You are on page 1of 13

Optimization of a Transit Services Model with a Feeder Bus

and Rail System Using Metaheuristic Algorithms

Mohammad Hadi Almasi 1; Ali Sadollah 2; Sina Mirzapour Mounes 3; and Mohamed Rehan Karim 4
Downloaded from by Fundacion Universidad Del Notre on 10/16/18. Copyright ASCE. For personal use only; all rights reserved.

Abstract: Nowadays, many passengers use transit systems to reach their destinations; however, the growing concern for public transit is its
inability to shift passenger’s mode from private to public transportation. By designing a well-integrated public transit system and improving
the cost-effectiveness network, the public transport could play a crucial role in passenger satisfaction and reducing the operating cost. The
main target of this paper is to present a new mathematical programming model and design an efficient transit system to increase the efficiency
of integrated public transit services through the development of feeder bus services and coordination of major transportation services with
the aim of minimizing the costs. In this study, optimized transit services and coordinated schedules are developed using metaheuristic algo-
rithms such as genetic algorithm, particle swarm optimization, and imperialist competitive algorithm. The data used and the coordination
were obtained from a case study widely provided in the literature. Finally, obtained numerical results of the proposed model including
optimal solution, statistical optimization results, and the convergence rate, and comparisons are discussed in detail using tables and figures.
DOI: 10.1061/(ASCE)CP.1943-5487.0000418. © 2014 American Society of Civil Engineers.
Author keywords: Public transportation; Transit network design; Feeder bus; Intermodal coordination; Routing and scheduling;
Metaheuristic algorithms.

Introduction The main target of this paper is to present a new model and to
design an efficient transit system to increase the efficiency of feeder
Transportation is a multimodal, multiproblem, and multispectral network designs and coordinated schedules to minimize costs.
system, as it involves different categories and activities such as An improved integrated intermodal system may lead to a reduction
policy making, planning, designing, infrastructure construction, in total cost and an increase in profit, which consequently leads to
and development. Nowadays, considering the significant develop- achieving an optimum transit network design. Furthermore, such a
ments in technology, economy, and society, an efficient transpor- system can provide higher quality services for passengers.
tation system plays a key role in passengers’ satisfaction and the The structure of this paper is organized as follows: first, the rel-
reduction of costs. evant literature regarding the transit services and their applications
An intermodal transit system is a type of transportation, which are summarized. Then, a brief description, definition of the prob-
has proved challenging and controversial in the field of public lem, and details of the new mathematical model are presented, fol-
transportation. To improve complicated public transportation sys- lowed by succinct representations of applied metaheuristic methods
tems, a well-integrated transit system in urban areas can play a cru- to optimize the considered transit system problem. Finally, the
cial role in passengers’ satisfaction and reducing operating costs. computational optimization results obtained by optimizers accom-
This system usually consists of integrated rail lines and a number panied by their comparisons and discussions are put forth in, and
of feeder routes connecting transfer stations. The rail line that pro- the concluding results are presented.
vides an effective and convenient mode of transportation can carry
large numbers of travelers, while the feeder bus routes transport
passengers from bus stops to train stations (Kuan et al. 2006). Literature Review

A large and growing body of literatures on public transit system has

investigated integrated rail and feeder services.
Ph.D. Candidate, Center for Transportation Research (CTR), Faculty
of Engineering, Univ. of Malaya, 50603 Kuala Lumpur, Malaysia
(corresponding author). E-mail:
Classification Based on the Approach
Postdoctoral Research Fellow, School of Civil, Environmental and In general, previous approaches of transit network design problems
Architectural Engineering, Korea Univ., Seoul 136-713, South Korea. (TNDPs) could be divided into two major groups, namely, analytic
E-mail: and network approaches. These approaches differ in their purposes
Ph.D. Candidate, Center for Transportation Research (CTR), Faculty
and have different advantages and disadvantages. They should be
of Engineering, Univ. of Malaya, 50603 Kuala Lumpur, Malaysia. E-mail:
considered as complementary rather than alternative (Kuah 1986).
Professor, Center for Transportation Research (CTR), Faculty of Analytic Approaches
Engineering, Univ. of Malaya, 50603 Kuala Lumpur, Malaysia. E-mail:
Analytic models were developed to derive optimum relations
Note. This manuscript was submitted on August 20, 2013; approved on between different components of the feeder bus network process.
May 27, 2014; published online on July 22, 2014. Discussion period open This approach starts by formulating the design objective as a con-
until December 22, 2014; separate discussions must be submitted for in- tinuous function with a set of design variables. It is assumed that
dividual papers. This paper is part of the Journal of Computing in Civil the design variables are continuous, and the optimal values are
Engineering, © ASCE, ISSN 0887-3801/04014090(13)/$25.00. obtained using the optimal conditions according to the objective

© ASCE 04014090-1 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(6): 04014090

function. The typical design variables are feeder route location, rail A large number of studies have been published in utilizing
station spacing, and service frequencies. An analytic model needs heuristic methods due to their flexibility features (Reeves 1993).
a prespecified shape of the road geometry, and a well-designed Shrivastav and Dhingra (2001) developed a heuristic algorithm
demand function presenting the distribution of demand in the ser- to integrate the suburban stations and bus services, along with
vice area. Numerous studies have attempted to explain analytic optimization of coordinated schedules of feeder bus services using
models (Wirasinghe 1977; Wirasinghe et al. 1977; Wirasinghe existing schedules of suburban trains. They used Dijkstra’s algo-
1980; Kuah and Perl 1988; Chien and Yang 2000; Chien et al. rithm for generating initial routes. A good selection for generating
2001; Chowdhury et al. 2002; Chien and Schonfeld 1998). This initial routes is able to provide an efficient alternative at a reason-
approach can process only small-sized or regularly shaped net- able computation time by applying local search algorithms. Parallel
works. Therefore, the number of the possible solutions increases implementations including performance analysis of the two promi-
substantially with the increase in the number of roads in network. nent graph algorithms (i.e., Floyd–Warshall and Dijkstra) were
If road network is simple, the usage in the model will only be lim- utilized by Pradhan and Mahinthakumar (2013) in a large-scale
Downloaded from by Fundacion Universidad Del Notre on 10/16/18. Copyright ASCE. For personal use only; all rights reserved.

