Transportation Research Part A 38 (2004) 305–328

www.elsevier.com/locate/tra

Ferry service network design: optimal fleet size,

routing, and scheduling
M.F. Lai, Hong K. Lo *

Department of Civil Engineering, Hong Kong University of Science and Technology, Clear Water Bay,
Kowloon, Hong Kong, PR China
Received 16 May 2003; received in revised form 3 August 2003; accepted 18 August 2003

Abstract

The study formulated a ferry network design problem by considering the optimal fleet size, routing, and
scheduling for both direct and multi-stop services. The objective function combines both the operator and
passengersÕ performance measures. Mathematically, the model is formulated as a mixed integer multiple
origin–destination network flow problem with ferry capacity constraints. To solve this problem of practical
size, this study developed a heuristic algorithm that exploits the polynomial-time performance of shortest
path algorithms. Two scenarios of ferry services in Hong Kong were solved to demonstrate the perfor-
mance of the heuristic algorithm. The results showed that the heuristic produced solutions that were within
1.3% from the CPLEX optimal solutions. The computational time is within tens of seconds even for
problem size that is beyond the capability of CPLEX.
Ó 2004 Elsevier Ltd. All rights reserved.

1. Introduction

Ferry services in Hong Kong play a supplementary role for serving the cross-harbor traffic but
are essential for serving the outlying islands. The government plays an important role in the
provision of these services by setting up a financially viable environment to entice private sector
participation and hence avoid public subsidies for their operation. The current practice of the
Hong Kong government is to bundle ferry services into packages. Each package, operated by one
company, consists of routes that are profitable and those that are not, thereby allowing cross-
subsidization of routes within each package. It is then up to the operator of each package to

*
Corresponding author. Tel.: +852-2358-8742; fax: +852-2358-1534.
E-mail address: [email protected] (H.K. Lo).

0965-8564/$ - see front matter Ó 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.tra.2003.08.003
306 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328

determine the service schedules, interlining strategies, ferry types (fast and ordinary), and fleet
size, so as to maximize its overall profit by providing an acceptable level of service.
This paper aims at developing a ferry fleet management model and solution algorithm to
optimize the ferry fleet size, ferry routing, and service schedules on one package of routes. The
model is formulated as a mixed integer multiple origin–destination network flow problem with
ferry capacity constraints. It combines the operator as well as usersÕ objectives, as is typically
considered in transit network design studies (see for example, Ceder and Wilson, 1986). UsersÕ
objectives are represented by the total waiting and journey time. The operatorsÕ objectives are
represented by the total system cost of providing the services. Due to limited alternatives for ferry
services to the outlying islands, as is the scenario that this study focuses on, the patronage and
hence total fare revenue in the short run are fixed, subject to the provision of enough capacity to
carry all the demand. In the long run, the demand is a function of many factors, including service
quality, travelersÕ residence and job relocation, and existence of alternative or competitive modes.
Relaxing this assumption on captive demand is our ongoing research. Eventually, by combining
these two measures, a single objective function is formed and optimized, subject to the ferry
capacity constraints.
The problem addressed in this study can be considered as a service network design problem,
which involves determining the service network and its passenger flows simultaneously, so as to
achieve a certain objective. Magnanti and Wong (1984) first formulated this problem as a mixed
integer linear program. Crainic and Laporte (1997) and Crainic (2000) presented a state-of-the-art
review on this topic. This problem finds applications in many contexts. For example, Yan and
Young (1996) developed strategic models for fleet routing and scheduling. Yan and Chen (2002)
studied the scheduling of inter-city bus carriers. Yan and Tseng (2002) developed a passenger
demand model for airline flight scheduling and routing. Kim et al. (1999) studied the service
network design problem for express package delivery. Gorman (1998) integrated various service
network design aspects into a scheduled operating plan for rail. Farvolden and Powell (1994)
presented the service network design model for less-than-truckload (LTL) transportation.
Generally, network design formulations are NP-hard and difficult to solve. Exact and efficient
solution methods generally do not exist, except for special variants of the problem formulation
(Crainic, 2000). A widely used solution method is Lagrangian relaxation. The approach yields a
lower bound on the optimal value of the original formulation by a sequence of multiplier
adjustments to the relaxed sub-problem. Subsequently, a heuristic is developed to find an upper
bound that constitutes a feasible solution. This process is iterated until the gap between the lower
and upper bounds falls within a specified tolerance or until the number of iterations reaches the
specified limit (Yan and Chen, 2002; Yan and Tseng, 2002). The solution quality, computing
efficiency, and convergence properties of the bounding procedures depend on the choice of the
non-differentiable optimization technique to solve the Lagrangian duals and the implementation
and calibration of these methods (Crainic et al., 2001). Kim et al. (1999) developed an approach
by synthesizing column and row generation techniques for a large-scale service network design
problem. Armacost et al. (2002) introduced a composite-based network design formulation for
solving the express shipment service problem, in which the composites are combinations of air-
craft routes that implicitly capture commodity flows. The authors showed that their new for-
mulation better approximates its linear program (LP) relaxation, and therefore is easier to solve
than a traditional network design formulation. Farvolden and Powell (1994) developed an add-
 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328 307

drop heuristic based on the sub-gradients derived from the optimal dual variables of the freight
routing problem. In many cases, heuristics and intelligent search techniques such as simulated
annealing or genetic algorithms are developed for complex problems.
In this study, instead of relying on LP relaxation, a two-phase heuristic algorithm is developed
to efficiently solve the service network problem. Phase I, called path set building, determines a set
of paths (constituting both the ferry routes and schedules) that is feasible and provides a good
upper bound of the optimal solution. The procedure first estimates the contribution of each
service link to be added, and then exploits the shortest path algorithm to determine the ferry route
that passes through the service links that offer the highest combined contribution. In phase II,
called path set rotation, the procedure explores combinations of paths from phase I as a base to
search for solution improvements. The contribution of this algorithm lies with its robustness and
capability of solving problems of practical size in efficient computation times. Its efficiency and
solution quality are illustrated by solving the problem of ferry service provision to outlying islands
in Hong Kong.
The structure of this paper is as follows. In Section 2, we formulate the ferry fleet management
model, discuss model assumptions, and depict the time–space network construction. In Section 3,
we develop and explain the two-phase heuristic solution algorithm. In Section 4, we present the
numerical studies, in which the heuristic is compared with the optimal solution, and the results of
a full-scale implementation of the problem. Finally, in Section 5, we provide some concluding
remarks.

2. Model formulation

2.1. Assumptions

(a) Passenger demand and arrival pattern. As this study focuses on ferry services to the outlying
islands, with limited alternative modes, we assume captive demands. The demands for each
origin–destination (OD) pair, as well as their arrival patterns at the piers are given. In the long
run, people might change their residence and job locations, rendering demand elastic. These
limitations in the present model can be relaxed, however. In the extensions that we are devel-
oping, we model both demand and departure time as elastic functions of the ferry service sche-
dule and quality. But this is outside the scope of this study.
(b) Journey time triangle inequality. That is for any three piers i, j, k, the journey time from i to j
plus from j to k is equal or longer than that from i to k. This assumption should work for
services with the same ferry types, implying that passengers will experience longer journey
times in multi-stop services as compared with direct services.
(c) Overnight empty ferry repositioning. At a specific pier, the number of ferries at berth at the
start of the day is not necessarily equal to that at the end of the day. This assumption is jus-
tified by the fact that the cost of ferry reposition at the end of the day is relatively insignificant
as compared to the total operation cost.
(d) Single fleet type. This study considers a single ferry type. The framework of the model and the
proposed solution algorithm, however, can be extended to include multiple ferry types with
different service characteristics. Such an extension needs to consider passenger preferences
308 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328

and choices on service fare, quality, and journey times (in the case of express and regular ser-
vices), which adds non-linearity to the model. Relaxing this assumption is our ongoing re-
search.

