Professional Documents
Culture Documents
Transportation Science
Publication details, including instructions for authors and subscription information:
http://pubsonline.informs.org
This article may be used only for the purposes of research, teaching, and/or private study. Commercial use
or systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisher
approval, unless otherwise noted. For more information, contact permissions@informs.org.
The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitness
for a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, or
inclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, or
support of claims made of that product, publication, or service.
INFORMS is the largest professional society in the world for professionals in the fields of operations research, management
science, and analytics.
For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org
informs ®
Vol. 43, No. 2, May 2009, pp. 129–143
issn 0041-1655 eissn 1526-5447 09 4302 0129 doi 10.1287/trsc.1080.0250
© 2009 INFORMS
Tadashi Yamada
Department of Urban Management, Kyoto University, Nishikyo, Kyoto 615-8540, Japan,
t.yamada@kiban.kuciv.kyoto-u.ac.jp
Eiichi Taniguchi
Department of Urban Management, Kyoto University, Nishikyo, Kyoto 615-8540, Japan,
taniguchi@kiban.kuciv.kyoto-u.ac.jp
D esigning multimodal freight transport networks can facilitate the economic development of regions and
countries as well as help to reduce negative environmental impacts. It is therefore crucial that such be
undertaken in areas where more priority is given on road-based freight transport systems. This paper proposes
a model for strategic transport planning, particularly in freight terminal development and interregional freight
transport network design. The model determines a suitable set of actions from a number of possible actions, such
as improving the existing infrastructure or establishing new roads, railways, sea links, and freight terminals.
Modelling is undertaken within the framework of bilevel programming, where a multimodal multiclass user
traffic assignment technique is incorporated within the lower-level problem, whilst the upper-level problem
determines the best combination of actions such that the freight-related benefit-cost ratio is maximised. The
upper-level problem involves combinatorial optimisation, and a heuristic approach based on genetic local search
is applied as a solution technique. Empirical results of the model as applied to an actual large-sized inter-
regional intermodal freight transport network show that genetic local search could provide better performance
as compared to other genetic algorithm-based, as well as tabu search-based, heuristics. The model is successfully
applied to transport network planning in the Philippines, where the development of a freight transport network
is necessary to increase the utilisation of other modes rather than road-based vehicles.
Key words: multimodal freight transport; discrete network design; genetic local search
History: Received: February 2007; revisions received: October 2007, April 2008; accepted: October 2008.
Published online in Articles in Advance May 7, 2009.
roads, railways, sea links, and freight terminals. The traffic assignment technique (e.g., Dafermos 1980;
most feasible set of infrastructure projects is selected Crainic, Florian, and Leal 1990; Guelat, Florian, and
for efficient design of the multimodal freight trans- Crainic 1990; Cascetta 2001; Nagurney and Dong
port network. The actions are represented by adding 2001). This allows the model to deal with more real-
new links or selecting existing links for improvement istic situations. Therefore, the model has the potential
in the transport network. The problem can therefore to be used as a tool for strategic level of multimodal
be considered as a discrete network design problem freight transport planning, particularly in freight ter-
Downloaded from informs.org by [190.144.171.70] on 07 August 2016, at 20:11 . For personal use only, all rights reserved.
(DNDP), which basically implies the selection of link minal development and freight transport network
additions to an existing transport network with given design. To our knowledge, there has been no research
demand from each origin to each destination. The that uses such traffic assignment techniques in DNDP,
DNDP is described in detail in Bruynooghe (1972); with the performance of the model being investi-
Steenbrink (1974); Poorzahedy and Turnquist (1982); gated with actual large-sized interregional intermodal
Boyce (1984); Magnanti and Wong (1984); Chen and freight transport networks. The application of the
Alfa (1991); Yang and Bell (1998); and Gao, Wu, and model to such transport networks is the highlight of
Sun (2005). the paper.
As indicated in Ravi and Sinha (2006), there has The optimisation process in determining the set
been very little effort to study the problem in an of network improvement actions is described within
integrated manner that would allow one to exploit the framework of bilevel programming, where the
the savings that may result from making both deci- multimodal multiclass user equilibrium traffic assign-
sions in a coordinated way to reduce the total cost of ment technique is incorporated within the lower level,
location and transportation. In this context, Arnold, whilst the combination of actions is approximately
Peeters, and Thomas (2004) deals with the prob- optimised using metaheuristics-based procedures in
lem of optimally locating rail/road terminals for the upper level. This type of problem also involves
freight transport and relates closely to our problem. a mathematical problem with equilibrium constraints
Their approach is based on fixed-charge network (MPEC) (Luo, Pang, and Ralph 1996). The advan-
design problems (e.g., Magnanti and Wong 1984; tage of adopting metaheuristic techniques is that such
Balakrishnan, Magnanti, and Wong 1989), which are techniques can handle complex problems and provide
also applied by Melkote and Daskin (2001) for simul- the flexibility of the design of bilevel programming
taneously optimising facility locations and the design problems if applied as optimisation techniques. These
of the underlying transport network. As such, net- techniques can also compute approximately efficient
work location theory, including the fixed-charge net- solutions in relatively shorter times. Therefore, there
work design problems, location-routing problems have been an increasing number of researches incor-
(e.g., Laporte 1988; ReVelle and Laporte 1996; Min, porating metaheuristic techniques into such problems
Jayaraman, and Srivastava 1998), and hub location (e.g., Bard 1998, Taniguchi et al. 1999, Shepherd and
problems (e.g., O’Kelly 1987; Aykin 1990; Campbell Sumalee 2004, Zhang and Yang 2004).