ited to theoretical applications and may not be applied to real-world transportation network.
situations (Kuan et al. 2004). Moreover, Chowdhury et al. (2002) proposed a model for better
coordination of the intermodal transit system. They applied a
Network Approaches numerical search algorithm (Powell’s algorithm) to solve their
Network approaches do not need prespecified shape of road geom- problem.
etry in the area. As a result, it is not limited to a simple network Steven and Chien (2005) suggested specific feeder bus service
structure, but it can be applied to more complicated networks. Kuah to provide a shuttle service between a recreation center and a major
and Perl (1989), as first developers of network approach, resolved the public transportation facility. They proposed an integrated method-
feeder bus network design and scheduling problem (FNDSP) by ology (analytical and numerical techniques) for development and
mathematical programming models. Some of the researchers optimization of decision variables.
adopted network approaches (Kuah and Perl 1989; Martins and
Pato 1998; Shrivastav and Dhingra 2001; Kuan et al. 2004, 2006; Metaheuristic Methods
Shrivastava and O’mahony 2007; Gholami and Mohaymany 2011). Transit route network design (TRND) problems are usually com-
Previous studies allocated discrete variables for demand and design plicated, and have three main challenges, in terms of managing
element in network approach, leading to its capability to deal with competing objectives, significant scalability, as well as significant
larger problem sizes and more realistic situations. combinatorial explosion when the network size grows (Baaj and
Mahmassani 1995). Because of this complexity in early efforts
Classification Based on Solution Methods of optimization through heuristics or analytical solutions, simpli-
fied versions of TRND problems were solvable. More recently,
To solve TNDPs, many studies have been carried out, which can be development of computing power offered using metaheuristic ap-
categorized into the following four groups: mathematical, heuristic, proaches, including genetic algorithms (GAs), simulated annealing
metaheuristic, and hybrid method. These groups are explained in (SA), and tabu search (TS). The TRND metaheuristics tend to
the following sections. follow one of the two different templates. The first approach begins
Mathematical Methods by generating a large set of possible routes and then iteratively
Several studies are performed on modeling feeder bus network selecting different subsets of the routes to create route networks.
design using mathematical approaches, which are based on The second approach begins by generating a potential route layout
mathematic concepts. A study by Wirasinghe et al. (1977) used and then one or more of the routes in the solution are changed in an
a mathematical method for designing a coordinated rail/bus transit attempt to find better solutions (Blum and Mathew 2011). Although
system that operated in peak hours between a metropolitan region metaheuristic methods are more time consuming than the early
and central business district. They obtained values of three inter- heuristics, they are capable of consistently producing high-quality
related parameters, called the station spacing, feeder bus zone solutions. In addition, metaheuristics, when compared with
boundary, and train headway to minimize the total cost. Further- classical heuristics, perform a much more thorough search of
more, Kuah and Perl (1988) presented an analytic model for the solution space, allowing inferior and sometimes infeasible
designing an optimal feeder bus network to access an existing rail moves, as well as recombination of solutions to create new ones
line. To decrease total cost, they applied mathematical method and (Cordeau et al. 2002).
avoided the simultaneous combining of stop spacing with route Kuan et al. (2004, 2006) applied GAs, ant colony optimization
spacing and operating headway variables. More rigorous problem (ACO), SA, and TS to resolve a feeder network design problem
statements and solid theoretical ground can usually be found in (FNDP) for a similar work conducted by Kuah and Perl (1989),
mathematical optimization approaches compared with other transit which improved previously proposed solutions. They generated
network design methods. However, such methods have two main several random tests to evaluate and compare the performance
disadvantages, being either a nonconvex problem or, in most cases, of their methods in terms of efficiency and accuracy of solutions.
a problem with unknown convexity. Another disadvantage stated Chien and Yang (2000) proposed an exhaustive search algo-
by Garey and Johnson (1979) is that the resultant mathematical rithm to optimize feeder bus route location and its operating head-
optimization systems derived from realistic combinatorial transit way in a given network. Moreover, in another study carried out
route network problems are usually at least nondeterministic poly- by Shrivastava and O’mahony (2006), optimum feeder routes and
nomial time hard (NP hard; Zhao and Zeng 2006). schedules of a suburban area were determined using the GAs.
The developed routes and schedules were optimized; however, it
Heuristic Methods failed to completely meet the demand. The reason was that some
To solve the problems at which classical methods are too slow, or of the nodes did not have good connection with other nodes in the
they fail to obtain any exact solution, heuristic methods are study area.
designed to accelerate the solving time of the problem, or to find Mohaymany and Gholami (2010) suggested an approach for
an approximate solution. solving multimodal FNDP with the objective of minimizing the

© ASCE 04014090-2 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(6): 04014090

total operator, user, and social costs. They used the ACO for can be determined by the network design problem. The route
constructing routes and modifying the optimization procedure to choice behavior of road users such that traffic flows satisfy user
identify the best mode and route in the service area. equilibrium conditions is influenced by this expansion decision
(Sharma et al. 2011).
Hybrid Methods The development of better integrated intermodal systems
Hybrid methods are categorized as another type of solution method improves service quality, which by extension increases passengers’
and combine the abilities of different computational techniques satisfaction as a corollary of better coverage, reduced access costs,
to solve complex problems. Recently, Shrivastava and O’mahony minimal delay, and shorter travel times. From the viewpoint of the
(2009) have developed the Shrivastava–O’mahony hybrid feeder transit operators, an overall coordination among the various public
route generation algorithm. The idea was to develop public bus transport modes can reduce their operating costs and increase their
routes and coordinate schedules in suburban area. In the proposed revenue by maintaining shorter routes and eliminating duplication
research, the GAs and the heuristic approach were combined to find of routes by the train and the buses.
Downloaded from by Fundacion Universidad Del Notre on 10/16/18. Copyright ASCE. For personal use only; all rights reserved.

optimized feeder routes, which have higher efficiency in compari- The focus of this study is designing a set of feeder bus routes
son with those developed by other researchers. and determining operating frequency on each route through min-
Many researchers have made an attempt to design a more effi- imizing the objective function including sum of operator, user,
cient feeder network and to provide feeder services connecting and social costs. The mathematical formulation of the improved
major transportation systems and welfare facilities. However, there model and its details for constraints are presented in the following
still are some limitations and gaps in previous studies. Table 1 sections.
illustrates a comparison of improved model in this study with some
of the previous models, and the gaps are expected to be filled by
the current improved model. Mathematical Formulation of the Model
The main objective of this research is to develop a mathematical
formulation model for designing and coordinating schedules of To formulate the cost function for the reported problem, total cost
integrated public transit services, which include development of function is expressed in the following equation. The total cost func-
feeder services and coordination with major transportation services tion is the sum of user, operator, and social costs that could be
and transfer time consideration between two modes. In the pro- formulated as follows:
posed improved model, the additional terms and constraints used CT ¼ Cu þ Co þ Cs ð1Þ
in objective function provide more accurate and efficient solutions
for various conditions of transit systems and this may lead to the where CT , Cu , Co , and Cs = total cost, user cost, operation cost,
creation of a more realistic model in simulating real-life problems. and social costs, respectively. Table 2 illustrates these costs more
Henceforth, each term of the improved model will be described
Problem Definition in detail in the following subsections. For nomenclature purposes,
all variables and parameters used in this paper for the modified
An integrated transit system including rail lines and a number of objective function are defined in the “Notation List” section.
feeder routes connected at different transfer stations is expected in
large metropolitan areas, where transit demand is high and widely
User Cost
needed. The problem involves designing a feeder network to pro-
vide access to an existing rail system and coordinate schedule of The user cost (Cu ) is the expense imposed on passengers using
transit service. the transit system (contains feeder and train services). This cost
Under a given budgetary constraint, the optimal capacity includes access, waiting, and in-vehicle traveling costs, which
improvements of existing road links of a transportation network are denoted by Ca, Cw , and Cui , respectively

Table 1. Comparison of Improved Model with Some of the Previous Models

User cost (Cu ) Operation cost (Co) cost (Cs )
Feeder Train Feeder Train Feeder
CuiF CoiF
References CaF CwF Crui Cdui CaT CwT CuiT CfF Croi Cdoi CmF CpF CfT CoiT CmT CpT CsF
p p p p p
Kuah and Perl (1989) — p p — — p p — p — — — — — — — —
Martins and Pato (1998) — p p — — p p — p — — — — — — — —
Kuan et al. (2004) — p p — — p p — p — — — — — — — —
Kuan et al. (2004) — p p — — p p — p — — — — — — — —
Kuan et al. (2006) — p — —
p — p — — — — — — — —
Shrivastava and O’mahony (2006) — — p — p — — — p — — — — — — — —
Shrivastava and O’mahony (2007) — — p — p — — — p — — — — — — — —
Shrivastava and O’mahony (2009) — —
p p — — — —
p p — —
p —
p — — — — —
Mohaymany and Gholami (2010) — p p — — — — p p — p p — — — — p
Gholami and Mohaymany (2011) —
p p p — —
p — — — p — — — —
Cipriani et al. (2012) p p p — p — — — — — p — — — — — —
Ciaffi et al. (2012) p p p —
p p —
p —
p —
p —
p —
p p —
p —
p —
p —
p —
p —
Proposed model

© ASCE 04014090-3 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(6): 04014090


Cu ¼ Ca þ Cw þ Cui ð2Þ

(Cs )c

In light of the user cost, which is the summation of feeder bus

and train cost, Eq. (2) can be rewritten as follows:

Cu ¼ ðCaF þ CaT Þ þ ðCwF þ CwT Þ þ ðCuiF þ CuiT Þ ð3Þ

In general, all elements of the user cost can be formulated as the
product of an hourly demand, average time spent in each travel time

category (i.e., access time, wait time, and in-vehicle time), and the
users’ value of time, all of which are explained in the following

Operating in-vehicle
Downloaded from by Fundacion Universidad Del Notre on 10/16/18. Copyright ASCE. For personal use only; all rights reserved.