2.2. Variable definitions

This service network design problem involves determining both the ferry routing and service
schedules for the planning horizon, which requires specifying the time dimension within the
formulation. For this purpose, we draw upon the convenience of a time–space network structure,
in which each node represents a specific location at a specific time, whereas each arc represents the
temporal and spatial connection between the two corresponding nodes. The problem involves the
determination of two types of arc variables: (i) ferry arc flows that specify the ferry routes and
departure schedules, and (ii) passenger arc flows that depict the passenger movements given a set
of ferry arc flows. The formulation, therefore, constitutes two types of time–space networks: the
ferry flow and passenger flow networks. Each of these networks can be further divided into a
group of sub-networks to handle different ferry types and different OD demands, as discussed
below.
First, we define the following notations.
Sets
R set of OD pairs
N f , Af sets of nodes and arcs, respectively, in the ferry flow network
N d , Ad sets of nodes and arcs, respectively, in the dth passenger flow network (for demand on
OD pair d)
f f
Nb ,Ne sets of nodes at the beginning and ending of the planning horizon, respectively, in the ferry
flow network, a subset of N f
f d
S ,S sets of service arcs in the ferry and the dth passenger flow network, respectively
W f ,W d sets of waiting arcs in the ferry and the dth passenger flow network, respectively
Od origin arcs in the dth passenger flow network
Dd destination arcs in the dth passenger flow network
Md an artificial node
Parameters
d an OD pair
F fixed cost associated with owning or hiring a ferry for 1 day
V maximum fleet size
d
Bi exogenous passenger demands on OD pair d for nodes i 2 N d at the origin pier
Q ferry capacity
Tij travel time between node i and j
Td travel time between OD pair d based on direct service
b time interval in the time–space network, which is set to be identical for both the ferry and
passenger flow networks and is equal to the time lapse between two consecutive service
arcs
ad fare for OD pair d
 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328 309

Cij operating cost per trip between node i and j

vw value of waiting time
vt value of travel time
Uijf upper bound of ferry flow (i; j) in the ferry flow network
Uijd upper bound of passenger flow (i; j) in the dth passenger flow network
Decision variables
Yij the ferry flow (i; j) in the ferry flow time–space network
Xijd the passenger flow (i; j) in the dth passenger flow time–space network

2.3. Network description

2.3.1. Ferry time–space network

The time–space network of ferry flow in Fig. 1 is defined by a graph GðN f ; Af Þ, in which f
specifies the ferry type. For situations where a single fleet size is considered, only one ferry flow
time–space network is needed. N f is the set of nodes in the time–space network, Af is the set of
arcs representing ferry movements. Af consists of two subsets: service arc set S f and wait arc set
W f , such that Af ¼ S f [ W f .
Each service arc describes a ferry trip, whose journey time, origins and destinations are specified
by the corresponding time–space nodes. Arc flow is represented by a binary variable, which equals
1 for a service arc, 0 otherwise. Arc cost encompasses operating costs, including fuel, mainte-
nance, and labor. The fixed cost of owning or hiring a ferry per day is imposed on arcs originating
from the beginning of the planning horizon. Each wait arc, or a vertical arc in the ferry network,
indicates ferries idling at a pier without providing service. It is represented by a non-negative
integer variable, denoting the number of ferries berthing at a pier. We assume that wait arcs have
negligible operating cost.

Pier I Pier II … Pier N

07:00

07:15 Legend
Time
….
07:30 Wait arc

…
Service arc
23:30

….
24:00

Fig. 1. The ferry time–space network schematic.

310 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328

2.3.2. Passenger time–space network

The time–space network of passenger flow is defined by a collection of graphs GðN d ; Ad Þ, where
d refers to a particular OD pair, N d the set of nodes and Ad the set of arcs representing passenger
movements. Similar to Af , Ad consists of two subsets, service arc set S d and wait arc set W d such
that Ad ¼ S d [ W d . Moreover, associated with every graph GðN d ; Ad Þ are one artificial node, M d
and two artificial arcs: origin arc Od and destination arc Dd . Fig. 2 schematically illustrates the
passenger time–space network.
Service arcs denote passenger trips between piers, whose journey times are as specified in the
corresponding nodes of the time–space network. The arc flow represents the number of onboard
passengers, subject to the capacity of the ferry. The flow on the wait arc, on the other hand,
describes the number of passengers waiting at the pier, which could be a result of either early
arrivals at the pier, or insufficient capacity of the departed ferry to carry all the demand. Similar to
wait arcs in the ferry network, passenger wait arcs are represented by vertical arcs. The formu-
lation accounts for passengersÕ waiting times at the origin piers, which form part of the objective
function to be minimized.
The passenger network associated with each OD pair has an artificial node M d . An origin arc is
constructed to connect M d to the origin node of the last time interval, as illustrated in Fig. 2,
whose arc flow represents the amount of passengers un-served at the origin at the end of the
planning horizon. One may interpret this flow as the unsatisfied or lost demand within the
planning horizon. If serving all demand is an important consideration, one may set a large penalty
for the origin arc, so that more frequent services are arranged to carry all the demand, at the
expense of a higher operating cost. In a similar manner, the passenger network includes a des-
tination arc that connects M d to the destination node of the last time interval. This destination arc
flow represents the total amount of passengers that arrive at the destination during the planning

Origin Intermediate … Destination

Time Pier Piers Pier
07:00
Legend

07:15 Wait arc

….
07:30 Service arc

…
Origin arc

23:30
Destination
arc
….
24:00
Demand
Md arrivals

Fig. 2. The passenger time–space network schematic.

 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328 311

horizon. One may interpret this flow as the total demand carried by the service. Multiplying this
flow by the corresponding OD fare gives the revenue for this OD pair.

2.4. Mathematical formulation

X X X X X
Minimize Z¼ Yij F þ Yij Cij þ Xijd vw b
f f ij2S f d2R ij2ðW d ;Od Þ
i2Nb j2N f nNb
!
X X X X
þ Xijd Tij
d
Xij2DdT
d
vt
Xijd ad ð1Þ
d2R ij2S d d2R ij2Dd
X X
subject to Yij
Yki ¼ 0 8i 2 N f n ðNbf [ Nef Þ ð2Þ
j2N f k2N f
X X
Yij 6 V ð3Þ
f f
i2Nb j2N f nNb
X X
Xijd
Xkid ¼ Bdi 8i 2 N d ; 8d 2 R ð4Þ
j2N d k2N d
X
Xijd 6 Yij Q 8ij 2 S f ð5Þ
d2R