1994, 1996) has been a powerful tool for transport net- This sort of bilevel programming is similar to that
work design and facility locations. Francis, McGinnis, used in Taniguchi et al. (1999). However, the model
and White (1992), Daskin (1995), and Drezner (1995) presented in this paper can handle both the improve-
provide a comprehensive overview of network loca- ment of link capacity and the addition of new links
tion models. to an existing multimodal transport network and can
Many useful findings have been provided by these deal with a variety of actions relating to the develop-
conventional network location models in efficiently ment of freight transport network, whereas the model
obtaining exact/approximate optimal solutions and described in Taniguchi et al. (1999) can only deal with
clarifying the basic relationships between the optimal the choice of nodes for pickup/delivery and line-haul
location and transport cost. However, such models trucks to determine the location and size of freight ter-
have not appropriately taken into account the change minals taking the behaviour patterns of motorists into
in traffic flow caused by expanding and improving consideration. This difference depends partly on the
the transport network. Traffic conditions on the trans- use of multimodal multiclass user equilibrium traf-
port network should be influenced by the network fic assignment technique to be formulated with vari-
improvement actions implemented, especially when ational inequality, which is not applied in Taniguchi
such actions are carried out on a larger scale as trans- et al. (1999). Furthermore, this paper utilises more
port initiatives in interregional or urban transport detailed representation of multimodal transport net-
planning. works and deliberates on the choice of metaheuristic
The model presented in this paper explicitly incor- techniques to be incorporated within the model.
porates traffic and freight flows on the transport net- The rest of the paper is organised as follows. In the
work using a multimodal multiclass user equilibrium following section, the overall modelling framework is
Yamada et al.: Designing Multimodal Freight Transport Networks
Transportation Science 43(2), pp. 129–143, © 2009 INFORMS 131
described, and the formulation of the model is given. Tobin, and Harker 1983; Crainic, Florian, and Leal
In §3, the performance of several heuristic procedures 1990; Guelat, Florian, and Crainic 1990; Tavasszy 1996;
to be incorporated within the upper level is investi- Southworth and Peterson 2000). The reason why such
gated on a test network with the objective of applying an aggregate approach was used is that the national
the model to an actual large-sized interregional inter- origin-destination and transport facility surveys in
modal freight transport network. The model is then the Philippines are almost viable for interregional
tested and applied in §4 to an actual transport net- freight transport modelling. These only offer aggre-
Downloaded from informs.org by [190.144.171.70] on 07 August 2016, at 20:11 . For personal use only, all rights reserved.
work in the Philippines, and explores the implications gate origin-destination information that identifies the
of how to expand the interregional freight transport mode shares of the total commodity but cannot pro-
network. Finally, in §5, the methodologies, results, vide information on the activities of individual ship-
and analyses in the paper are summarised. pers as well as those of carriers. Thus, neither the influ-
ence of shipper-carrier behaviour nor their interaction
2. The Model in the freight transport decision is explicitly taken into
Modelling is undertaken using bilevel programming. account in this model.
The lower level describes the multimodal multiuser The model allows freight and passengers to be
equilibrium flow on the transport network, whereas treated as multiclass users, with modal split and route
the upper level determines the best combination of choice carried out simultaneously by converting the
network improvement actions. Overall, the model can multimodal network into a unimodal abstract mode
be represented as a leader-follower game (Stackelberg network. A detailed representation of freight move-
1934), where transport planning departments are the ment within terminals is developed based on network
leaders, and the followers are the transport network descriptions proposed by Guelat, Florian, and Crainic
users (i.e., freight and passengers). This model can (1990); Tavasszy (1996); and Southworth and Peterson
simply be formulated as follows (where · · denotes (2000), incorporating loading/unloading, storage and
the inner product): administrative processes (see Figure 1).
max f y b x∗ x0∗ (1) The lower-level problem involves user equilib-
y rium conditions with a nonseparable and asymmetric
s.t. cx∗ x − x∗ ≥ 0 ∀ x ∈ K (2) Jacobian matrix cost function among user types
that can be stated using variational inequality (e.g.,
cx0∗ x0 − x0∗ ≥ 0 ∀ x0 ∈ K 0 (3) Dafermos 1980, Nagurney 2000). This can be formu-
where, lated as follows:
∗
Find xai ∈ such that:
y: set of freight network improvement actions,
b: vector of investment/operation costs, p
∗
x: vector of link flows, cai x̃∗ × xai − xai ≥ 0 ∀ x̃ ∈ (4)
x∗ :vector of equilibrium link flows, i=1 a∈A
assumed to have a fixed value and does not depend ta : free travel time on link a,
on volume, the time cost component, particularly the r0a : existing link capacity on link a,
time spent on the link, is a function of volume and ra : additional link capacity on link a if the
differs by link type. The time spent on the link is rep- action is implemented,
resented by a continuous function in the form of poly- 1 , 2 , : parameters to be calibrated.
nomial approximation presented by Crainic, Florian, The objective function shown in Equation (5) rep-
and Leal (1990) to keep the link cost function mono- resents the BCR value, which is the ratio of re-
tonically increasing (see Equation (7)). duced total generalised freight cost and investment/
2.2. Upper-Level Problem operation cost required for implementing the set of
The upper-level problem approximately optimises actions. The reduced total generalised freight cost is
the combination of freight network improvement considered as the benefit produced by the imple-
actions based on the ratio of the reduced total gen- mented set of actions and can be computed as the
eralised freight cost and the investment/operational difference between the total generalised cost incurred
cost incurred for implementing the actions. This can with and without the actions implemented. Combin-
be undertaken using the benefit-cost ratio. Therefore, ing the actions can generally yield synergistic effects
the objective function for the upper-level problem depending on the actions included. The generalised
is to maximise the BCR value of a combination of cost incurred on each link can be estimated by Equa-
actions to be implemented. tion (6), which is composed of a fare component and a
Let A be defined as A = A1 ∪ A2 ∪ A3 , where A1 = time cost component, with the time spent on each link
a a = 1 2
n is the set of existing links without being assessed by Equation (7). Budget constraints
modifications, A2 = a a = n + 1 n + 2
n + m is on the investment/operation cost are not taken into
the set of existing links with possible actions to be account in this paper, even though this could be pos-
implemented, and A3 = a a = n + m + 1
n + 2m sible by incorporating additional procedures within
is the updated set of A2 after the associated action is the solution techniques for the upper-level problem.