Access Costs
The access cost (Ca ) is mainly incurred by feeder and train passen-


gers who have access to stops and stations. Thus, the access cost

can be formulated as follows:

Ca ¼ μa ðQK × taF þ QK × taT Þ ð4Þ
Operation cost (Co )b

The users’ value of time (μa ) is an important parameter in de-


termining the user cost, and is usually dependent on the economical

situation (e.g., annual income).

Waiting Cost
The waiting cost (Cw ) includes passengers waiting for the buses and
trains, which is the product of average wait time, demand, and the

value of users’ wait time (μw ). Average wait time can be estimated


by the fraction of the headway, which in this model is considered as

half the headway on average. Hence, the user waiting cost can be
represented using Eq. (5)


1 1
Operating in-vehicle

C w ¼ μw þ × QK ð5Þ
2Fk 2FT


User In-Vehicle Cost


Similarly, the user in-vehicle cost (Cui ) can be defined by the

product of demand, in-vehicle time, and value of time. The Cui
is formulated based on the average journey time and is calculated
for two main parts, namely, the run time and the dwell time. The

dwell time stands for the boarding and alighting time at the stops
(tdF ) and stations (tdT ). Therefore, the in-vehicle cost including
Running Dwell

in-bus and in-train cost, for each route k, is given as follows:

User in-vehicle

Cui ¼ Crui þ Cdui ð6Þ

Table 2. Illustration of Total Cost with All Terms in Proposed Model

Co ¼ ðCfF þ CfT Þ þ ðCoiF þ CoiT Þ þ ðCmF þ CmT Þ þ ðCpF þ CpT Þ.



|fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl}


 I  X  
1 X

Crui ¼ μI qi × Lijk þ tTj ð7Þ

V k i¼1 j¼1

Cu ¼ ðCaF þ CaT Þ þ ðCwF þ CwT Þ þ ðCuiF þ CuiT Þ.

|fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflffl}

Cdui ¼ μI ðn þ 1Þ × QK × tdF þ ðQK × tdT Þ ð8Þ

2 K
User cost (Cu )a


The first and second terms in Eqs. (7) and (8), respectively, de-
note the feeder bus and train user cost. In Eq. (6), Crui represents the
Running Dwell
User in-vehicle

run time for all passengers, which is equal to the link travel distance
from stop i to station j in route k (Lijk ) divided by average bus real



speed (V k ). The average cost of dwell time (Cdui ) is determined


by demand times, the passenger boarding, and alighting rate.


Moreover, tTj denotes riding time between station j and destination

of the train regardless of boarding and alighting times, and nk

stands for the number of stops in route k.


Because of the differences in spent time, which is dependent on

Cs ¼ CsF .

the dwell time at each bus stop, the geometric series equation has

been adopted to develop a more accurate model for distributing


dwell cost at the bus stops along the routes.


© ASCE 04014090-4 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(6): 04014090

Operation Cost Social Costs
The operating cost (Co ) is the summation of railway and feeder bus Social cost (Cs ) consists of many parameters that nonusers pay
operation costs. It can be described by the unit time or distance cost indirectly. For example, accident costs, pollution costs, infrastruc-
(such as hour or kilometer) in connection with the transit service ture costs, noise, and greenhouse gases. This cost is assumed to be
provided. Thus, Co can be formulated as follows: dependent on in-vehicle operating costs for feeder services and is
formulated as follows:
Co ¼ Coi þ Cm þ Cp þ Cf ð9Þ

These costs include the cost of trains and buses, and therefore, Cs ¼ λs ð2Fk × Lk Þ ð17Þ
it can be rewritten as follows:

Co ¼ ðCoiF þ CoiT Þ þ ðCmF þ CmT Þ þ ðCpF þ CpT Þ

Downloaded from by Fundacion Universidad Del Notre on 10/16/18. Copyright ASCE. For personal use only; all rights reserved.

Each cost term consists of several parameters and items, which

þ ðCfF þ CfT Þ ð10Þ consequently have differential influence in total cost, and determin-
Operating In-Vehicle Cost ing some of parameters and items needs cooperation of other sci-
Operating in-vehicle cost (Coi ) is dependent on the travel time and ences. Therefore, the interrelationship of some of cost terms with
round-trip distance. Coi (bus or train) is formulated based on the other related costs is considered. It is assumed that there is an
running cost (Croi ) and dwell cost (Cdoi ). The running cost for interrelationship between some of the costs such as social cost and
bus is formulated according to the round-trip distance against feeder operating in-vehicle cost (Mohaymany and Gholami 2010).
the rail, which is the round-trip time. It is assumed that the stop Thus, based on the previous studies, to moderately simplify the
delay time incurred at bus stops and also at intersections should proposed model, the social cost is assumed to be 20% of feeder
be taken into consideration. Thus, the Coi can be formulated as operating in-vehicle cost in this study.
given in Eqs. (11)–(13) in which the first terms in Eqs. (12)
and (13) correspond to the feeder bus cost, and the second terms Total Cost for a Route
denote the train cost
After calculating all cost components for route k, the total cost
Coi ¼ Croi þ Cdoi ð11Þ function CTk for route k is expressed as given in Eq. (18)

Croi ¼ λl ð2Fk × Lk Þ þ λIT ðFT × T T Þ ð12Þ   

1 1
CTK ¼ μa ðQK × taF þ QK × taT Þ þ μw þ × QK
Cdoi ¼ λI ðQK × tdF Þ þ λIT ðQK × tdT Þ ð13Þ 2Fk 2FT
 I  X 
1 X J
Indeed, railway operating cost is summarized in this subsection. þ μ1 qi × Lijk
V k i¼1
The term λIT represents all elements of operating cost including j¼1
fixed, maintenance, personnel, and in-vehicle costs ($=vehicle-h). 1
þ tTj ðnK þ 1Þ × QK × tdF þ ðQK × tdT Þ
Maintenance Cost 2
Maintenance cost (Cm ) consists of maintenance, repair, and tire þ λl ð2Fk × Lk Þ þ λIT ½ðQK × tdT Þ þ ðFT × T T Þ
costs. This cost depends on the fleet size and round-trip distance
þ λI ðQK × tdF Þ þ λm ð2Fk × LK Þ
formulated as follows:   
Cm ¼ λm ð2Fk × LK Þ ð14Þ þ λp × Lk þ ðQK × tdF Þ þ ðFk × Skj Þ
Personnel Cost 2Fk
Personnel cost (Cp ), which includes the drivers and administrative þ λf × LK þ λs ð2Fk × LK Þ ð18Þ
costs, is dependent on the fleet size, hourly pay, and insurance rate.
It can be formulated as follows:
Cp ¼ λp × LK þ ðQK × tdF Þ þ ðFk × Skj Þ ð15Þ Objective Function and Constraints