0 6 Xijd 6 Uijd 8ij 2 Ad ; Od ; Dd ; 8d 2 R ð6Þ

0 6 Yij 6 Uijf 8ij 2 Af ð7Þ
Yij 2 Integer 8ij 2 Af ð8Þ
The objective function (1) seeks to minimize the total system costs, comprising five terms: (i)
fixed cost associated with owning or hiring a ferry for the service period; (ii) trip operating cost;
(iii) passenger waiting cost; (iv) penalty cost of multi-stop trip; and finally (v) revenue (i.e.,
expressed in negative terms to offset the costs). All the variable definitions are provided in
Section 2.2. This objective function combines the operator as well as usersÕ perspectives, as is
typically considered in transit network design studies (see for example, Ceder and Wilson, 1986).
In reality, even the objective is to maximize profit; the operator must also consider its service
performance in terms of waiting time and travel time, which have implications on modal shifts
to its competitors in general, and on renewal of its franchised operation for the case of ferry
services in Hong Kong to be specific. In any case, the values of travel time and waiting time can
be considered as parameters or weights to maintain an appropriate tradeoff between these two
objectives.
The firstP
double-sum
P in the objective function gives the total fixed cost. The summation of ferry
flows, i.e., i2N f f Y , measures the number of ferries in operation at the beginning of the
j2N f nNb ij
b
planning horizon. The second summation gives the total trip operating cost. The next double-sum
defines the total passenger waiting cost. The product vw b refers to the cost of waiting for each
passenger on each wait arc. The waiting cost is set to 0 for those wait arcs at the destination pier as
passengers have already arrived at their destination. The fourth summation represents the total
312 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328

multi-stop trip penalty as measured by the cost of additional P travel time experienced by pas-
d
sengers in the multi-stop trip. The term inside the bracket, ij2S d Xij Tij , measures the total
d
experienced travel times in the passenger network for OD pair d. The destination arc flow Xij2D d
d d
represents the total passengers who reach their destination. The product Xij2Dd T represents the
total
P travel time of passengers had they all used direct services. Therefore, the difference between
d d d
ij2S d Xij T ij and Xij2DdT measures the total additional travel time due to passengers using multi-
stop or indirect services. If there is no multi-stop trip, i.e., Tij ¼ T d ; f8ijjXijd > 0 and ij 2 S d g, then
this penalty cost is 0. The last double-sum gives the total revenue, where ad is the fare for OD pair
d.
There are four sets of constraints. Constraint (2) denotes the conservation of ferry flows at each
node i in the ferry network. Constraint (3) requires that the ferries in operation be subject to the
maximum allowable fleet size. Constraint (4) states the passenger conservation condition at every
node in the passenger network after considering the exogenous demand. Constraint (5) combines
the passenger flows of all OD pairs between (i; j) and requires that the total passenger volume be
subject to the ferry capacity on each service arc (i; j). Constraints (6) and (7) provide the upper
bounds of passenger flows and ferry flows between (i; j), respectively. Constraint (8) defines the
ferry flows variables to be integers. Uijf is set to be either 1 or 0 according to whether service is
provided, and a positive integer for wait arcs to account for the situation of multiple ferries idling
at a pier. The former requires that at most one ferry trip is provided between a specific (i; j) at a
specific departure time, whereas in the latter, one may impose a maximum according to the
berthing capacity of the pier.
In this service network design problem, the objective is to determine the ferry flow variable
Yij and the passenger flow variable Xijd (segregated by OD pair d) so as to minimize the total
cost function (1). Exactly how passengers on each OD travel toward their destinations, whether
on direct or multi-stop services, is not explicitly represented. The formulation also does not
explicitly capture whether a ferry service is direct or multi-stop. All of these are to be sorted
out after the solution to (1)–(8) is determined. All in all, the formulation constitutes a mixed
integer linear program, which is hard to solve exactly when the problem size is large, as often
is the case for practical applications. It is, therefore, important to develop efficient heuristics
for solutions.

3. Two-phase heuristic algorithm

A two-phase heuristic solution algorithm is developed to efficiently solve this mixed integer
network flow problem. Phase I, known as path set building, determines a set of paths (ferry routes
and schedules) that gives a feasible and good upper bound of the optimal solution. The procedure
involves the development of a passenger-loading scheme to approximate the change in the
objective function for each change in service provision. It determines the potential contribution of
adding a service arc for both the cases of direct and multi-stop services. A shortest path algorithm
is then applied to the ferry network, whose arcs are represented by their potential contributions.
The shortest path thus obtained constitutes the ferry itinerary that produces the maximum
objective value reduction. Phase II, known as path set rotation, then searches for solution
improvements using the phase I path set as a base. These two phases are operated iteratively, as
 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328 313

discussed below. In these heuristic procedures, we do not need to solve any mixed-integer
problem. The most computationally intensive procedures involve the passenger-loading scheme
and the shortest path algorithm.

3.1. Passenger-loading scheme

In the formulation, the ferry and passenger networks do not represent in an explicit manner
the service types: direct or multi-stop. By solving the formulation (1)–(8), one obtains the ferry
service schedules but without explicitly defining the routing of each ferry. As is customary, a
flow decomposition method is needed to trace the path of each ferry. In a similar manner, the
solution produces the combined passenger loading for each service arc but without distin-
guishing among those taking direct versus multi-stop services. On the other hand, given the ferry
service schedules (i.e., Yij ), one can determine the passenger loading (i.e., Xijd ) on each service arc
by solving the same formulation (1)–(8) but treating Yij as fixed, which reduces the formulation
to a linear program.
In determining the contribution of a particular service arc addition, or in general studying the
sensitivity of a particular service change, one must solve the passenger-loading linear program
once. When the network gets large, due to the huge feasible space in service arc provisions or
modifications, one may need to solve the corresponding linear program many times, which could
be computationally demanding. An alternative is to follow the approach of Farvolden and Powell
(1994) which developed a sub-gradient logic to estimate the potential total cost saving if a link is
added to or dropped from the network. The accuracy of the sub-gradients depends on the
problem nature, which could be problematic in the presence of active capacity constraints as in
this study. Moreover, the inclusion or exclusion of a link may create a new set of feasible paths
that make this saving estimation complicated. To avoid this complexity, we simplify the total cost
saving (known as service arc contribution in this study) by working with the service provisions
directly, as explained in the following. This simplification avoids any linear program resolutions,
which significantly enhances the computation time.
To explicitly represent the services, we label the ferry services as either direct or multi-stop. In
terms of notation, we use uppercase variables (e.g. I; N ) to denote piers in general and corre-
sponding lower case variables (e.g. i; n) to include both the pier location and time in the time–
space network. Let the direct service serving OD pair d ¼ ði; nÞ be represented by apð1Þ;pð2Þ where
the subscript refers to the origin pier pð1Þ ¼ i and the destination pier at pð2Þ ¼ n. Denote DSdf to
be the set of direct services serving OD pair d. More generally, we use apð1Þ;...;pðkÞ to denote a k-stop
service serving OD pair (i; n), where pð1Þ ¼ i, pðkÞ ¼ n, and the intermediate stops are: pðjÞ,
f
2 6 j 6 k
1. Denote MSd;k to be the set of k-stop services connecting OD pair d. Actually the term
direct or multi-stop is specific to the particular OD pair. For OD pair (I; N ) as shown in Fig. 3, the
service arc apð1Þ;pð2Þ , pð1Þ ¼ i, pð2Þ ¼ n1 or ai;n1 constitutes a direct service. On the other hand,
apð1Þ;pð2Þ;pð3Þ , pð1Þ ¼ i, pð2Þ ¼ ii, pð3Þ ¼ n2 or ai;ii;n2 constitutes a three-stop service. Note that this
three-stop service passes through Pier II; hence demands for OD pairs (I; II) and (II; N ) many also
benefit. In this representation, each service arc represents the routing or trajectory of a particular
ferry (not just a service schedule).
Once the service provision (either direct or multi-stop) is specified, one can apply the following
passenger-loading scheme to determine its contribution:
314 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328

Pier I Pier II … Pier N

i

ii
time Multi-stop arc a i,ii,n2

Direct arc a i,nl

n1
n1

n2
n2

Fig. 3. Direct and multi-stop service arcs of OD pair (I; N ).