implemented. Links in A2 and A3 can be numbered BCR values are commonly estimated using the total
such that if a ∈ A2 is selected (i.e., if the action asso- benefits and costs generated for the expected duration
ciated with it is implemented), link a + m in A3 will of the action to be implemented. However, in the sub-
replace a; otherwise, a + m in A3 will be discarded. sequent testing and application of the model, because
Here, let y be defined as the set of possible actions the future OD flows for both freight and passengers
associated with A3 (or A2 , y = ya a = n + m + 1 have hardly been estimated accurately yet, the equiv-
n + 2m (or y = ya a = n + 1 n + m, where ya is the alent one-year benefit and cost are used. In this con-
action implementation indicator with a binary value text, the analyses presented in this paper are only
of 1 if the action related to link a on the set of A2 is preliminary, and more detailed BCR analysis should
implemented, and 0 if it is otherwise, the objective be conducted in the future, taking into account the
function, f y, can be represented as follows: project life and social discount rates.
max f y The upper level can be considered as a combi-
i∗ i i∗
i∗ i i∗
i∗ i i∗
natorial optimisation problem. Metaheuristics-based
i∈F a∈A1 ∪A2 x0a ca x0a − a∈A1 xa ca xa − a∈A2 xa ca xa ya
= procedures are used to solve the problem. In the
a∈A2 ba ya past decade, several metaheuristics have been devel-
(5) oped and applied in the field of soft computing.
where, A vital role of these techniques is to solve complex
ca xai ya = ia + i dai xai ya (6) and difficult mathematical programming problems,
which often involve NP-hard problems. These can-
xai not ensure obtaining exact optimal solutions, but can
dai xai ya = ta 1 + 1 xai + 2 (7)
r0a + ya ra provide reasonable and practical solutions. Hence,
Yamada et al.: Designing Multimodal Freight Transport Networks
Transportation Science 43(2), pp. 129–143, © 2009 INFORMS 133
these have been commonly applied to combinatorial Evaluate y and select N − K individuals
optimisation problems wherein exact optimal solu- according to their fitness to obtain y2
tions are hard to determine. Ribeiro and Hansen Draw two solutions N − K times from y2
(2001), Michalewicz and Fogel (2002), Glover and as parent solutions
Kochenberger (2003), Herz and Widmer (2003), and Generate N − K child solutions (i.e.,
Resende and Pinho de Sousa (2004) provided excel- offspring) to obtain y3 by applying uniform
lent introductions and basic concepts of metaheuris- crossover procedure with predefined
Downloaded from informs.org by [190.144.171.70] on 07 August 2016, at 20:11 . For personal use only, all rights reserved.
Existing roads
Existing railway 5 1 2 11 9 6 12
Sea links
New expressway
Road widening 13 10 8
New railway
Rail terminal improvement
Sea port improvement
Downloaded from informs.org by [190.144.171.70] on 07 August 2016, at 20:11 . For personal use only, all rights reserved.
4 3 15 16
7
14
these procedures can provide better performance in the GA-AO procedure, and then the best crossover
combinatorial optimisation problems. and mutation rates in GLS are explored.
(iii) Genetic Local Search (GLS) The model is initially tested on a relatively small
GLS is a procedure that integrates local search tech- transport network in West Java, Indonesia (see Fig-
niques with GA. As shown in the previous section, ure 2) to investigate the performance of the three
a local search operator is inserted after the mutation types of GA-based procedures. This network is com-
process of the GA-AO procedure. Here, a local search posed of 422 links, 105 nodes, 46 OD pairs for freight,
operator for swapping the neighbourhood is utilised, and 52 OD pairs for passenger. There are 16 alterna-
where a gene is randomly selected in every individual, tive actions, as shown in Figure 2 and Table 1, includ-
and each of the neighbouring genes is swapped for ing planned and hypothesised infrastructure projects.
its allele. This can produce three individuals, includ- Improvements, including road widening, are set on
the assumption that the overall capacity is improved
ing the original one. The best individual among these
1.5 times the initial capacity. In the case of seaport
three is determined and preserved for the next gener-
and rail terminal improvement, increased capacities
ation. This can be considered as a kind of “best admis-
sible move strategy” (e.g., Yagiura and Ibaraki 2001).