The first and second terms in Eq. (15) rely on the feeder running This transit network model must satisfy users, operators, and social
time and dwell time in route k, respectively. Accordingly, the third terms. Thus, the objective function is defined as sum of user,
term denotes the personnel cost while they are in the rest time or operator, and social costs, and is shown in Eq. (19)
queue. Hence, to increase the accuracy of cost function, adding
slack time (Skj ) into the schedule of bus route k at station j and K 
X User
average rest time are considered for each bus in stations. minimize CT ¼ ðCa þ Cw þ Cui Þ
Fixed Costs 
Operating Social
Fixed cost (Cf ) contains initial fleet costs such as vehicle owner- zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{ z}|{
ship costs, license, insurance, and so forth. It is formulated accord- þ ðCf þ Coi þ Cm þ Cp Þ þ Cs ð19Þ
ing to the fleet size and hourly fixed cost for the vehicle given for
route k
  Thus, according to the “Mathematical Formulation of the
Cf ¼ λf × Lk ð16Þ Model” section, objective function can be formulated after substi-
Vk tution of cost terms as follows:

© ASCE 04014090-5 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(6): 04014090

 X X X  X
1 1
MinimizeCT ¼ μa taF qi þ taTj qi × Y ij þ μw þ × QK
i¼1 j¼1 i¼1 k¼1
2Fk 2FT
K  I  X   
1 X J
þ μI qi × Lijk þ ðnK þ 1Þ × QK × tdF
V k i¼1 j¼1
XJ  X I    X K   X K 
þ qi × Y ij × ½tdT × ðJ − j þ 1Þ þ tTj  þ λf 2 × LK þ λl 2 Fk × LK
j¼1 i¼1
k¼1 k k¼1
XK  X I   X K 
þ λI QK × tdF þ λIT qi × tdT þ ðλIT × FT × T T Þ þ λm 2 Fk × LK
k¼1 i¼1 k¼1
Downloaded from by Fundacion Universidad Del Notre on 10/16/18. Copyright ASCE. For personal use only; all rights reserved.

K     X 
2Fk K
þ λp × LK þ ðQK þ tdF Þ þ ðFk × Skj Þ þ λs 2 Fk × LK ð20Þ
Vk k¼1

Subject to μw Qk
Fopt;K ¼
X H 4lk ½ðλl þ λm þ λs Þ þ V1k ðλf þ λp Þ þ ð2Skj × λp Þ
lih × X ihk ¼ 1 i ¼ 1; : : : ; I ð21Þ ð30Þ
k¼1 h¼1

In addition, the minimum required frequency of each route k is

H given as
lih × X ihk ≤ 1 k ¼ 1; : : : ; K ð22Þ
i¼1 h¼1
Freq;K ¼ ð31Þ
LF × C
Thus, the given frequency for each route k is obtained by select-
X ihk − X mik ≥ 0 i ¼ 1; : : : ; I k ¼ 1; : : : ; K ð23Þ ing the maximum value out of optimum frequency (Fopt;K ) and
h¼1 m¼1
required frequency (Freq;K ).
Some limitations are considered for the TNDP to represent an
K effective transit network model satisfying route feasibility, fre-
X ihk ≥ 1 ∀ H ð24Þ quency, and so forth. Eqs. (21)–(25) determine the route feasibility
i∈=H h∈H k¼1 in the network design. Eq. (21) explains that each bus stop should
be placed in a single route (many-to-one pattern). In addition,
I Eq. (22) ensures that each generated route must be connected to
X ihk þ X mik − Y ij ≤ 1 i ¼ 1; : : : ; I only one railway station. Accordingly, in Eq. (23), each bus is
h¼1 m¼1 assumed to pass at all the stops in its route node. Eq. (24) explains
j ¼ I þ 1; : : : ; I þ J k ¼ 1; : : : ; K ð25Þ that each feeder bus route should be linked to only one railway
station. Eq. (25) specifies that a bus stop can be assigned to a station
in which the corresponding route terminates at one of the rail sta-
lmin ≤ Lk ≤ lmax k ¼ 1; : : : ; K ð26Þ tions. Constraints on the minimum and maximum length of feeder
routes are given in Eq. (26). Similarly, the minimum and maximum
fmin ≤ Fk ≤ fmax k ¼ 1; : : : ; K ð27Þ frequencies are indicated in Eq. (27). Eq. (28) shows the maximum
number of vehicles in the fleet, and Eq. (29) represents the mini-
mum frequency for satisfying the demand.
× Lk þ ðQK × tdF Þ þ ðFk × Skj Þ ≤ N ð28Þ
Applied Optimization Methods

QK A lot of methods are used to solve TNDPs. Based on literature,

≤ Fk k ¼ 1; : : : ; K ð29Þ there are pros and cons for all these optimization methods. For ex-
ample, mathematical methods have some limitations such as being
where decision variables contain two binary variables, called Y ij either a nonconvex problem or a problem with unknown convexity.
and X ihk standing for definition of network and continues variable Moreover, some heuristic and analytical approaches have some
(Fk ), respectively. other limitations in the early efforts of optimization in solving com-
Determination of Fk , as one of the decision variables depends on plex problems. More recently, development of computing power
the transit network configuration. Thus, the optimal feeder bus fre- offered using metaheuristic approaches (Blum and Mathew 2011).
quency using analytical solution can be determined by setting the Some of the reasons for applying these three approaches in this
first derivative of the total cost function (CTk ) with respect to the paper are (1) to adopt some well-known and powerful methods
feeder bus frequency, equating it to zero and solving it. Therefore, for optimization of TNDPs [e.g., GA and particle swarm optimi-
the optimal bus frequency can be taken as zation (PSO)]; (2) the third method is imperialist competitive

© ASCE 04014090-6 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(6): 04014090

algorithm (ICA), a comparatively new metaheuristic method (first The PSO concept consists of, at each step, changing the velocity
proposed by Atashpaz-Gargari and Lucas in 2007) that is not (i.e., accelerating) of each particle toward its pBest and gBest
mostly utilized in transportation field so far. locations. Acceleration is weighted by a random term with separate
random numbers being generated for acceleration toward pBest
and gBest locations. The position and velocity of basic swarm
GA parameters are updated using the following equations (Kennedy
GAs are members of a collection of methodologies known as evo- and Eberhart 1995):
lutionary computation. These techniques are based on the princi-
ples of natural selection and evolution processes that are met in V iþ1 ¼ wV i þ c1 r1 ðpBesti − X i Þ þ c2 r2 ðgBesti − X i Þ ð32Þ
nature. The efficiency of the numerous evolutionary algorithms
in comparison with other heuristic techniques has been tested in
both generic (Youssef et al. 2001; Elbeltagi et al. 2005) and engi- X iþ1 ¼ X i þ V iþ1 ð33Þ
Downloaded from by Fundacion Universidad Del Notre on 10/16/18. Copyright ASCE. For personal use only; all rights reserved.

neering design (Giraud-Moreau and Lafon 2002) problems.