A. Each service arc addition is intended to serve a particular OD pair, defined to be the first and
final stops of the service. Direct services will only carry the corresponding OD demands. For
multi-stop services, they will also carry demands that are along the ferry route, subject to the
availability of space and loading priority. For example, referring to Fig. 3, the multi-stop ser-
vice arc ai;ii;n2 is intended to serve the OD pair (I; N ), or more specifically, the OD pair (i; n2) in
the time–space network. Nevertheless, OD pairs (i; ii), (ii; n2) in the time–space network may
also benefit.
B. Passenger arrivals at the origin pier are loaded to the first available ferry that heads toward
their destinations (in an either direct or multi-stop manner). Computationally, this heuristic
procedure, as we will show later, actually approximates the performance of the optimal solu-
tion quite well. In reality, this assumption is reasonable with published service schedules. As
one expects, most commuters would come to the pier according to the departure time of
the service desired.
C. Based on the nodal conservation condition (4) at the origin pier, the demand to be loaded onto
the added service is determined.
D. Due to the ferry capacity constraint (5), and depending on the demand situation, some pas-
sengers may not be able to board the added ferry service. Because of this, a set of loading pri-
orities is required.
E. Loading priority follows these rules: for the k-stop service apð1Þ;...;pðkÞ ,
iii. The highest priority goes to passengers on the OD (pð1Þ; pðkÞ). Denote ‘‘’’ to be the pri-
ority indicator. The priority of loading follows this order:
ðpð1Þ; pðkÞÞ  ðpð1Þ; pðk
1ÞÞ      ðpð1Þ; pð2ÞÞ
iii. The next set of priority is given to the demand starting from the origin pð2Þ:
ðpð2Þ; pðkÞÞ      ðpð2Þ; pð3ÞÞ
iii. In general, referring to the following expression with l ¼ 0; . . . ; k
1, the set with a smaller
l has a higher priority; and for a given l, their priority decreases in this manner:
ðpðl þ 1Þ; pðkÞÞ  ðpðl þ 1Þ; pðk
1ÞÞ      ðpðl þ 1Þ; pðl þ 2ÞÞ
iv. The principle is to fully utilize the additional ferry service along its multi-stop route to the
greatest extent possible, starting from the first stop of the service, and then proceeding to
 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328 315

the next stop. Generally, passengers who are able to use the service will lower their waiting
times (as compared with the previous situation without the service). The more passengers
this additional arc serves, the more benefit the system gains, contributing to a larger drop
in the objective function value.

For exposition purposes, we illustrate the procedure with two examples. One involves a direct
service, the other a multi-stop service. Fig. 4 shows the addition of a direct service serving Pier I to
I;N I;N I;N
Pier N from node i2 to n1. The variables Xi1;i2 ; Xi2;i3 ; Xi2;n1 denote passenger flows between Pier I
I;N
and Pier N at the specified nodes, and Bi2 the exogenous demand. The ferry capacity is repre-
sented by Q. The ferry capacity and conservation conditions at node i2 can be expressed as:
I;N I;N
Xi2;n1 ¼ minðQ; Xi1;i2 þ BI;N
i2 Þ
I;N I;N
Xi2;i3 ¼ Xi1;i2 þ BI;N I;N
i2
Xi2;n1

By solving this set of conservation and ferry capacity conditions recursively, one for each node at
Pier I, starting from the beginning of the planning horizon, we obtain the resultant changes in
passenger flows, and hence the change in the objective function (1). Note that we are not solving a
system of simultaneous equations that involve the minimum operator, but merely executing the
equations sequentially from top to bottom, and recursively from the start of the planning horizon.
Computationally this is rather straightforward, involving no tedious operations.
In the above example of providing a direct service, only the target OD pair (I; N ) directly
benefits. The following example illustrates the case of a multi-stop service, as shown in Fig. 5,
going from Pier I to Pier N via Pier II. This multi-stop service is intended to serve OD pair (I; N ).
However, OD pairs (I; II), (II; N ) may also benefit. Following the same notations for passenger
flows and exercising the priority rules, one obtains these sets of conservation and ferry capacity
conditions, one set for each OD pair, starting from the one with the highest priority:
8
< X I;N I;N I;N
i2;ii2;n1 ¼ minðQ; Xi1;i2 þ Bi2 Þ
ðaÞ ðI; NÞ I;N I;N
: Xi2;i3 ¼ Xi1;i2 þ BI;N I;N
i2
Xi2;ii2;n1

Pier I Pier N

i1
I,N
X
il, i2
I,N
BiIN
B 22
,
i2
I,N
I,N X i2, nl
X
i2, i3

i3
n1
n1

Fig. 4. Illustrative example of a direct service.

316 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328

Pier I Pier II Pier N

ii1
1

I,N
Bi 2
B , BiI2, I I ii2
2 iii1
i1

I I, N
ii3
3 iii2
i2 Bii2
B

iii3
i3 nn1
1

Fig. 5. Illustrative example of a multi-stop service.

(
I;II I;N I;II
Xi2;ii2 ¼ minðQ
Xi2;ii2;n1 ; Xi1;i2 þ BI;II
i2 Þ
ðbÞ ðI; IIÞ I;II I;II I;II I;II
Xi2;i3 ¼ Xi1;i2 þ Bi2
Xi2;ii2

(
II;N I;N II;N
Xii2;n1 ¼ minðQ
Xi2;ii2;n1 ; Xii1;ii2 þ BII;N
ii2 Þ
ðcÞ ðII; NÞ II;N II;N II;N II;N
Xii2;ii3 ¼ Xii1;ii2 þ Bii2
Xii2;n1
The passengers on the OD pair (I; N ) are loaded to the ferry first, followed by (I; II), and then by
(II; N ). In loading passengers for OD pair (I; II), one must account for the space that has been
I;N
occupied by passengers on OD pair (I; N), Xi2;ii2;n1 , which is subtracted from the ferry capacity.
Similar considerations apply to the OD pairs that are of lower priority and located at downstream
parts of the ferry route. Similar to the case of the direct service arc, these equations are executed
from top to bottom or from top priority to low priority. They only involve algebraic operations
plus the minimum operator of two terms. After determining the resultant changes in passenger
flows due to the introduction of this multi-stop service arc, one can determine the change in
objective function value, which provides a measure for the contribution of adding this multi-stop
arc.

3.2. Phase I––path set building

Phase I determines a set of ferry paths that is feasible and forms a good upper bound on the
optimal solution. It involves an iterative process that adds ferry services based on this criterion:
can the additional ferry service contribute to reducing the total system cost? This question is
answered through the passenger-loading scheme that estimates the service arc contributions.
After which, a shortest path algorithm is applied to the network comprising these service arc
contributions, to determine the combination of service arcs that produces the maximum
objective value reduction. This shortest path on the time–space network constitutes the actual
routing or trajectory of the ferry to be added. This whole process is repeated until no more
paths can be added. Fig. 6 shows a flow chart of this iterative procedure, whose steps are ex-
plained in the following.
 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328 317

Fig. 6. Flow chart of the two-phase heuristic.