Table 1 Listing of Actions (West Java Case)
Parameter values for these operators are deter-
mined based on the findings of existing research by Investment/operation
Goldberg (1989), Taniguchi et al. (1999), and Yamada, No. Type of action Location cost (bil. Rp.) BCR∗
Taniguchi, and Noritake (1999). The length of the 1 Road widening Jakarta-Merak 147481 020
chromosome is assumed to be 16 (i.e., m = 16), which 2 Road widening Jakarta-Bogor 61158 052
3 Road widening Padalarang-Bogor 190793 002
is the number of actions used in the succeeding 4 Rail terminal Sukabumi 501297 002
model application to the transport network in the improvement
Philippines. The crossover rate is set to 0.6 and the 5 New seaport Ciwandan 1503890 018
6 New seaport Karawang 1503890 004
mutation rate to 0.03. The number of elites preserved
7 Rail terminal Bandung 401037 016
in each generation is 10 (i.e., K = 10), whereas 50 indi- improvement
viduals (i.e., N = 50) are set in each generation. The 8 Seaport Cirebon 1002593 00001
number of generations (i.e., stopping criterion) is set improvement
9 Road widening Subang-Wado 115699 020
to 100. Existing research indicates that these param- 10 New expressway Bandung-Cirebon 395824 002
eter values are reliable for SGA and GA-AO. How- 11 Road widening Karawang-Jakarta 176958 056
ever, GLS might have more appropriate values for 12 Road widening Cirebon-Karawang 170541 0003
13 New expressway Cikampek-Cirebon 785532 004
crossover and mutation rates. Therefore, the same 14 Road widening Sumedang-Cirebon 167032 013
crossover and mutation rate values as those used in 15 New expressway Cikampek-Padalarang 253656 036
existing research (i.e., crossover rate is 0.6 and muta- 16 Road widening Bandung-Subang 75194 00003
tion rate 0.03) are applied to investigate the effective- ∗
BCR of each action represents the value in case it is individually
ness of incorporating the local search operator within implemented.
Yamada et al.: Designing Multimodal Freight Transport Networks
Transportation Science 43(2), pp. 129–143, © 2009 INFORMS 135
BCR
1.0
relatively cheaper compared with the other types. GLS
Data obtained from the 2001 National OD Survey of 0.5 GA-AO
SGA
Indonesia are used as inputs for interregional freight 0
and passenger movement. Parameter values relating 0 10 20 30 40 50
Downloaded from informs.org by [190.144.171.70] on 07 August 2016, at 20:11 . For personal use only, all rights reserved.
Number of generation
to link performance are mainly estimated using sta-
tistical data on roads, rails, ports, and terminals col- Figure 3 Highest BCR Value in Each Generation
lected from various sources.
Road capacities and speed data are acquired from
the database of the Indonesian Inter-urban Road Man- computational results. In contrast, both Figures 3
agement System (IRMS) and the Indonesian toll road and 4 indicate typical results for 20 computations.
operator PT Jasa Marga. Railway data and related Table 2 reports the BCR values for the best, aver-
information are obtained from the Department of age, and worst solutions found across the 20 runs, as
Communications and the semiprivate railway com- well as the average computational times in seconds to
pany PT KAI. Port information and other sea network complete the search.
data are collected from the Directorate General of GLS provides consistently better performance in
Sea Communication under the Department of Com- terms of the best, average, and worst solutions found
munications. Capacities for loading/unloading links (Table 2). Although both GA-AO and SGA have the
are derived from the ship-handling capacities of vari- potential to produce the same best value as GLS, their
ous ports and terminals. Time components for inspec- performance is not stable. In addition, it can be seen
tion, inventory, and administration of terminals are from Figure 3 that although GLS has already found
assumed, ranging from 6 to 48 hours depending the best solution in the earlier generation, SGA has
on terminal type. The number of berths in ports is not yet reached the best solution even in the 30th
derived from the port’s berth length and the average generation.
ship length for loading/unloading at sea terminals, The average value of the objective function illus-
whereas for rail terminals it is equal to the number of trates the convergence level of the procedures (Fig-
yards. ure 4). GLS has nearly reached convergence after the
The best solution is found to contain action num- 15th generation, whereas other procedures are still far
bers 2 and 9. The test network is relatively small, just from convergence even in the 30th generation and do
a part of Java Island, and road-based freight trans- not appear to sufficiently converge even around the
port has mainly been undertaken in this area. Hence, 100th generation.
road widening could be more effective to improve Average computational times spent for 100 gener-
freight transport in this case. These actions could ations were measured with a PC of 3.4 GHz CPU
result in the alleviation of severe traffic congestion and 1.0 GB RAM. Computational times are in part
around Jakarta, the capital of Indonesia, because of influenced by the number of individuals evaluated.
implementation of action number 2, as well as con- GLS needs more individuals to be assessed, because
tributing to efficient transport movements between it incorporates the local search algorithm. However,
the eastern and western parts of the region because the accumulated number of evaluated individuals
of action number 9. for GLS is 1,808 in the 100th generation (1,476 in
The performance of all three procedures are com- the 30th generation), which is only 2.8% (2.3%) of
pared in Table 2 and Figures 3 and 4. The results the total number of possible combinations for the
shown in Table 2 are obtained after each procedure 16 alternatives.
is tested for a total of 20 runs with different seeds The results indicate that GLS can provide the best
of random values, because all three procedures incor- performance among the three GA-based schemes and
porate several randomised processes that affect their
2.0
GLS GA-AO SGA
Table 2 Computational Results for GA-Based Procedures 1.5
(20 Runs)
BCR
1.0
GLS GA-AO SGA 0.5
leads to a sufficiently good solution even in the 30th Table 3 Performance Comparison in GLS (20 Runs)
generation. This implies that the local search operator
Crossover rate
embedded within the GA-AO procedure works well
to supplement the performance of GA in searching Mutation rate 0.4 0.5 0.6 0.7 0.8
local areas. When the total number of actions is 16, the 0.01
number of generations (i.e., stopping criterion) in GLS Best 1.45 1.45 1.45 1.45 1.45
can therefore be set to 30, which is the case for subse- Average 1.20 1.41 1.36 1.30 1.31
Downloaded from informs.org by [190.144.171.70] on 07 August 2016, at 20:11 . For personal use only, all rights reserved.