Through these tests, GAs are identified as robust heuristic tools where w = inertia weight for velocities (previously set between
capable of delivering efficient and robust solutions to diverse de- 0 and 1); X i = current value particle i; V i = updated velocity of
sign problems. In general, a GA begins its search with a population particle i; pBesti = best solution found by particle i; gBesti = best
of random individuals. Each member of the population possesses a solution found by the swarm; r1 and r2 = uniform random numbers
chromosome, which consists of genes. A gene may take on one of in the [0,1] range; c1 = cognitive component (self-confidence of
two allele values, either a 1 or a 0. All genes within a chromosome the particle); and c2 = social component (swarm confidence). The
are assembled as a binary string of ones and zeros, often with dis- terms c1 and c2 are constants that influence how each particle is
tinct segments. directed toward good positions taking into account personal best
The first step in creating an offspring population for the GA is to and global best information, respectively.
construct a mating pool. The mating pool contains N individuals, They usually are set as c1 ¼ c2 ¼ 1.5. The role of w is crucial
which are copied from the parent population that will be utilized to for the PSO convergence. It is used to control the impact of pre-
create the finished offspring population. To create the mating pool, vious velocities on the current particle velocity. A general rule of
a crowded tournament selection operator may be utilized. thumb indicates to set a large value initially to make the algorithm
To begin, two solutions from the parent population are selected explore the search space and then gradually reduce it to get refined
at random to compete in a tournament to construct the mating pool. solutions (Dong et al. 2005; Bergh and Engelbrecht 2006).
A copy of the winner of the tournament is kept in the mating
pool. This process is repeated until each solution in the parent
population has competed twice and the N spots in the mating pool ICA
have been filled. The resulting mating pool contains more copies of The ICA is inspired from the social–political process of imperial-
the more desirable solutions in the parent population and fewer ism and imperialistic competition. Similar to many optimization
copies of the less desirable solutions from the parent population. algorithms, the ICA starts with an initial population. Each individ-
With the mating pool complete, the task of constructing the off- ual of the population is called a country. Some of the best countries
spring population can commence. To generate two offspring solu- with the minimum cost are considered as the imperialist states and
tions, one begins by selecting two individuals from the mating pool the rest will be the colonies of those imperialist states. All the
at random. Once selected, the two individuals undergo crossover to colonies are distributed among the imperialist countries based on
create two offspring solutions, which are placed in the offspring their power.
population. To increase the diversity of the search, mutation oper- To define the algorithm, first of all, initial countries of size
ator is applied on randomly selected chromosome (Goldberg 1989). N Country are produced. Then, some of the best countries (with the
size of N imp ) in the population are selected to be the imperialist
states. Therefore, the rest with the size N col will form the colonies
PSO that belong to imperialists. Then, the colonies are divided among
PSO is an evolutionary computation technique for solving global imperialists according to their power (Atashpaz-Gargari and Lucas
optimization problems developed by Kennedy and Eberhart (1995). 2007). In such a way, the initial number of each empire’s colonies
It is a computation technique through individual improvement plus has to be proportional to its power. Hence, the initial number of
population cooperation and competition, which is based on the colonies of the nth empire will be (Khabbazi et al. 2009)
simulation of simplified social models such as bird flocking, fish ( )
schooling, and the swarm theory. Cost
Researchers found that the synchrony of animal’s behavior was NCn ¼ round PN imp × N col ; n ¼ 1,2; : : : ; N imp
through maintaining optimal distances between individual mem- i¼1 i

bers and their neighbors (Kennedy and Eberhart 1997). The PSO ð34Þ
algorithm exhibits common evolutionary computation attributes in-
cluding initialization with a population of random solutions and where NCn = initial number of colonies of the nth empire; N col =
searching for optima by updating generations. total number of initial colonies. To divide the colonies, NCn of the
Potential solutions, called birds or particles, are then flown colonies are randomly chosen and given to the nth imperialist
through the problem space by following the current optimum par- (Khabbazi et al. 2009).
ticles. Each particle keeps track of its coordinates in the problem After dividing all colonies among imperialists and creating the
space, which are associated with the best solution (fitness) it has initial empires, these colonies start moving toward their relevant
achieved so far. This value is called pBest. Another best value that imperialist country. This movement is a simple model of assimila-
is tracked by the global version of the PSO is the overall best value tion policy. In addition, the total power of an empire is defined by
and its location obtained so far by any particle in the population. the sum of the cost of the imperialist, and some percentage of the
This location is called gBest. mean cost of its colonies is given as follows (Khabbazi et al. 2009):

© ASCE 04014090-7 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(6): 04014090

TCn ¼ Costðimperialistn Þ þ ξfmean½Costðcolonies of empiren Þg

where TCn = total power of the nth empire; ξ = positive small

number. After computing the total power of empires, usually the
weakest colony (or colonies) of the weakest empire is chosen by
other empires and the competition is started on possessing this
colony. Each imperialist participating in this competition, accord-
ing to its power, has a probable chance of possessing the cited
To start the competition, at first, the weakest empire is chosen
and then the possession probability of each empire is estimated.
Downloaded from by Fundacion Universidad Del Notre on 10/16/18. Copyright ASCE. For personal use only; all rights reserved.

The possession probability Pp is related to the total power of the

empire (TC). To evaluate the normalized total cost of an empire
(NTC), the following equation is used (Khabbazi et al. 2009):

NTCn ¼ maxfTCi g − TCn n; i ¼ 1,2; 3; : : : ; N imp ð36Þ


During the imperialistic competition, the weak empires will

slowly lose their power and become weak as the time progresses.
At the end of a process, just one empire will remain that governs the
whole colonies (Khabbazi et al. 2009).

Computational Results and Discussion

As it is shown in Fig. 1 the improved model was applied to the
analysis of transit services including bus feeder services connecting
the rail stations in the case study by Kuah and Perl (1989).
The used network has 59 nodes, which include 55 bus stops
(1–55), and four rail stations (56–59) covering a service area of
3.22 × 4.02 km2 with an hourly demand density is approximately
200 passengers per stop. The values for the other parameters are Fig. 1. Schematic view for configuration of rail stations and stops for
given in Table 3. The dimensions and inputs of each node (Fig. 1) benchmark problem
are extracted from Kuah and Perl (1989). The transit network was
designed with feeder bus and fixed rail lines.
The TNDPs are categorized as the NP-hard problems with non-
linear objective function and constraints. Searching for the best
Less NFEs mean less time to reach the global optimum. This
feasible routes to minimize the cost function is a crucial mission
for solving the TNDP. Therefore, the optimization approaches, feature returns back to the structure of the algorithms.
which are mostly metaheuristics are of great importance. Best solution represents the accuracy of the method. The NFEs
The proposed improved model was carried out using the GA, and the best solution are dependent on each other. The ideal situa-
PSO, and ICA for a case study widely used in the literature (Kuah tion is the less NFEs and more accurate solution. To insure the fair-
and Perl 1989). The GA, PSO, and ICA have shown great poten- ness of comparison with other algorithms, a maximum 50,000 of
tials for solving optimization problems as they conducted a global NFEs [NFEs = number of iterations (generation) × number of
stochastic search (Giraud-Moreau and Lafon 2002; Dong et al. population] were imposed for all reported optimizers. The maxi-
2005; Atashpaz-Gargari and Lucas 2007). The applied methods mum number of iterations was 1,000 for the GA, PSO, and ICA.
were coded and implemented in MATLAB version 7.4 program- The detailed settings of applied approaches are tabulated in Table 4
ming software and run on Pentium IV 2.53-GHz CPU with for more clarification.
4-GB RAM. In terms of complexity level of the TNDP, the number of gen-
The task of optimizing the TNDP was executed in 15 indepen- erated routes is also considered as design variable, which vary from
dent runs for all the optimizers under consideration. The initial one individual to another in a generated population. Indeed, the
parameters of the GA included the population size of 50 individ- total number of design variables can vary in this problem. This
uals, scattered crossover fraction of 0.8, stochastic uniform as a means that the considered problem is categorized as a dynamic
selection function, and rank as a scaling function. optimization problem, which has various numbers of design vari-
Accordingly, for the PSO, the initial parameters consisted of the ables based on each individual.
population size of 50 individuals, the inertia weight for velocities Table 5 shows the comparison of the best solution attained for
of 0.8, and cognitive and social components (c1 and c2 ) of 1.5. all cost terms using applied optimization engines for the improved
In addition, the initial parameters for the ICA were chosen as num- model. The obtained total cost (CT ) is highlighted in bold in the
ber of country of 50, number of imperialist country of 4, revolution Table 5 for all algorithms. Similarly, Table 6 demonstrates the
rate of 0.3. comparison ways of obtaining statistical results for three reported
The number of function evaluations (NFEs) determines the optimizers for the TNDP. Table 7 presents the gained statistical op-
speed (computational effort) and the robustness of the algorithm. timization results for each cost term for the TNDP using the GA.