3.2.1. Step 1. Service arc contribution determination

In this step, by using the passenger-loading scheme discussed above, we determine the con-
tribution of each service arc, including both direct and multi-stop services. Mathematically, each
individual arcÕs contribution is obtained by approximating the change in the objective function (1)
by setting the corresponding value of Yij ¼ 1 for direct services serving OD pair d, or
Ypð1Þ;pð2Þ ¼    ¼ Ypðk
1Þ;pðkÞ ¼ 1 for k-stop services, notated as Yq ¼ 1, where q defines the routing of
the k-stop service. The service arc contributions for direct and multi-stop services, respectively, are
denoted as:

C t ðaij Þ ¼ ZðYt
1 ; Yij ¼ 1Þ
ZðYt
1 Þ 8aij 2 DSdf ð9Þ

f
C t ðapð1Þ;...;pðkÞ Þ ¼ ZðYt
1 ; Yq ¼ 1Þ
ZðYt
1 Þ 8apð1Þ;...;pðkÞ 2 MSd;k ð10Þ

where the variable definitions are:

C t ðaij Þ contribution of direct service aij 2 DSdf on OD pair d at iteration t
f
C t ðapð1Þ;...;pðkÞ Þ contribution of k-stop service apð1Þ;...;pðkÞ 2 MSd;k on OD pair d at iteration t
ZðÞ objective function value as determined with the passenger-loading scheme
Yt
1 ferry flow variables fixed at iteration t
1
Yq ferry flow variables associated with the k-stop service serving OD pair d
318 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328

The argument ðYt
1 ; Yij ¼ 1Þ of the objective function ZðÞ in (9) denotes updating the value of
Yij to be 1 while holding the rest of Yt
1 to be the same as in the last iteration. The same notation
applies to the argument in (10), that is, updating the elements in Yq to be 1 while holding the rest
of Yt
1 unchanged.
To initiate the procedure, we let t ¼ 1 and Y0 ¼ 0 represent the network without any ferry
services. The procedure is repeated for all service arcs that are not assigned (i.e., Yij ¼ 0 or Yq ¼ 0)
at the current iteration. At the end of Step 1, the service arc contributions of all unassigned arcs
are determined.

3.2.2. Step 2. Shortest path determination

Based on the service arc contributions determined in (9) and (10), Step 2 determines the path
that connects the service arcs to obtain the maximum combined objective value reduction. By
expressing service arc contributions in negative units, the path with the lowest (or most negative)
service arc contributions gives the maximum objective value reduction. Such a path can be
conveniently determined with the shortest path algorithm on the service arc contributions. This
ferry service network (Figs. 1 and 2) contains no cycles as time is explicitly captured in the nodes,
therefore common shortest path algorithms can work with these negative values in arc costs.
The shortest path determined in this way is merely a construct to find the ferry itinerary that
gives the maximum reduction in objective function. The system is free to choose where to start the
ferry at the beginning of the planning horizon and where to finish at the end of the planning
horizon. In Step 2, we solve the shortest path tree N times, where N is the number of origin piers at
the beginning of the planning horizon. Specifically, for this purpose, the origin and destination sets
are, respectively, {i 2 Nbf } and {j 2 Nef }. Let the shortest path associated with OD pair (i; j) at the
tth iteration be denoted as Pijt . Among the set of shortest paths associated with different (i; j)Õs as
OD pairs, the final shortest path Prst chosen is the one that produces the largest contribution, i.e.,
Prst ¼ arg minfF ðPijt Þ; 8i 2 Nbf ; 8j 2 Nef g ð11Þ
where F ðPijt Þ is the contribution (in negative terms) of each shortest path Pijt .

3.2.3. Step 3. Ferry deployment decision

In Step 2, we determine the best contributing path at the current iteration. As the iteration
proceeds, more and more ferries are added. Paths chosen in later iterations will have lower
marginal contributions due to the existence of services introduced in prior iterations. On the other
hand, each additional ferry must pay for its fixed and operation costs, regardless of its passenger
volume. The system will come to a point when adding more ferry services will not be cost-effective.
In deciding whether a ferry serving the path Prst (determined in Step 2) should be added, one must
ensure that:
ZðYt
1 ; fYa ¼ 1; 8a 2 Prst gÞ 6 ZðYt
1 Þ ð12Þ
or the additional ferry introduces a further drop in the objective function. The notation for the
argument of ZðÞ on the left hand side of (12) follows from the earlier definition, i.e., setting
{Ya ¼ 1; 8a 2 Prst } while holding the rest of Yt
1 unchanged. If (12) is satisfied and the maximum
allowable fleet size is not violated, the procedure goes to Step 4. Otherwise, no more ferries will be
considered for addition. The procedure proceeds to phase II.
 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328 319

3.2.4. Step 4. Variable updating

After deploying an additional ferry, the additional ferry path Prst is appended to the end of the
deployable ordered path set Pk , where k is the index associated with the phase II procedure to be
discussed later. To initialize the phase I procedures when t ¼ 1, k is set to one and the deployable
path sets P1 and P0 are empty, i.e., P1 ¼ P0 ¼ ;, where P0 is used to initialize the phase II pro-
cedure. The ferry flow variables are updated by setting {Ya ¼ 1; 8a 2 Prst } in Yt
1 to form Yt and
the corresponding passenger flows are updated in accordance with the passenger-loading scheme.
The process continues to Step 1.
Through the above four steps, a feasible and good upper bound of the optimal solution is
obtained. This solution satisfies all the constraints, including flow conservation (passenger and
ferry), capacity limitation, and integrality conditions. Furthermore, the ferry deployment decision
in Step 3 ensures that any additional deployment must produce a reduction in the value of the
objective function. Phase I stops when no more ferry paths can be added, as determined in Step 3.

3.3. Phase II––path set rotation

Phase II improves the solution by altering the path set determined in phase I. In phase I, we
iteratively add shortest paths while holding the prior shortest paths selected as fixed. Thus, at a
given iteration t in phase I, the prior path set forms a base for generating the next shortest path.
The service arc contribution, which measures the marginal contribution of a service arc, depends
on the prior services introduced. On the other hand, each shortest path introduced takes away
passengers from the prior path set, hence reducing its contribution. In the end, with paths being
added gradually, the actual contributions of paths added early on are lower than their initial
estimations. That is, in estimating path contributions, this sequential process of adding paths
favors the earlier paths and biases against the later paths. The purpose of phase II, therefore, is to
explore whether the solution can be improved by rotating the prior path set in a systematic
manner.
One may recall that at the end of phase I, no more ferry paths can be added. Let k be the
index for phase II iterations, Pk be the ordered path set generated at the end of phase I,
wherein the order captures the sequence of the paths being added, and B be the ordered path
set used to form the new base. At the beginning of phase II, if the objective value is improved,
B is updated to be the path set Pk . Then, the first path from B is removed to form a new base
for the next iteration k þ 1, which is fed back into phase I to generate new shortest paths. The
rationale, as explained above, is to remove the earlier paths sequentially to explore whether the
solution can be further improved. The heuristic procedures for both phases I and II are shown
in Fig. 6.
Phase II terminates when the predefined number of iterations is reached or when the path set at
iteration k, Pk , repeats an earlier set, implying that the procedure will repeat itself in a cyclic
manner. The final solution P is obtained from selecting the path set that gives the minimum
objective function among all the phase II path sets, i.e.:

P ¼ arg minfZðYðPk ÞÞ; 8k 2 Phase II path setsg ð13Þ

where YðPk Þ ¼ fYa ¼ 1; 8a 2 Prst ; 8Prst 2 Pk ; Ya ¼ 0 otherwiseg

320 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328

4. Numerical studies

We implement the heuristic algorithm for a ferry route package in Hong Kong. To compare
computational performance, the commercial optimization package CPLEX-6.0-MIP (ILOG,
1998) is applied to solve the problem formulation (1)–(8) exactly. The computations are imple-
mented on a Pentium III 900 PC. We first solve the problem for the morning peak. The for-
mulation involves about 100 integer and binary variables, and 404 real variables. The purpose of
this comparison is to assess the solution quality of the heuristic as compared with the exact
solution.
To apply the heuristic to a problem of practical size, we solve the ferry fleet assignment for the
whole day, which involves about 850 integer and binary variables, and over 3000 real variables.
This problem size, in terms of the number of integer and binary variables, is beyond the capability
of CPLEX. In fact, we were unable to solve it with CPLEX using one week of computational time.
On the other hand, we are able to solve this whole-day problem with the heuristic in tens of
seconds. The solution itinerary compares well with the peak-period itinerary.

4.1. Problem setting

The problem contains two ferry routes: CBD–Mui Wo (C-MW) and CBD–Peng Chau (C-PC).
Both MW and PC are outlying islands. These two routes share similar characteristics in terms of
patronage, journey time, and fare. The average daily patronage of the routes C-MW and C-PC
are around 6500 and 5500 person-trips, respectively. The patronage ratio between the peak and
off-peak periods is 1.26 according to the Hong Kong Third Comprehensive Transport Study
(Transport Department, 1999). Table 1 shows the OD demand pattern in the morning peak. Table
2 presents the trip operating cost matrix. The travel times are shown in Table 3. Table 4 sum-
marizes the parameter values, including ferry fixed cost, fare, value of time, network time interval,
maximum fleet size, and ferry capacity.

Table 1
Demand distribution in the morning peak period
OnD CBD Mui Wo Peng Chau
CBD – 182 154
Mui Wo 1631 – –
Peng Chau 1381 – –

Table 2
Trip operating cost (HK$)
OnD CBD Mui Wo Peng Chau
CBD – 900 675
Mui Wo 900 – 225
Peng Chau 675 225 –
 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328 321

Table 3
Origin–destination travel times (min)
OnD CBD Mui Wo Peng Chau
CBD – 50 45
Mui Wo 50 – 15
Peng Chau 45 15 –

Table 4
Parameter values
F ad vw and vt b V Q

1
$5000 $10.5 $40 h 15 min 6 ferries 1000 passengers

Demand
450

400 PC to C

MW to C
350
C to PC
300
C to MW
250

200

150

100

50

0
Time
07:00 07:15 07:30 07:45 08:00 08:15 08:30 08:45 09:00

Fig. 7. The demand patterns over the 2-h morning peak period.

First, we solve the problem for the 2-h morning peak (7:00–9:00 a.m.). As shown in Fig. 7, the
demand patterns from both islands are positively skewed over time, as work trips from the
outlying islands concentrate more between 7:00 and 8:00 a.m. than between 8:00 and 9:00 a.m. As
for the demand generated from Central, it is normally/symmetrically distributed over the 2-h
period.
After solving the peak period problem, we solve the problem for the whole day. Two scenarios
are included: scenario I contains only direct services; scenario II includes both direct and multi-
stop services. These two scenarios assess whether the operator and passengers can benefit from
multi-stop services.

4.2. Results for the peak-period problem

The formulation involves 36 integer variables, 62 binary variables, and 404 real variables, and a
total of 252 constraints. The optimal ferry routes and schedules solved by CPLEX are presented in
Table 5.
Scenario II contains four multi-stop journeys: three from Mui Wo via Peng Chau to CBD and
one from CBD via Peng Chau to Mui Wo. Comparing with scenario I, the combined objective
value is reduced by over 14%, demonstrating the advantage of multi-stop services. Specifically,
322 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328

Table 5
Optimal ferry routes and schedules for scenarios I and II
Scenario I Scenario II
CBD Peng Chau CBD Peng Chau
8:15 a.m. 7:15 a.m. 8:15 a.m. 7:15 a.m.a
9:00 a.m. 7:30 a.m. 9:00 a.m.b 7:30 a.m.
8:00 a.m. 8:00 a.m.a
9:00 a.m. 8:30 a.m.a
9:00 a.m.
CBD Mui Wo CBD Mui Wo
8:00 a.m. 7:30 a.m. 8:00 a.m. 7:00 a.m.a
8:30 a.m. 8:00 a.m. 8:30 a.m. 7:30 a.m.
9:00 a.m. 9:00 a.m. 9:00 a.m.b 7:45 a.m.a
8:15 a.m.a
9:00 a.m.
Z  ¼ 27,070.5 Z  ¼ 23,135.5
Optimal fleet size ¼ 5 Optimal fleet size ¼ 5
Average passenger waiting time ¼ 12.43 min Average passenger waiting time ¼ 7.29 min
Percentage of multi-stop trip users ¼ 23.21%
a
From Mui Wo via Peng Chau to CBD.
b
From CBD via Peng Chau to Mui Wo.

they lead to a substantial reduction in average passenger waiting time (from 12.43 to 7.29 min), at
the expense of a modest increase in average travel time (3.48 min). Although the current practice is
to offer only direct services, the result indicates that multi-stop services offer a potential strategy
for substantial cost reductions.
On the performance of the heuristic, for scenario I, Fig. 8 shows how the objective value and
other cost components vary with the number of ferries deployed in phase I of the heuristic
algorithm. The total passenger waiting time decreases with the number of ferry paths added,
whereas the total trip operating cost and fixed cost increase. Revenue is shown with a negative

Cost (HKD)
350000

322930
300000 Objective value
Total passenger waiting cost
250000 Total trip operating cost
Total fixed cost
Total negative revenue
200000 198302.5