quent testing, comparisons, and applications. A lower Worst 0.38 0.68 0.44 0.68 0.20
number of generations and individuals can offer sav- 0.02
ings in computational times. This is effective partic- Best 1.45 1.45 1.45 1.45 1.45
Average 1.41 1.41 1.37 1.41 1.22
ularly when the model is applied to a larger-sized Worst 0.68 0.68 0.68 0.68 0.68
transport network, because the lower-level problem
0.03
requires relatively considerable computational times. Best 1.45 1.45 1.45 1.45 1.45
Existing GLS methods have not always encom- Average 1.41 1.45 1.45 1.41 1.37
passed both the processes of mutation and local Worst 0.74 1.45 1.45 0.68 0.68
search. Therefore, the performance of GLS without 0.04
mutation should be tested to compare with that of Best 1.45 1.45 1.45 1.45 1.45
GLS with mutation. The test is conducted with the Average 1.41 1.45 1.41 1.45 1.37
Worst 0.68 1.45 0.68 1.45 0.68
same parameter values as those used in the experi-
0.05
ment for the three GA-based procedures, apart from
Best 1.45 1.45 1.45 1.45 1.45
the number of generations being set to 30. The results Average 1.45 1.45 1.45 1.41 1.37
of the BCR values for the best, average, and worst Worst 1.45 1.45 1.45 0.68 0.68
solutions found across the 20 runs are 1.45, 1.30, and
0.45, respectively, and the average computational time
for GLS without mutation is 478 seconds (cf., the good and sound solutions for this type and size of
average computational time for GLS with mutation is problems. Higher crossover and mutation rates are,
481 seconds for 30 generations, as shown in Table 4). however, likely to entail more computational times
These results indicate the effectiveness of incorporat- because of the increased number of computational
ing mutation process within GLS in terms of robust- operations required for both the crossover and muta-
ness and superiority of solutions, as compared to tion procedures. Hence, a crossover rate of 0.5 and a
the previous results of GLS with mutation shown in mutation rate of 0.03 are hereafter used in GLS.
Table 2. This also implies a sort of synergistic effect of
using both mutation and local search operators. With 3.2. GLS, TS-Based Procedures, and RSM
more genes being selected in the local search, the per- Performance comparison with other metaheuristics
formance of GLS without mutation could be differ- and approximate techniques is helpful for further
ent from the above. However, this could lead to an investigating the performance of GLS. In general, the
increase in computational time, which could not be performance of metaheuristics can often be greatly
suitable for larger-sized freight transport networks. improved using adjusted parameters and procedure-
The proper crossover and mutation rate values of specific operators. Such comparisons, therefore, might
GLS with mutation are then examined for subse- not be necessarily useful. However, Arroyo and
quent applications of the model to actual large-sized Armentano (2005) point out that suitable implemen-
transport networks. The test is undertaken with the tation designs of tabu search (TS) can lead to algo-
same parameter values as those applied except for rithms that are very competitive to genetic algorithms
the crossover and mutation rates (i.e., length of chro- in single-objective optimisation. Thus, this paper also
mosomes: 16, number of individuals: 50, number of investigates the performance of TS-based procedures
generations: 30, number of elites: 10). Table 3 com- as compared to that of the GLS proposed. The com-
pares the best, average, and worst solutions found parison is undertaken with random search method
across the 20 runs within the range of 0.4 and 0.8 for (RSM) apart from the TS-based procedures.
crossover rate and 0.01 and 0.05 for mutation rate. As noted in Rochat and Taillard (1995, p.
Each combination of crossover and mutation rates 149), “A fundamental principle of TS is to exploit the
given in Table 3 can provide the best solution of 1.45, interplay between diversification and intensification
but the average and worst solutions vary depend- where diversification drives the search to examine new
ing on the crossover and mutation rates adopted. It regions, and intensification focuses more intently on
can empirically be concluded that when the crossover regions previously found to be good.” Potentially, TS
rate is between 0.5 and 0.7 and the mutation rate has many ways of improving its performance on
is between 0.03 and 0.05, GLS can offer relatively specific problems, with the procedures associated with
Yamada et al.: Designing Multimodal Freight Transport Networks
Transportation Science 43(2), pp. 129–143, © 2009 INFORMS 137
the diversification or with the intensification being Table 4 Computational Results for GLS, TS-Based Procedures, and
incorporated. Glover and Laguna (1997) presents a RSM (20 Runs)
variety of such improvements. This paper focuses on GLS TS-B TS-MSM TS-MSM&LM RSM
the intensification-oriented procedures rather than
the diversification-oriented ones, because the GLS Best 1.45 1.45 1.45 1.45 1.02
Average 1.45 0.42 0.82 1.45 0.44
schemes presented above can provide better solutions Worst 1.45 0.18 0.16 1.45 0.20
searching in relatively small space (i.e., small number Ave. CPU time (sec) 481 1,230 1,209 560 1,271
Downloaded from informs.org by [190.144.171.70] on 07 August 2016, at 20:11 . For personal use only, all rights reserved.
Road (R2 = 0.60) and 16), because of the archipelagic nature of the
100 Philippines. These actions would help establish the
corridor running the length of the Philippines, which
Observed link flow (’000t/day)
40
(i.e., value in the numerator in Equation (5)) for the
best solution (i.e., actions 4, 6, 15, and 16), as well
as other sets of actions around it. Either of the road-
20
widening actions of 4 and 6 or seaport improvement
actions of 15 and 16 could provide significant ben-
0
0 20 40 60 80 100 efits for freight transport. However, combining both
Estimated link flow (’000t/day) types of actions apparently brings about synergistic
Rail (R2 = 0.98) effects as manifested by a considerably higher amount
40 of benefits and BCR values, surpassing an increase in
investment cost.