© ASCE 04014090-8 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(6): 04014090

Table 3. Selected Values for Parameters of Used Benchmark Problem Table 6. Comparison of Statistical Results Gained by Optimizers under
Parameter Unit Value
Best Mean Worst
μa $=Passenger-h 8
Optimizers solution solution solution SD
μw $=Passenger-h 8
μI $=Passenger-h 4 GA 31,994 32,644 33,353 323
λf $=Vehicle-h 14.37 PSO 31,293 31,725.6 32,021 222.98
λl $=Vehicle-km 0.357 ICA 29,549 29,964.13 30,590 250.84
λI $=Vehicle-h 11.44
λm $=Vehicle-km 0.75
λp $=Vehicle-h 10.2
λs $=Vehicle-km 0.07 Table 7. Gained Statistical Optimization Results for Each Cost Term for
V km=h 32 the TNDP Using the GA
Downloaded from by Fundacion Universidad Del Notre on 10/16/18. Copyright ASCE. For personal use only; all rights reserved.

Skj Min 15
taF Min 7.5 Parameters Best Mean Worst SD
taTj Min 4 CT 31,994 32,664 33,353 323
tdT Min=passenger 0.03 Cu 26,512 27,039 27,556 241.24
tdf Min=passenger 0.096 Co 5,318.5 5,495.6 5,678.3 93.69
VT km=h 40 Cw 5,782 5,921.3 6,060 77.611
FT Vehicle=h 20 Cui 3,854.6 4,309.8 4,688.2 211.44
f min Vehicle=h 2 Cf 661.3 705.57 762.2 27.24
f max Vehicle=h 20 Coi 727 762.28 807.34 21.67
N Vehicle=h 140 Cm 1,104.4 1,178.4 1,273 45.49
LF Passenger=seat 1 Cp 1,313 1,398 1,429 29,626
C Passenger=veh 50 Cs 103.1 110. 118.8 4.21
lmin km No constrain AF 11.33 11.84 12.38 0.27
lmax km 4 T PK 329,472 359,158 401,357 21,232
λlt $=Vehicle-h 180

From Table 6, the ICA has obtained the best statistical results
among others for the TNDP with the minimum cost of $29,549.
Table 4. Detailed Setting for Applying Proposed Methods The PSO and GA are placed in the second and third ranks, respec-
tively, whereas the PSO showed better solution stability in terms of
Method Detailed setting undergoing less standard deviation (SD) compared with reported
GA Number of independent run 15 algorithms.
Number of iterations 1,000 Tables 8–10 show the detailed statistical optimization results
Population size 50 for each term of modified cost function using applied algorithms.
Maximum NFEs 50,000 By observing Tables 8–10, it can be concluded that the ICA is
Scattered crossover 0.8
superior to other optimizers for finding all cost terms (except the
Mutations percentages 0.3
Number of parents for tournament selection 3 CP and AF ) having minimum statistical optimization results
PSO Number of independent run 15 (i.e., the best, the mean, and the worst solutions). It is worth men-
Number of iterations 1,000 tioning that the ICA has obtained the minimum cost for the CP ,
Population size 50 whereas the PSO has offered better mean, worst, and SD solutions
Maximum NFEs 50,000 (Tables 9 and 10). Regarding the low value of SD in the PSO algo-
Cognitive and social components (c1 and c2 ) 1.5 rithm, it may be due to selected values for the inertia factor and
Inertia weight(w) 0.8 acceleration coefficients (c1 and c2 ), which imply high tuning
ICA Number of independent run 15
dependency of PSO that can be counted as a drawback for this
Number of iterations 1,000 and
5,000 method. In other words, PSO in this context seems to be a more
Number of country(N Country ) 50 parameter-dependent method than ICA. Consequently, ICA has
Maximum NFEs 50,000 and provided better results than PSO.
250,000 Fig. 2 demonstrates the convergence rate and cost history (cost
Number of imperialist country(N ipm ) 4 reduction) among considered optimization engines. As it can be
Revolution rate 0.3 seen from Fig. 2, the ICA has faster and more accurately reached
Total number of initial colonies 46 its optimum solutions (solution quality) compared with the PSO
Number of independent run 15
and GA. The GAs cost reduction was not considerably high
Total number of initial colonies 46
compared with its early iterations.

Table 5. Comparison of the Best Obtained Solutions for the TNDP Using Reported Methods
Method Cw Cui Cf Coi Cm Cp Cu Co Cs CT AF T PK
GA 5,849.8 3,854.6 672.1 735.64 1,122.5 1,396 26,512 5,377 104.8 31,994 12.08 329,472
PSO 5,681.80 3,406.98 673.81 737.02 1,125.37 1,304.27 25,896.78 7,948.1 105.03 31,293 12.676 358,618
ICA 5,127.78 2,853.56 471.21 575.95 786.98 1,401.06 24,789.34 4,685.80 73.45 29,549 15.125 215,482
Note: Obtained total cost (CT ) is highlighted in bold.

© ASCE 04014090-9 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(6): 04014090

Table 8. Obtained Statistical Optimization Results for Each Cost Term for
the TNDP Using the PSO
Parameters Best Mean Worst SD
CT 31,293 31,725.6 32,021 222.98
Cu 25,896.8 26,209.2 26,551.3 218.03
Co 7,895.2 8,139.9 8,355 160.64
Cw 5,565.9 5,806 6,034 155.477
Cui 3,286.38 3,595.17 4,143.5 277.59
Cf 529.09 674.01 727.85 43.84
Coi 726.9 747.78 780.95 16.817
Cm 1,104.1 1,147.98 1,217.7 35.33
Cp 1,292.6 1,375.39 1,434.4 54.14
Downloaded from by Fundacion Universidad Del Notre on 10/16/18. Copyright ASCE. For personal use only; all rights reserved.