150000

100000
82617.5
36455 31169.5 30709.5 30304.5 32449.5
50000
45290

0
0 1 2 3 4 5 6 7 8 9
Number of ferry
-50000

Fig. 8. The objective value and cost components against ferry deployment for scenario I.
 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328 323

value, as the objective function is expressed in cost terms. Total negative revenue decreases and
reaches a constant value when all the passengers are served. The objective value is decreasing until
the 8th ferry path addition, indicating that it is not cost-effective to add this 8th ferry. Therefore,
the phase I procedure stops, with the first seven ferry paths generated constituting a good upper
bound for the optimal solution.
At the end of phase I, the path set rotation procedure of phase II starts, whose results are
summarized in Table 6. At the start of phase II, ZðYðP1 ÞÞ is smaller than ZðYðP0 ÞÞ since P0 is an
empty set. Therefore, B is set to be P1 (paths 1–7). Path 1 is then removed from B to form the new
base for k ¼ 2 (paths 2–7), which is fed back into phase I. When the next iteration of phase I
terminates, a new ferry path 8 is generated. Since ZðYðP2 ÞÞ is not smaller than ZðYðP1 ÞÞ, no
update is implemented on path set B. The new base for k ¼ 3 contains paths 3–7 only, with path 2
removed. When k ¼ 5, ZðYðP5 ÞÞ is smaller than ZðYðP4 ÞÞ, B is updated to contain paths 5–10. The
new base for k ¼ 6 contains paths 6–10 after removing path 5 from B. The procedure is repeated
until the predefined number of k iterations is reached. Or, in this example, the path set at k ¼ 18 is
the same as the path set at k ¼ 1. 1 Continuing the procedure further will not result in further
improvement, as the result for k ¼ 19 is equivalent to that of k ¼ 2. In this case, the phase II
procedure stops. The final solution is the path set that attains the minimum Z in Table 6, or P13
whose objective value is 27,410.50.
For scenario II containing both direct and multi-stop services, the heuristic produces a set of
similar results. In the interest of space, the detailed figure and table are not shown. The phase II
procedure produces repeated results at k ¼ 22 and therefore is terminated. The solution path set
P6 attains the minimum objective value of 23,423.00 (at k ¼ 6).
Table 7 compares the solutions between the heuristic algorithm and CPLEX. The heuristic
algorithm produces solutions whose objective functions are within 1.3% from those of the optimal
solutions, while using around 0.5% of CPLEXÕs runtime or around 10 s. Moreover, according to
our computational experience, the CPLEX runtime is highly sensitive to the maximum allowable
fleet size V in budget constraint (3). Increasing the maximum allowable fleet size to seven in these
scenarios, for example, results in a substantial increase in runtime as compared with the ones
shown in Table 7. In some cases, the problem is not solvable by CPLEX within one week.
However, this dramatic runtime variation does not happen in the heuristic algorithm.
The ferry schedule produced by the heuristic follows closely that produced by CPLEX, as
shown in Tables 8 and 9. The performance measures also match closely, in terms of average
passenger waiting time (12.99 min versus 12.43 min) for scenario I, average passenger waiting time
(7.41 min versus 7.29 min) and multi-stop passenger proportion (23.1% versus 23.2%) for scenario
II. Also, the fleet sizes for both solutions are 5.

4.3. Results for the whole-day problem

The problem size of the formulation for 1 day (7:00 a.m.–11:30 p.m.) increases substantially,
consisting of 293 non-negative integer variables, 527 binary variables and 3425 real variables. It is

1
The paths in Table 6 are numbered according to the iteration index t when they are added, not according to their
physical characteristics or itineraries.
 324
M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328
Table 6
Phase II path set rotation results for scenario I
k Path set P k Z
1 1 2 3 4 5 6 7 30,304.50
2 2 3 4 5 6 7 8 30,304.50
3 3 4 5 6 7 8 9 10 32,274.50
4 4 5 6 7 8 9 10 11 32,274.50
5 5 6 7 8 9 10 28,604.50
6 6 7 8 9 10 11 28,604.50
7 7 8 9 10 11 12 28,539.50
8 8 9 10 11 12 13 27,979.50
9 9 10 11 12 13 14 27,979.50
10 10 11 12 13 14 15 27,979.50
11 11 12 13 14 15 16 27,979.50
12 12 13 14 15 16 27,464.00
13 13 14 15 16 17 27,410.50
14 14 15 16 17 18 27,410.50
15 15 16 17 18 19 27,410.50
16 16 17 18 19 20 28,570.50
17 17 18 19 20 21 28,570.50
18 18 19 20 21 22 23 24 30,304.50
Note: paths in bold italics form the new base for the current iteration k.
 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328 325

Table 7
Solution performance comparisons
Objective value Computation time (s)
CPLEX Heuristic Difference CPLEX Heuristic Heuristic/CPLEX
Scenario I 27,070.5 27,410.5 1.3% 1749 11 0.6%
Scenario II 23,135.5 23,423.0 1.2% 2990 13 0.4%

Table 8
The ferry schedule from CPLEX and the heuristic for scenario I
Scenario I (CPLEX) Scenario I (Heuristic)
CBD Peng Chau CBD Peng Chau
8:15 a.m. 7:15 a.m. 8:15 a.m. 7:15 a.m.
9:00 a.m. 7:30 a.m. 9:00 a.m. 7:30 a.m.
8:00 a.m. 8:00 a.m.
9:00 a.m. 9:00 a.m.
CBD Mui Wo CBD Mui Wo
8:00 a.m. 7:30 a.m. 8:00 a.m. 7:30 a.m.
8:30 a.m. 8:00 a.m. 9:00 a.m. 8:00 a.m.
9:00 a.m. 9:00 a.m. 9:00 a.m.
Optimal fleet size ¼ 5 Fleet size ¼ 5
Average pass. waiting time ¼ 12.43 min Average pass. waiting time ¼ 12.99 min

Table 9
The ferry schedule from CPLEX and the heuristic for scenario II
Scenario II (CPLEX) Scenario II (Heuristic)
CBD Peng Chau CBD Peng Chau
a
8:15 a.m. 7:15 a.m. 8:15 a.m. 7:15 a.m.a
9:00 a.m.b 7:30 a.m. 8:45 a.m.b 7:30 a.m.
8:00 a.m.a 8:00 a.m.a
8:30 a.m.a 8:30 a.m.a
9:00 a.m. 9:00 a.m.
CBD Mui Wo CBD Mui Wo
a
8:00 a.m. 7:00 a.m. 8:00 a.m. 7:00 a.m.a
8:30 a.m. 7:30 a.m. 8:30 a.m. 7:30 a.m.
9:00 a.m.b 7:45 a.m.a 8:45 a.m.b 7:45 a.m.a
8:15 a.m.a 8:15 a.m.a
9:00 a.m. 9:00 a.m.
Optimal fleet size ¼ 5 Resulted fleet size ¼ 5
Average pass. waiting time ¼ 7.29 min Average pass. waiting time ¼ 7.41 min
Percentage of multi-stop trip users ¼ 23.2% Percentage of multi-stop trip users ¼ 23.1%
a
From Mui Wo via Peng Chau to CBD.
b
From CBD via Peng Chau to Mui Wo.
326 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328

beyond the capability of CPLEX to determine the optimal solution. Therefore, we solve this
whole-day problem only with the heuristic algorithm. For this whole-day problem, we take the
fixed cost to be $15,000 per day per ferry. The operational costs are already expressed on a per
journey basis. The OD demand matrix for the whole day is obtained from the Hong Kong Third
Comprehensive Transport Study (Transport Department, 1999).
The computation result of the heuristic algorithm is highlighted in Table 10. The heuristic
produces the result that the fleet size for the whole day is the same as the optimal fleet for the peak
period. This is reasonable as the morning peak is typically the busiest. Therefore, one would not
expect a fleet size calculated for the whole day to be larger than that for the morning peak. The
average passenger waiting times for both scenarios I and II are longer than those of the peak
period. This is a reasonable result as the headways of off-peak services are generally longer,
causing longer average daily waiting time. More passengers take the multi-stop services for the
whole-day problem. Again, this is reasonable as multi-stop services are more cost-effective for the
off-peak period when the demand is low. In order to further assess the quality of this heuristic
solution, the peak schedules extracted from the whole-day timetables are compared with the
optimal schedules previously solved with CPLEX, as presented in Tables 11 and 12. The tables
indicate that the extracted schedules are similar to the optimal ones in general, with some minor
shifts in departure times.
The problem size increases from around 100 integer and binary variables for the peak-period
problem to 850 integer and binary variables for the whole-day problem. Yet the computation

