes into account both continuous and discrete de- used in solving NDPs (Wang and Zhang 2017, Wang et al. 2014).
cision variables of link capacity expansion. Farahani et al. (2013) Luathep et al. (2011) extended this to investigate a MILP formu-
gave a comprehensive survey of the literature and detailed lation for the MNDP. Liu and Wang (2015) used a stochastic UE to
discussions. reformulate the NDP as a MILP and solved it by using a range
Numerous studies have investigated the network design problem reduction technique. To the best of our knowledge, finding a fast
under different assumptions. If the imperfect rationality of users is global optimization algorithm for the DNDP remains an open
considered, then the SUE principle is a reasonable extension of the problem.
UE principle and can be adopted to formulate users’ route choice
behavior. Based on the SUE principle, Long et al. (2010) proposed
Objectives and Contributions
a network design model and a branch-and-bound–based method to
tackle the turning restriction design problem. Considering the multi- The first objective of this work was to develop an optimization
dimensional measures of network design strategy (NDS), some re- model for the UFTNDP and thus find a network design scheme
cent studies proposed multiobjective optimization models for the with an optimal number of lanes and road classifications for each
network design problem. Szeto et al. (2010) simultaneously consid- candidate link. The second objective was to propose an improved
ered social, economic, and environmental impacts on humans and SO relaxation–based solution algorithm for the proposed UFTNDP
proposed a multiobjective optimization model for the road-network model.
design problem. To achieve these goals, we first formulated the UFTNDP as a
In addition to the elements considered in existing network de- bilevel programming model. A road planner at the upper level op-
sign studies, road damage also is an important problem facing timizes the urban freight transportation network capacity expansion
freight networks. Moreno-Quintero (2006) pointed out that in terms scheme in terms of road classification and the number of lanes for
of road-network quality, the major impact of freight transportation each candidate link, whereas freight and auto users at the lower
is not congestion but rather road damage. Road damage is highly level make their respective route choices by following the UE prin-
dependent on road classification and the weights of freight trucks. ciple. This model is computationally challenging, because the bi-
Martin and Thoresen (2015) pointed out that marginal road damage level programming problem is a nonconvex programming problem.
cost caused by equivalent standard axles and fixed damage cost are Because heuristics and metaheuristics cannot guarantee a global
very different for various road classifications. Volovski et al. (2017) solution, we preferred the exact algorithm, and improved the one
indicated that the annual maintenance expense costs of roads highly proposed by Wang et al. (2013) from the perspective of searching
depend on road pavement types. strategy. Wang et al. (2013)’s SO relaxation–based method always
Although existing studies have considered different assump- generates a global solution to the DNDP. However, it always
tions and proposed various network design models for urban searches several nonoptimal NDS and cuts off only one in each
transportation networks, almost all network design models iteration, so it needs many iterations and takes a long time to con-
implicitly suppose that a road can never perish and will serve verge. We improved the SO relaxation–based method through a
users forever. As argued previously, road damage should be solver technique called solver callback, which can cut off multiple
can be chosen for each candidate link. Constraint Eq. (3) ensures
first urban transportation network design model that simultane-
that exactly one capacity can be chosen for each candidate link.
ously addresses travel time cost, road damage cost, and road
Constraints Eqs. (4) and (5) ensure that link level parameters xa;m
classification.
and lane parameters ya;n are binary decision variables. Constraint
2. We propose an improved SO relaxation–based method to solve
Eq. (6) guarantees that the road construction cost does not exceed
the UFTNDP model, which worked very well in the tested in-
the total budget.
stances. Moreover, it can be easily extended to solve a generic
Because objective function Eq. (1) and constraint Eq. (6) are
DNDP.