Cs 103.05 108.14 113.65 3.24

AF 11.47 12.27 13.07 0.535
T PK 306,112 347,107 386,765 25,371.5

Table 9. Attained Statistical Optimization Results for Each Cost Term for
the TNDP Using the ICA
Parameters Best Mean Worst SD
CT 29,549 29,964.1 30,590 250.84
Cu 24,789.3 25,128.2 25,523.4 215.37
Co 4,629.1 4,758.3 5,007.5 97.90
Cw 5,074.46 5,200.65 5,462.55 97.56
Cui 2,815.54 3,119.59 3,601.11 218.63
Cf 470.49 496.11 554.69 23.23
Coi 575.38 595.75 642.31 18.47
Cm 785.79 828.58 926.4 38.80
Cp 1,269.89 1,387.24 1,459.35 70.96
Cs 73.34 77.33 86.46 3.62
AF 13.5 14.72 15.28 0.455
T PK 196,456 227,059 253,375 17,648.2

Table 10. Comparison between Applied Metaheuristic and the Heuristic

Results Fig. 2. Comparison of convergence rate and cost history ($) with
respect to the number of iterations (generations) for the (a) GA;
Methods Total cost ($=h)
(b) PSO; (c) ICA
Heuristic 29,010
GA 31,994
PSO 31,293
ICAb 29,549 heuristic method used in literature by Kuah and Perl (1989) as a
ICAc 28,875 benchmark is shown in Table 10. It can be seen that ICA with
a 1,000 iterations produces results close to the best solution obtained
From Kuah and Perl (1989).
ICA with 1,000 iteration. by the heuristic method, and ICA with 5,000 iterations yields com-
ICA with 5,000 iteration. paratively better solution.
As discussed in the “Literature Review” section and summa-
rized in Table 1, there were some limitations and gaps in the
In addition, considering the convergence rates at approaching previous studies. This paper provided more accurate and effi-
1,000 iterations in Fig. 2, it can be seen that by increasing the num- cient solutions of various conditions of transit systems by using
ber of iterations of the ICA method, lower cost might be achieved. additional terms and constraints in objective function. In other
Therefore, ICA method was executed in 5,000 iterations, and words, the effort has been made to widen the scope of the
total cost of $28,875.18 was achieved (Fig. 3). Considering the research by considering all aspects of satisfaction (i.e., the user
presented objective function and the imposed constraints, the cost satisfaction, operation cost satisfaction, and social cost
achievement of these levels of total costs shows that the proposed satisfaction). The outputs of these solutions have demonstrated
model can be regarded as a potential method to overcome current that the presented model has been verified and the proposed ICA
difficulties in public transit system. For the two other optimization can be considered as a suitable approach to gain moderate-
algorithms, the use of maximum number of iteration exceeding quality solutions with reasonable computational cost. This model
1,000 did not improve the optimum solutions may lead to the creation of more realistic model in simulating
A comparison of the best results achieved by three metaheuristic real-life problems by providing fresh empirical data as a future
algorithms applied for optimizing the improved model and the work.

© ASCE 04014090-10 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(6): 04014090

Downloaded from by Fundacion Universidad Del Notre on 10/16/18. Copyright ASCE. For personal use only; all rights reserved.

Fig. 3. Convergence rate and cost history ($) with respect to the maximum iteration number of 5,000 for ICA

Conclusion CpT = train personals cost ($=h);

Croi = running operating in-vehicle cost ($=h);
In this paper, an improved model is suggested for the TNDP. Crui = running user in-vehicle cost ($=h);
The main purpose of this paper was to develop a real-life model Cs = social cost ($=h);
(actualizing the cost function and adding additional constraints) CT = total system cost ($=h);
for handling the TNDPs. The case study used in this paper was CTk = total cost function for route k ($=h);
extracted from the literature. Finding the optimum feasible routes Cu = user cost ($=h);
to reduce the cost function is a vital and difficult task for solving Cui = user in-vehicle cost ($=h);
the TNDP classified as NP-hard problem. For this reason, the im- CuiF = feeder user in-vehicle cost ($=h);
portance of optimization techniques, particularly metaheuristics is CuiT = train user in-vehicle cost ($=h);
understood. Therefore, three well-known optimization algorithms, Cw = waiting cost ($=h);
namely, GA, PSO, and ICA were used. The obtained statistical CwF = feeder waiting cost ($=h);
optimization results acquired by the ICA were superior to those CwT = train waiting cost ($=h);
attained by the PSO and GA. In terms of solution stability, the Fk = frequency of feeder bus on route k (vehicle=h);
PSO slightly outperformed the ICA. In terms of solution quality Fopt;k = optimum frequency of feeder bus on route (vehicle=h);
and statistical results, the GA ranked third. The obtained optimum Freq;k = required frequency of feeder bus on route k (vehicle=h);
number of routes using ICA was 20. Applying optimum network FT = frequency of train (vehicle=h);
resulted in the lowest level of total cost $28,875 by ICA, whereas f max = maximum frequency (vehicle=h);
the corresponding costs obtained by PSO and GA, are respectively fmin = minimum frequency (vehicle=h);
approximately 8.4 and 10.8% greater than that of ICA. H = all nodes containing stops and stations;
I = number of stops;
J = number of stations;
Notation k = number of routes;
K = set of routes belonging to the network of feeder;
The following symbols are used in this paper: Lijk = link travel distance from each stop i to station j in route
AF = average frequency of feeder buses system (vehicle-h); k (km);
C = capacity of feeder bus (passenger=vehicle); Lk = length of route k for the feeder bus (km);
CaF = feeder access cost ($=h); lih = distance from node i to h (km);
CaT = train access cost ($=h); lmax = the maximum length of one route (km);
Cdoi = dwell operating in-vehicle cost ($=h); lmin = the minimum length of one route (km);
Cdui = dwell user in-vehicle cost ($=h); LF = load factor of feeder bus (passenger=seat);
Cf = fixed costs ($=h); N = total fleet size of feeder bus (veh);
CfF = feeder fixed cost ($=h); nk = number of stops in route k;
CfT = train fixed cost ($=h); Qk = demand of route k (passenger=h);
Cm = maintenance cost ($=h); qi = demand of node i in mode (passenger=h);
CmF = feeder maintenance cost ($=h);
Skj = slack time route k at station j (h);
CmT = train maintenance cost ($=h);
Co = operation cost ($=h); T PK = total passenger-km (passenger-km);
Coi = operating in-vehicle cost ($=h); TT = train link travel time from 59 to 56 (h);
CoiF = feeder operating in-vehicle cost ($=h); taF = average access time to reach the feeder station (h);
CoiT = train operating in-vehicle cost ($=h); taTj = average access time to the rail station j (h);
Cp = personnel cost ($=h); tdf = dwell time for boarding and alighting to the feeder bus
CpF = feeder personals cost ($=h); (h=passenger);

© ASCE 04014090-11 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(6): 04014090

tdT = dwell time for boarding and alighting to the train Elbeltagi, E., Hegazy, T., and Grierson, D. (2005). “Comparison among
(h=passenger); five evolutionary-based optimization algorithms.” Adv. Eng. Inf., 19(1),
tih = linked in-vehicle time between nodes i and h of feeder 43–53.
bus (h); Garey, M. R., and Johnson, D. S. (1979). Computers and intractability:
tTj = linked riding time between station j and destination of the A guide to the theory of NP-completeness, Freeman, New York.
Gholami, A., and Mohaymany, A. S. (2011). “Economic conditions
train (h);
for minibus usage in a multimodal feeder network.” Transp. Plann.
V = average operating speed of feeder bus (km=h); Technol., 34(8), 839–856.
V T = average operating speed of train (km=h); Giraud-Moreau, L., and Lafon, P. (2002). “Comparison of evolutionary
X ihk = binary variable; value of 1 if stop i precedes stop h on bus algorithms for mechanical design components.” Eng. Optim., 34(3),
route k; 307–322.
Y ij = binary variable; value of 1 if stop i is assigned to station j; Goldberg, D. (1989). Genetic algorithms in optimization, search and
λf = fixed cost of feeder bus ($=vehicle-h); machine learning, Addison Wesley, New York.
Downloaded from by Fundacion Universidad Del Notre on 10/16/18. Copyright ASCE. For personal use only; all rights reserved.