Table 10
Computation results for the whole-day problem
Scenario I Scenario II
Objective value 136,961 129,356
Fleet deployment 5 5
Average passenger waiting time (min) 14.5 9.5
Revenue 122,199 122,199
Percentage of multi-stop trip users N/A 31.50%
Computation time (s) 21 41

Table 11
The ferry schedule from CPLEX and the heuristic solved for the whole day (scenario I)
Scenario I (CPLEX) Scenario I (Heuristic)a
CBD Peng Chau CBD Peng Chau
8:15 a.m. 7:15 a.m. 7:45 a.m. 7:00 a.m.
9:00 a.m. 7:30 a.m. 8:30 a.m. 7:15 a.m.
8:00 a.m. 9:15 a.m. 7:45 a.m.
9:00 a.m. 8:30 a.m.
CBD Mui Wo CBD Mui Wo
8:00 a.m. 7:30 a.m. 8:00 a.m. 7:30 a.m.
8:30 a.m. 8:00 a.m. 9:00 a.m. 8:00 a.m.
9:00 a.m. 9:00 a.m. 9:00 a.m.
a
Extracted from the whole-day solution.
 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328 327

Table 12
The ferry schedule from CPLEX and the heuristic solved for the whole day (scenario II)
Scenario II (CPLEX) Scenario II (heuristic algorithm)a
CBD Peng Chau CBD Peng Chau
b
8:15 a.m. 7:15 a.m. 8:15 a.m. 7:15 a.m.b
9:00 a.m.c 7:30 a.m. 9:00 a.m.c 7:30 a.m.b
8:00 a.m.b 7:45 a.m.b
8:30 a.m.b 8:30 a.m.b
9:00 a.m. 9:00 a.m.
CBD Mui Wo CBD Mui Wo
b
8:00 a.m. 7:00 a.m. 8:00 a.m. 7:00 a.m.b
8:30 a.m. 7:30 a.m. 8:30 a.m. 7:15 a.m.b
9:00 a.m.c 7:45 a.m.b 9:00 a.m.c 7:30 a.m.b
8:15 a.m.b 8:00 a.m.
9:00 a.m. 8:15 a.m.b
9:00 a.m.
a
Extracted from the whole-day solution.
b
From Mui Wo via Peng Chau to CBD.
c
From CBD via Peng Chau to Mui Wo.

times increase only from 11 to 21 s for scenario I with direct services, and from 13 to 41 s for
scenario II with both direct and multi-stop services. This clearly demonstrates the advantage of
the polynomial-time performance of the heuristic algorithm.

5. Concluding remarks

The study formulated the ferry network design problem by considering the optimal fleet size,
routing, and scheduling for both direct and multi-stop services. The objective function combines
the operator and passengersÕ performance measures. Mathematically, the model is formulated as a
mixed integer multiple OD network flow problem with capacity constraints. The heuristic algo-
rithm, which differs from the conventional relaxation approach, was developed to efficiently solve
the problem. The ferry flow variables are determined by exploiting the polynomial-time perfor-
mance of shortest path algorithms on the network of service arc contributions, which are
themselves estimated by the passenger-loading scheme to avoid repeated linear program solutions.
Two examples of ferry services in Hong Kong were examined to demonstrate the performance
of the heuristic algorithm. The results of the numerical studies showed that the heuristic produced
solutions that were within 1.3% from the CPLEX optimal solutions, whereas the computational
times were around 0.5% of CPLEXÕs. For the whole-day problem that is beyond the capability of
CPLEX, the heuristic obtained a solution in tens of seconds. These results demonstrate the effi-
ciency in solving problems of practical size and its ability to solve problems in polynomial times.
In this study, we consider that the demands for each OD pair are fixed and their arrival patterns
given. This assumption is reasonable for ferry services with limited competitive modes in the short
run, such as ferry services serving the outlying islands, where waterborne transport is the primary
transport mode for commuters. These assumptions, though somewhat restrictive, simplify the
328 M.F. Lai, H.K. Lo / Transportation Research Part A 38 (2004) 305–328

analysis substantially. In the long run, people might change their residence and job locations,
rendering demand elastic. Travelers may also change their demand patterns and departure times
in response to the service schedule and their desired arrival times at destinations. So there are
interactions between the supply and demand patterns that need to be considered. In addition,
operators may offer differential services, fast-expensive versus slow-inexpensive, to accommodate
the different market segments. These extensions would require modeling the service preferences
and utilities as perceived by travelers, the fare structure between the different services, and their
respective service schedules and travel times, making the formulations non-linear and more
complex. Our current research focus is on addressing these extensions within the framework
developed herein. By drawing upon the heuristic procedure developed in this study, we solve these
more complex formulations through a sequence of relaxed sub-problems. This is outside the scope
of this paper. We hope to be able to report these results in a future study.

Acknowledgement

The helpful comments of the anonymous referee are gratefully acknowledged. This study is
sponsored by the Competitive Earmarked Research Grants, HKUST6083/00E and HKUST6161/
02E, of the Hong Kong Research Grant Council.

References

Armacost, A.P., Barnhart, C., Ware, K.A., 2002. Composite variable formulations for express shipment service
network design. Transportation Science 36, 1–20.
Ceder, A., Wilson, N., 1986. Bus network design. Transportation Research 20B, 331–344.
Crainic, T.G., 2000. Service network design in freight transportation. European Journal of Operational Research 122,
272–288.
Crainic, T.G., Laporte, G., 1997. Planning models for freight transportation. European Journal of Operational
Research 97, 409–438.
Crainic, T.G., Frangioni, A., Gendron, B., 2001. Bundle-based relaxation methods for multi-commodity capacitated
fixed charge network design problems. Discrete Applied Mathematics 112, 73–99.
Farvolden, J.M., Powell, W.B., 1994. Subgradient methods for the service network design problem. Transportation
Science 28, 256–272.
Gorman, M.F., 1998. An application of genetic and tabu searches to the freight railroad operating plan problem.
Annals of Operations Research 78, 51–69.
ILOG, 1998. CPLEX 6.0 UserÕs Manual. ILOG, Inc., Incline Village, NV.
Kim, D., Barnhart, C., Ware, K., Reinhardt, G., 1999. Multimodal express package delivery: a service network design
application. Transportation Science 33, 391–407.
Magnanti, T.L., Wong, R.T., 1984. Network design and transportation planning: models and algorithms.
Transportation Science 18, 1–55.
Transport Department, 1999. Technical Report of Third Comprehensive Transport Study (CTS-3), Hong Kong.
Yan, S., Chen, H.L., 2002. A scheduling model and a solution algorithm for inter-city bus carriers. Transportation
Research A 36, 805–825.
Yan, S., Tseng, C.H., 2002. A passenger demand model for airline flight scheduling and fleet routing. Computers &
Operations Research 29, 1559–1581.
Yan, S., Young, H.F., 1996. A decision support framework for multi-fleet routing and multi-stop flight scheduling.
Transportation Research A 30, 379–398.