nonlinear, we introduce auxiliary variable za;m;n ¼ xa;m ya;n and re-
3. We also contribute some managerial insights. First, we found
place constraint Eq. (6) with equivalent linear constraints
that Braess’s paradox applies in numerical examples. This indi-
cates that larger link capacity expansion is not always better and X X X
that policy makers should optimize NDS carefully. Second, in τ a;m;n za;m;n ≤ B ð7Þ
addition to link capacity expansion, a proper choice of road clas- a∈A2 m∈M a n∈N a
sification for candidate links can help improve further the per-
formance of an urban freight transportation network.
za;m·n ≤ xa;m ; ∀ a ∈ A2 ; m ∈ M a ; n ∈ N a ð8Þ
X X X
vFa;s − vFa;s ¼ qFi;s ; ∀ i ∈ N; s ∈ D ð16Þ vPa ¼ vPa;s ; ∀a∈A ð20Þ
a∈iþ a∈i−
s∈D
X
vFa ¼ vFa;s ; ∀a∈A ð17Þ
s∈D
vPa;s ≥ 0; ∀ a ∈ A; s ∈ D ð21Þ
Downloaded from ascelibrary.org by Universidad Militar Nueva Granada on 08/18/20. Copyright ASCE. For personal use only; all rights reserved.
where
8 P
>
> − DFrs i¼s
>
< r∈O
qFi;s ¼ F
Drs i¼r; ∀ i ∈ N; r ∈ O; s ∈ D
>
>
>
:0 otherwise
8 P
>
> − DPrs i¼s
>
< r∈O
qPi;s ¼ P
Drs i ¼ r; ∀ i ∈ N; r ∈ O; s ∈ D
>
>
>
:0 otherwise
Constraints Eqs. (7)–(11) ensure that the total construction cost s.t.
does not exceed the total budget. Objective function Eq. (13) X
represents the equilibrium of the freight trucks and private cars. vFa ¼ vFa;m ; ∀ a ∈ A2 ð25Þ
m∈Ma
Constraint Eq. (14) ensures aggregate flow conservation on each
link. Constraint Eq. (15) ensures that there is no flow on a candidate 0 ≤ vFa;m ≤ Uxa;m ; ∀ a ∈ A2 ; m ∈ M a ð26Þ
link if that link has not been constructed. Constraints Eqs. (16)–(18)
ensure freight vehicle flow conservation. Constraints Eqs. (19)–(21) va ¼ vFa þ vPa ; ∀a∈A ð27Þ
ensure private car flow conservation.
We reformulated both the upper- and lower-level programs as va ≤ Uð1 − ya;0 Þ; ∀ a ∈ A2 ð28Þ
equivalent mixed-integer convex programs (MICPs) to handle them
easily. Wang et al. (2013) used similar preprocessing. We intro- 0 ≤ va;m;n ≤ Uza;m;n ; ∀ a ∈ A2 ; n ∈ N a ð29Þ
duced new variables va;m;n , ∀a ∈ A2 ; m ∈ M a ; n ∈ N a , vFa;m , and X X
∀a ∈ A2 ; m ∈ M a ; va;m;n denotes vehicle flow on link a if link a va ¼ va;m;n ; ∀ a ∈ A2 ð30Þ
is level m with n lane(s); vFa;m denotes freight vehicle flow on link a m∈M a n∈N a
if link a is type m. The bilevel programming Eqs. (1)–(21) can be
reformulated as follows. with constraints Eqs: ð16Þ–ð21Þ ð31Þ