Observed link flow (’000t/day)
Investment/operation
No. Type of action Location cost (mil. Peso) BCR∗
The benefit gained from action set 4 6 15 16 though the costs required for seaport improvement
increases as the value of increases because of are relatively lower. The higher BCR value could
increased traffic congestion on the network. This also be explained by more-congested conditions on
action set reduces the road mode share by about 10% the network because of increased OD freight volume
in each case. However, the best set of actions is not that makes action set 12 15 more efficient. Addi-
necessarily 4 6 15 16 in terms of BCR, except for tional capacity in these seaport improvement actions
the case of = 1
0, because the required investment could considerably mitigate traffic congestion aris-
cost has a significant influence on the BCR value. ing within and around the seaports where increased
In the case of = 0
8, action set 6 15 leads to freight movement exceeds their original capacity.
almost the same amount of benefit as in action set The best set of actions differs between = 0
9 and
4 6 15 16 with lower investment costs. This could = 1
1 as well. These sets commonly include two
largely be attributed to less congested conditions on road-widening actions in the central part of Luzon
the network caused by the reduction in OD freight Island and two seaport improvement actions along the
volume. It would therefore be crucial not to oversup- transport corridors of the Philippines. These actions
ply infrastructure in less congested conditions. can decrease total freight movement in terms of ton-
In the case of = 1
2, action set 12 15 also kilometres resulting from the reduction of road-based
indicates higher BCR value with fewer actions even freight transport. Overall, it can be summarised that
though the benefit obtained from it is significantly road widening in the central part of Luzon as well
lower than action set 4 6 15 16. Thus, invest- as seaport improvement along the Nautical Highway
ment cost greatly influences the BCR value, even could be efficient in terms of freight-related BCR.
(mil. Peso/year)
2.0E+06 16.0
Cost Benefit BCR
1.6E+06 12.0
1.2E+06
8.0
8.0E+05
4.0
4.0E+05
0.0E+00 0.0
)
5)
6)
6)
6)
)
16
16
,6
16
16
,1
,1
,1
,1
(4
5,
5,
5,
5,
,6
,6
15
5
,1
,1
(1
,1
,1
(4
(4
,
(4
(6
,6
,8
,6
,4
,6
(4
(1
(4
Action set
(mil. Peso/year)
1.6E+06 40.0
Cost Benefit BCR
1.2E+06 30.0
8.0E+05 20.0
4.0E+05 10.0
Downloaded from informs.org by [190.144.171.70] on 07 August 2016, at 20:11 . For personal use only, all rights reserved.
0.0E+00 0.0
(6, 15) (4, 6, 15, 16) (6, 7, 15, 16) (4, 6, 15, 16) (4, 6, 15, 16) (4, 6, 14, 15) (4, 6, 15, 16) (12, 15) (4, 6, 15, 16)
β = 0.8 β = 0.9 β = 1.0 β = 1.1 β = 1.2
Action set
5. Conclusions References
This paper presented a model that can help efficiently Aarts, E. H. L., J. H. M. Korst. 1989. Simulated Annealing and Boltzmann
expand the multimodal freight transport network sys- Machines A Stochastic Approach to Combinatorial Optimization
and Neural Computing. John Wiley & Sons, Chichester, UK.
tem. It incorporates GLS, which is an evolutionary
Ackley, D. H. 1987. A Connectionist Machine for Genetic Hillclimbing.
computation technique based on GA, for solving a Kluwer Academic Publishers, Boston.
discrete network design problem. The best combina- Arnold, P., D. Peeters, I. Thomas. 2004. Modelling a rail/road
tion of freight network improvement actions can be intermodal transportation system. Transportation Res. Part E 40
selected from among a feasible set of actions that 255–270.
includes improvement of existing infrastructure and Arroyo, J. E. C., V. A. Armentano. 2005. Genetic local search for
multi-objective flowshop scheduling problems. Eur. J. Oper. Res.
establishment of new roads, rail, and sea links. 151 717–738.
Modelling was undertaken within the framework Aykin, T. 1990. On a quadratic integer program for the location of
of the bilevel programming problem, where the interacting hub facilities. Eur. J. Oper. Res. 46 409–411.
upper-level problem approximately optimises the Balakrishnan, A., T. L. Magnanti, R. T. Wong. 1989. A dual-ascent
combination of actions with metaheuristics-based pro- procedure for large-scale uncapacitated network design. Oper.
cedures such that the freight-related benefit-cost ratio Res. 37 716–740.
Bard, J. F. 1998. Practical Bilevel Optimization Algorithms and
is maximised. Preliminary testing results with the
Applications. Kluwer Academic Publishers, Dordrecht, The
transport network in West Java, Indonesia revealed Netherlands.
that when the size of the transport network is rela- Boyce, D. E. 1984. Urban transportation network equilibrium and
tively large and the search space is relatively small, design models: Recent achievements and future prospectives.
GLS can offer the best performance among the GA- Environ. Planning 16 A 1445–1474.
based and TS-based procedures tested, considering its Braess, D., A. Nagurney, T. Wakolbinger. 2005. On a paradox of
traffic planning. Transportation Sci. 39 446–450.
robustness and faster searching ability.