λI = vehicle operating cost of feeder bus ($=vehicle-h); Kennedy, J., and Eberhart, R. (1995). “Particle swarm optimization.”
λl = vehicle operating cost of feeder bus ($=vehicle-km); Proc., Neural Networks, 1995. IEEE Int. Conf. on Neural Networks,
λlT = vehicle operating cost of train ($=vehicle-h); Vol. 4, IEEE, Piscataway, NJ, 1942–1948.
λm = maintenance cost of feeder bus ($=vehicle-km); Kennedy, J., and Eberhart, R. (1997). “A discrete binary version of the
λp = personals cost of feeder bus ($=vehicle-h); particle swarm algorithm.” IEEE Sys Man Cybern, Computational
λs = social cost of feeder bus ($=vehicle-km); Cybernetics and Simulation, Vol. 5, Orlando, FL, 4104–4108.
μa = passenger access cost ($=passenger-h); Khabbazi, A., Atashpaz-Gargari, E., and Lucas, C. (2009). “Imperialist
competitive algorithm for minimum bit error rate beam forming.”
μI = passenger riding cost on transit mode
Int. J. Bio-Inspired Comput., 1(1/2), 125–133.
($=passenger-h); and
Kuah, G. K. (1986). “Feeder bus route design problem.” Ph.D. thesis,
μw = passenger waiting cost for arrival of transit mode Maryland Univ., College Park, MD.
($=passenger-h). Kuah, G. K., and Perl, J. (1988). “Optimization of feeder bus routes and bus
stop spacing.” J. Transp. Eng., 10.1061/(ASCE)0733-947X(1988)
114:3(341), 341–354.
Kuah, G. K., and Perl, J. (1989). “The feeder-bus network-design problem.”
J. Oper. Res. Soc., 40(8), 751–767.
Kuan, S., Ong, H., and Ng, K. (2004). “Applying metaheuristics to feeder
The authors would like to acknowledge the Ministry of Higher
bus network design problem.” Asia-Pac. J. Oper. Res., 21(4), 543–560.
Education of Malaysia and the University of Malaya for the
Kuan, S., Ong, H., and Ng, K. (2006). “Solving the feeder bus network
financial support under flagship FL020/2012. design problem by genetic algorithms and ant colony optimization.”
Adv. Eng. Software, 37(6), 351–359.
Martins, C., and Pato, M. (1998). “Search strategies for the feeder bus
References network design problem.” Eur. J. Oper. Res., 106(2–3), 425–440.
MATLAB version 7.4 [Computer software]. Mathwork, Natick, MA.
Atashpaz-Gargari, E., and Lucas, C. (2007). “Imperialist competitive Mohaymany, A. S., and Gholami, A. (2010). “Multimodal feeder network
algorithm: An algorithm for optimization inspires by imperialistic design problem: Ant colony optimization approach.” J. Transp. Eng.,
competition.” IEEE Congress on Evolutionary Computation, Singapore, 10.1061/(ASCE)TE.1943-5436.0000110, 323–331.
4661–4667. Pradhan, A., and Mahinthakumar, G. (2013). “Finding all-pairs shortest
Baaj, M. H., and Mahmassani, H. S. (1995). “Hybrid route generation path for a large-scale transportation network using parallel Floyd-
heuristic algorithm for the design of transit networks.” Transp. Res., Warshall and parallel Dijkstra algorithms.” J. Comput. Civ. Eng.,
Part C Emerging Technol., 3(1), 31–50. 10.1061/(ASCE)CP.1943-5487.0000220, 263–273.
Bergh, F., and Engelbrecht, A. P. (2006). “A study of particle swarm Reeves, C. R. (1993). Modern heuristic techniques for combinatorial
optimization particle trajectories.” Inf. Sci., 176(8), 937–971.
problems, Wiley, New York.
Blum, J. J., and Mathew, T. V. (2011). “Intelligent agent optimization of
Sharma, S., Mathew, T. V., and Ukkusuri, S. V. (2011). “Approximation
urban bus transit system design.” J. Comput. Civ. Eng., 10.1061/
techniques for transportation network design problem under demand
(ASCE)CP.1943-5487.0000095, 357–369.
uncertainty.” J. Comput. Civ. Eng., 10.1061/(ASCE)CP.1943-5487
Chien, S., and Schonfeld, P. (1998). “Joint optimization of a rail transit line
.0000091, 316–329.
and its feeder bus system.” J. Adv. Transp., 32(3), 253–284.
Shrivastav, P., and Dhingra, S. (2001). “Development of feeder routes for
Chien, S., and Yang, Z. (2000). “Optimal feeder bus routes on irregular
street networks.” J. Adv. Transp., 34(2), 213–248. suburban railway stations using heuristic approach.” J. Transp. Eng.,
Chien, S., Yang, Z., and Hou, E. (2001). “Genetic algorithm approach for 10.1061/(ASCE)0733-947X(2001)127:4(334), 334–341.
transit route planning and design.” J. Transp. Eng., 10.1061/(ASCE) Shrivastava, P., and O’mahony, M. (2006). “A model for development
0733-947X(2001)127:3(200), 200–207. of optimized feeder routes and coordinated schedules: A genetic
Chowdhury, S. M., Steven, I., and Chien, J. (2002). “Intermodal transit algorithms approach.” Transp. Policy, 13(5), 413–425.
system coordination.” Transp. Plann. Technol., 25(4), 257–287. Shrivastava, P., O’mahony, M. (2007). “Design of feeder route network
Ciaffi, F., Cipriani, E., and Petrelli, M. (2012). “Feeder bus network design using combined genetic algorithm and specialized repair heuristic.”
problem: A new metaheuristic procedure and real size applications.” J. Public Transp., 10(2), 99–123.
Procedia Soc. Behav. Sci., 54, 798–807. Shrivastava, P., and O’mahony, M. (2009). “Use of a hybrid algorithm for
Cipriani, E., Gori, S., and Petrelli, M. (2012). “Transit network design: modeling coordinated feeder bus route network at suburban railway
A procedure and an application to a large urban area.” Transp. Res., stations.” J. Transp. Eng., 10.1061/(ASCE)0733-947X(2009)135:1(1),
Part C Emerging Technol., 20(1), 3–14. 1–8.
Cordeau, J. F., Gendreau, M., Laporte, G., Potvin, J. Y., and Semet, F. Steven, I., and Chien, J. (2005). “Optimization of headway, vehicle size
(2002). “A guide to vehicle routing heuristics.” J. Oper. Res. Soc., and route choice for minimum cost feeder service.” Transp. Plann.
53(5), 512–522. Technol., 28(5), 359–380.
Dong, Y., Tang, J., Xu, B., and Wang, D. (2005). “An application of Wirasinghe, S. (1977). “Assignment of buses in a coordinated rail and bus
swarm optimization to nonlinear programming.” Comput. Math. Appl., transit system.” Proc., 7th Int. Symp. on Transportation and Traffic
49(11-12), 1655–1668. Flow Theory, Kyoto, Japan, 673–96.

© ASCE 04014090-12 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(6): 04014090

Wirasinghe, S. (1980). “Nearly optimal parameters for a rail/feeder-bus Youssef, H., Sait, S. M., and Adiche, H. (2001). “Evolutionary algorithms,
system on a rectangular grid.” Transp. Res., Part A Gen., 14(1), simulated annealing and tabu search: A comparative study.” Eng. Appl.
33–40. Artif. Intell., 14(2), 167–181.
Wirasinghe, S. C., Hurdle, V. F., and Newell, G. F. (1977). “Optimal param- Zhao, F., and Zeng, X. (2006). “Simulated annealing–genetic algorithm for
eters for a coordinated rail and bus transit system.” Transp. Sci., 11(4), transit network optimization.” J. Comput. Civ. Eng., 10.1061/(ASCE)
359–374. 0887-3801(2006)20:1(57), 57–68.
Downloaded from by Fundacion Universidad Del Notre on 10/16/18. Copyright ASCE. For personal use only; all rights reserved.

© ASCE 04014090-13 J. Comput. Civ. Eng.

J. Comput. Civ. Eng., 2015, 29(6): 04014090