For the original problem (OP), the upper-level programming is
Define
X X X X
la ðcfixed þ cdam Ωx;y
a va Þ þ
F
min a la cfixed
a;m;n za;m;n
x;y
a∈A1
X X X
a∈A2 m∈Ma n∈N a ≔fðx; yÞ∶x and y satisfy constraints Eqs: ð2Þ − ð5Þ and ð7Þ − ð11Þg
þ la cdam F
a;m va;m þα va ta ðva Þ
a∈A2 m∈M a a∈A1 Ωv ðx; yÞ
X X X
þα va;m;n ta;m;n ðva;m;n Þ ð22Þ ≔ fv∶v satisfies constraints Eqs: ð25Þ − ð31Þ under NDSðx; yÞg
a∈A2 m∈M a n∈N a
The bilevel programming can be abbreviated as follows:
X X X X
s:t: Eqs. ð2Þ−ð5Þ and ð7Þ−ð11Þ ð23Þ min la ðcfixed
a þ cdam
a va Þ þ
F
la cfixed
a;m;n za;m;n
ðx;yÞ∈Ωx;y
a∈A1 a∈A2 m∈Ma n∈N a
The lower-level programming is X X X
þ a;m va;m þ α
la cdam F va ta ðva Þ
XZ X X XZ
a∈A2 m∈M a a∈A1
va va;m;n X X X
min ta ðθÞdθ þ ta;m;n ðθÞdθ ð24Þ þα va;m;n ta;m;n ðva;m;n Þ ð32Þ
v 0 0
a∈A1 a∈A2 m∈Ma n∈N a a∈A2 m∈M a n∈N a
5. for a ∈ A1 do
5. Exclude strategy ðx ; y Þ by adding lazy constraint to
a ðva Þ − T̄ a
T SO εSO
SO
½RP − copy 6. if va > v^ a and > then
6. else
SO
T̄ a 2
0
7. // Save incumbent solution for NDP 7. add lazy constraint T̄ SO a ≥ ½ta ðva Þ þ ta ðva Þva ðva −
8. UB ≔ SC va Þ þ ta ðva Þva to ½RP − copy
9. ðxopt ; y opt Þ ≔ ðx ; y Þ 8. end if
10. end if 9. end for
10. for a ∈ A2 do
11. for m ∈ Ma do
12. for n ∈ N a do
Dynamic OA Cutting Plane Method T SO ðv Þ− T̄ SOa;m;n ε
SO
13. if va;m;n > v^ a;m;n and a a;m;n >
Wang et al. (2013)’s method generates a set of linear constraints to SO
T̄ a;m;n 2
tighten the linear relaxation of the RP at each iteration. We refer to then
these as OA cutting planes. Our method considers generating OA 14. add lazy constraint
cutting planes at each searched solution to tighten the lower bound
of the initial problem, which can make use of the solver callback a;m;n ≥ ½ta;m;n ðva;m;n Þ
T̄ SO
function. The dynamic OA cutting plane method is implemented as 0
þ ta;m;n ðva;m;n Þva;m;n ðva;m;n
Algorithm 3.
− va;m;n Þ þ ta;m;n ðva;m;n Þva;m;n
to ½RP − copy
15. end if
16. end for
17. end for
18. end for
Numerical Tests
This section used the Sioux Falls network to illustrate the effective-
ness of the proposed model and solution method. This method uses
CPLEX version 12.8 to solve the linear relaxation of the RP with a
callback function. The relative optimality tolerance εSO was set at
0.001 for the dynamic OA method. The relative optimality toler-
ance ε for the bilevel programming also was set at 0.001. The tests
were performed on a PC with an Intel Core i5 CPU and 8G RAM.