Bruynooghe, M. 1972. An optimal method of choice of invest-
The model was then applied to the transport ments in a transport network. Presentation, Planning & Trans-
network in the Philippines, where the develop- port Research & Computation Seminars on Urban Traffic Model
ment of a multimodal freight transport network is Reasearch, London, UK.
highly desired. The performance of the modal split- Campbell, J. 1994. Integer programming formulations of discrete
hub location problems. Eur. J. Oper. Res. 72 387–405.
assignment model incorporated within the lower level
Campbell, J. 1996. Hub location and the p-hub median problem.
was examined, and the model adequately represented Oper. Res. 44 923–935.
the major freight flows based on available data. Cascetta, E. 2001. Transportation Systems Engineering Theory
Results also indicated that improvement actions relat- and Methods. Kluwer Academic Publishers, Dordrecht, The
ing to road widening at heavy-loaded links and sea- Netherlands.
port improvement could be more effective than the Chen, M., A. S. Alfa. 1991. A network design algorithm using a
development of new roads and rail links. This result stochastic incremental traffic assignment approach. Transporta-
tion Sci. 25 215–224.
could be significant in the development approach Crainic, T. G., M. Florian, J. Leal. 1990. A model for the strategic
of multimodal freight transport network planning in planning of national freight transportation by rail. Transporta-
archipelagic countries. Road widening in heavily con- tion Sci. 24 1–24.
gested areas, as well as seaport improvements, could Dafermos, S. C. 1980. Traffic equilibrium and variational inequality.
be more effective in establishing the transport corri- Transportation Sci. 14 43–54.
dors running along and across the country. Daskin, M. S. 1995. Network and Discrete Location Models, Algorithms,
and Applications. John Wiley & Sons, New York.
These findings prove that the model proposed in
Davidor, Y. 1991. Genetic Algorithms and Robotics A Heuristic Strategy
this paper could provide adequate results in terms of for Optimization. World Scientific Publishing, Singapore.
the effective design of multimodal freight transport Davis, L. 1991. Handbook of Genetic Algorithms. Van Nostrand,
networks. New York.
Yamada et al.: Designing Multimodal Freight Transport Networks
142 Transportation Science 43(2), pp. 129–143, © 2009 INFORMS
Dell’Amico, M., A. Lodi, F. Maffioli. 1999. Solution of the cumu- Michalewicz, Z., D. B. Fogel. 2002. How to Solve It Modern
lative assignment problem with a well-structured tabu search Heuristics. Springer-Verlag, Berlin.
method. J. Heuristics 5 123–143. Min, H., V. Jayaraman, R. Srivastava. 1998. Combined location-
Diaz, J. A., E. Fernadez. 2001. A tabu search heuristic for the gen- routing problems: A synthesis and future research directions.
eralized assignment problem. Eur. J. Oper. Res. 132 22–38. Eur. J. Oper. Res. 108 1–15.
Dorigo, M., T. Stutzle. 2004. Ant Colony Optimization. MIT Press, Murata, T., H. Ishibuchi. 1994. Performance evaluation of genetic
Boston. algorithms for flowshop scheduling problems. Proc. 1st IEEE
Dorigo, M., G. Di Caro, L. M. Gambardella. 1999. Ant algorithms Internat. Conf. Evolutionary Comput. IEEE, Orlando, FL, 812–817.
Downloaded from informs.org by [190.144.171.70] on 07 August 2016, at 20:11 . For personal use only, all rights reserved.
for discrete optimization. Artificial Life 5 137–172. Nagurney, A. 2000. Sustainable Transportation Networks. Edward
Drezner, Z. 1995. Facility Location A Survey of Applications and Elgar, Northampton, UK.
Methods. Springer-Verlag, Heidelberg. Nagurney, A., J. Dong. 2001. A multiclass, multicriteria traffic net-
Florian, M., H. Spiess. 1982. The convergence of diagonaliza- work equilibrium model with elastic demand. Transportation
tion algorithms for asymmetric network equilibrium problems. Res. Part B 36 445–469.
Transportation Res. Part B 16 477–483. O’Kelly, M. 1987. A quadratic integer program for the location of
Francis, R. L., L. F. McGinnis, J. A. White. 1992. Facility Layout and interacting hub facilities. Eur. J. Oper. Res. 32 393–404.
Location An Analytical Approach. Prentice-Hall, Upper Saddle Poorzahedy, H., M. A. Turnquist. 1982. Approximate algorithms for
River, NJ. the discrete network design problem. Transportation Res. Part
Friesz, T. L., R. L. Tobin, P. T. Harker. 1983. Predictive intercity 16 45–55.
freight network models: The state of the art. Transportation Res. Priemus, H. 1999. On modes, nodes and networks: Technological
Part A 17 409–417. and spatial conditions for a breakthrough towards multimodal
Galinier, P., J. K. Hao. 1999. Hybrid evolutionary algorithms for terminals and networks of freight transport in Europe. Trans-
graph coloring. Combin. Optim. 3 379–397. portation Planning Tech. 23 83–103.
Gao, Z., J. Wu, H. Sun. 2005. Solution algorithm for the bi-level Radcliffe, N. J., P. D. Surry. 1994. Formal memetic algorithms.
discrete network design problem. Transportation Res. Part B 39 T. Fogarty, ed. Evolutionary Computing. Springer-Verlag, Berlin.
479–495. Ravi, R., A. Sinha. 2006. Approximation algorithms for problems
Glover, F., G. A. Kochenberger. 2003. Handbook of Metaheuristics. combining facility location and network design. Oper. Res. 54
Kluwer Academic Publishers, Boston. 73–81.
Glover, F., M. Laguna. 1997. Tabu Search. Kluwer Academic Pub- Reeves, C. R. 1997. Genetic algorithms for the operations researcher.
lishers, Boston. INFORMS J. Comput. 9 231–250.
Glover, F., C. McMillan. 1986. The general employee scheduling Resende, M. G. C., J. Pinho de Sousa. 2004. Metaheuristics Computer
problem: An integration of management science and artificial Decision-Making. Kluwer Academic Publishers, Dordrecht, The
intelligence. Comput. Oper. Res. 15 563–593. Netherlands.
Goldberg, D. E. 1989. Genetic Algorithms in Search, Optimization, and ReVelle, C. S., G. Laporte. 1996. The plant location problem: New
Machine Learning. Addison Wesley, Reading, MA. models and research prospects. Oper. Res. 44 864–874.
Guelat, J., M. Florian, T. G. Crainic. 1990. A multimode multiprod- Ribeiro, C., P. Hansen. 2001. Essays and Surveys on Metaheuristics.
uct network assignment model for strategic planning of freight Kluwer Academic Publishers, Dordrecht, The Netherlands.
flows. Transportation Sci. 24 25–39. Rochat, Y., E. D. Taillard. 1995. Probabilistic diversification and
Herz, A., M. Widmer. 2003. Guidelines for the use of metaheuristics intensification in local search for vehicle routing. J. Heuristics 1
in combinatorial optimization. Eur. J. Oper. Res. 151 247–252. 147–167.
Holland, J. H. 1975. Adaptation in Natural and Artificial Systems. Uni- Rothengatter, W. 1991. Cost-benefit-analyses for goods transport on
versity of Michigan Press, Ann Arbor, MI. roads. M. Kroon, R. Smit, J. Van Ham, eds. Freight Transport
Jaszkiewicz, A. 2002. Genetic local search for multi-objective com- and the Environment. Elsevier, Amsterdam, 187–213.
binatorial optimization. Eur. J. Oper. Res. 137 50–71. Russ, B. F., T. Yamada, J. Castro, H. Yasukawa. 2005. Optimising the
Jaszkiewicz, A., P. Kominek. 2003. Genetic local search with dis- design of multimodal freight transport network in Indonesia.
tance preserving recombination operator for a vehicle routing J. Eastern Asia Soc. Transportation Stud. 6 2894–2907.
problem. Eur. J. Oper. Res. 151 352–364. Sheffi, Y. 1985. Urban Transportation Networks Equilibrium Anal-
Kirkpatrick, S. C., D. Gellat, Jr., M. P. Vecchi. 1983. Optimization by ysis with Mathematical Programming Methods. Prentice Hall,
simulated annealing. Science 220 671–680. Englewood Cliffs, NJ.
Laporte, G. 1988. Location-routing Problems. B. L. Golden, A. A. Shepherd, S. P., A. Sumalee. 2004. A genetic algorithm based
Assad, eds. Vehicle Routing Methods and Studies. North- approach to optimal toll level and location problem. Networks
Holland, Amsterdam, 163–198. Spatial Econom. 4(2) 161–179.
Luo, J. S., Z. Q. Pang, D. Ralph. 1996. Mathematical Programs Southworth, F., B. E. Peterson. 2000. Intermodal and international
with Equilibrium Constraints. Cambridge University Press, Cam- freight modeling. Transportation Res. Part C 8 147–166.
bridge, UK. Stackelberg, H. V. 1934. Marketform and Gleichgewicht. Julius
Magnanti, T. L., R. T. Wong. 1984. Network design and transporta- Springer, Vienna.
tion planning: Models and algorithms. Transportation Sci. 18 Steenbrink, A. 1974. Transport network optimization in the Dutch
1–55. integral transportation study. Transportation Res. Part B 8 11–27.
Melkote, S., M. S. Daskin. 2001. An integrated model of facility Syswerda, G. 1989. Uniform crossover in genetic algorithms. Proc.
location and transportation network design. Transportation Res. Third Internat. Conf. Genetic Algorithms, Morgan Kaufmann
Part A 35 515–538. Publishers, Inc., Fairfax, VA, 2–9.
Merz, P., B. Freisleben. 1997. Genetic local search for the TSP: New Taniguchi, E., M. Noritake, T. Yamada, T. Izumitani. 1999. Optimal
results. Proc. 1997 IEEE Internat. Conf. Evolutionary Comput. size and location planning of public logistics terminals. Trans-
IEEE Press, New York, 159–164. portation Res. Part E 35 207–222.
Yamada et al.: Designing Multimodal Freight Transport Networks
Transportation Science 43(2), pp. 129–143, © 2009 INFORMS 143
Taniguchi, E., R. G. Thompson, T. Yamada, J. H. R. Van Duin. Yamada, T., E. Taniguchi, M. Noritake. 1999. Optimal location plan-
2001. City Logistics Network Modelling and Intelligent Transport ning of logistics terminals based on multiobjective program-
Systems. Pergamon, Oxford, UK. ming method. L. J. Sucharov, ed. Urban Transport V. WIT Press,
Tavasszy, L. A. 1996. Modeling European freight transport flows. Southampton, UK, 449–458.
Unpublished doctoral dissertation, Delft University of Technol-
ogy, Delft, The Netherlands. Yang, H., M. G. H. Bell. 1998. Models and algorithms for road net-
work design: A review and some new developments. Transport
Thomas, R. 1991. Traffic assignment techniques. Avebury Technical.
Rev. 18 257–278.
Yagiura, M., T. Ibaraki. 2001. On metaheuristic algorithms for com-
Downloaded from informs.org by [190.144.171.70] on 07 August 2016, at 20:11 . For personal use only, all rights reserved.
binatorial optimization problems. Systems Comput. Japan 32 Zhang, X., H. Yang. 2004. The optimal cordon-based network con-
33–55. gestion pricing problem. Transportation Res. Part B 38 517–537.