Fig. 1 shows the network and new candidate links. Private car
demands and the data in the initial Sioux Falls network have the
same settings as those used by the Transportation Networks for Re-
search Core Team (2020), and we adjusted the unit of link data as
done by Luathep et al. (2011). We chose A2 ¼ fð13; 14Þ; ð14; 13Þ;
ð11; 15Þ; ð15; 11Þ; ð11; 9Þ; ð9; 11Þ; ð7; 16Þ; ð16; 7Þ; ð19; 22Þ; ð22; 19Þg
as the candidate link set. The value of time was set as $10/h. Freight
vehicle demands were set as 10% of private car demands. The
travel time function of links followed Bureau of Public Roads
(BPR) equation
ta ðva Þ ¼ t0a ½1 þ 0.15ðva =ca Þ4
Table 2. SO relaxation–based and improved SO relaxation–based methods when five candidate links are chosen
SO relaxation–based method Improved SO relaxation–based method
Average time to Average time to
CPU Relative generate one CPU Relative NDS cutting generate one
Cases time (s) Iterations gap (%) NDS cut (s) time (s) Iterations gap (%) planes NDS cut (s)
Case 1 >3,600 103 5.449 34.95 314.32 3 0.000 1386 0.23
Case 2 >3,600 228 2.382 15.79 99.36 3 0.000 473 0.21
Case 3 >3,600 259 3.498 13.90 136.09 2 0.000 858 0.16
Case 4 >3,600 169 3.493 21.30 193.43 3 0.000 1037 0.19
Case 5 >3,600 487 0.793 7.39 106.65 3 0.000 604 0.18
Case 6 >3,600 107 5.010 33.64 254.34 3 0.000 1352 0.19
Case 7 >3,600 333 2.735 10.81 51.63 2 0.000 679 0.08
Case 8 >3,600 185 2.598 19.46 69.47 3 0.000 642 0.11
Case 9 >3,600 220 2.454 16.36 74.47 3 0.000 592 0.13
Case 10 >3,600 180 3.666 20.00 84.14 2 0.000 679 0.12
This paper proposes a bilevel network design model for urban This study has been substantially supported by the National Natural
freight transportation networks to reduce the network’s system cost, Science Foundation Council of China through projects (Grant
which consists of road damage cost and travel time cost. Whereas Nos. 71601142, 71531011, and 71890970/71890973), and a project
the upper-level road planner chooses the optimal road classification sponsored by the program of Shanghai Academic Research Leader.
and capacity design to minimize the system cost, lower-level network
users make route choices in a UE manner. After that, we propose an
improved SO relaxation–based method that outperforms the existing Notation
SO relaxation–based method in solving the bilevel model. Numerical
tests indicated the feasibility of our model and method. The main The following symbols are used in this paper:
findings are as follows: A = A1 ∪ A2 ;
1. The optimal NDS will not construct all candidate links as ad- A1 = set of existing links in freight transportation network;
vanced types with maximum lanes without budget constraints. A2 = set of new candidate links that can be constructed;
Thus, they should be constructed with proper levels and lanes. B = total construction budget;
This instance of Braess’s paradox also urges development of an cdam
a = damage cost per unit distance if freight vehicle travels
efficient algorithm to solve the optimal NDS. on link a ∈ A1 ;
2. Link capacity design for candidate links can improve network cdam
a;m = Damage cost per unit distance if freight vehicle travels
performance. However, choosing proper types for candidate on link a ∈ A2 when link is type m ∈ M a ;
links can further improve the performance of the urban freight cfixed
a = fixed cost per unit distance of link a ∈ A1 ;
transportation network.
a;m;n = fixed cost per unit distance of link a ∈ A2 when link a is
cfixed
Based on our proposed model and method, we make the follow- level m with n lanes;
ing suggestions: D = destination set in freight transportation network;
1. This paper proposes a network design model that considers travel
DFw = demands of freight trucks for OD pair w ∈ W;
time cost, road damage cost, and road classification design.
DPw = demands of private cars for OD pair w ∈ W;
Future research can extend this idea in various freight network
environments (NDP for urban freight transportation with demand i− = fðk;iÞ ∈ A∶k ∈ Ng;
uncertainty, NDP for link design and maintenance planning, and iþ = fði;kÞ ∈ A∶k ∈ Ng;
so forth). la = length of link a ∈ A;
2. We suggest that government agencies consider not only road M a = road classification set of candidate link a ∈ A2 ;
capacity expansion, but also traffic components and road damage N = set of nodes in freight transportation network;
caused by freight trucks. If road classification is not considered, N a = set of numbers of lanes for candidate link a ∈ A2 ;