Computers & Operations Research 47 (2014) 27–41

Contents lists available at ScienceDirect

Computers & Operations Research

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

A Stackelberg hub arc location model for a competitive environment

Mihiro Sasaki a, James F. Campbell b,n, Mohan Krishnamoorthy c,e, Andreas T. Ernst d
a
Department of Information Systems and Mathematical Sciences, Nanzan University, Japan
b
College of Business Administration, University of Missouri – St. Louis, USA
c
IITB-Monash Research Academy, IIT Bombay, Powai, Mumbai 400076, India
d
CSIRO Mathematical and Information Sciences, Australia
e
Department of Mechanical and Aerospace Engineering, Clayton, VIC 3800, Australia

art ic l e i nf o a b s t r a c t

Available online 4 February 2014 In this paper, we consider the design of large-scale multiple allocation hub-and-spoke transportation
Keywords: networks in a competitive environment. We adopt a generic hub arc location model that locates arcs
Hub location with discounted transport costs connecting pairs of hub facilities. Two firms compete for customers in a
Competitive location Stackelberg framework where the leader firm locates hub arcs to maximize its revenue, given that the
Stackelberg competition follower firm will subsequently locate its own hub arcs to maximize its own revenue. We present an
optimal solution algorithm that allocates traffic between the two firms based on the relative utility of
travel via the competing hub networks. Results for each competing firm with up to three hub arcs show
the important role of competition in designing hub-based transportation systems.
& 2014 Elsevier Ltd. All rights reserved.

1. Introduction carriers compete throughout a geographic region. This competition is

Hub-and-spoke networks play an important role in many trans-
portation systems. These networks provide efficient transportation
between many origins and destinations (e.g., cities) via a set of hubs
that serve as switching and flow consolidation points, hub arcs that
connect two hubs with a discounted travel cost, and access arcs that
connect the non-hub nodes and hubs. Hub networks use fewer arcs
than in a point-to-point network and thus can reduce transportation
costs by exploiting the economies of scale from consolidated flows.
Studies on various hub location models have attracted much attention
since O'Kelly [45]. Reviews of hub location research include [6,13,16].
Recent research has extended solutions to larger problems [22,23,53]
and addressed multiple capacity levels [24], service considerations
[11,17,54], dynamic location [20] and stochasticity [21,54]. Some
illustrative applications of the wide range of settings for hub location
models are LTL (less-than-truckload) trucking [25], truckload trucking
[56], high speed rail [7], postal operations [29], liner shipping [31,32],
and airlines ([45] is one of the earliest examples).
The vast majority of hub location research has been directed at
finding an optimal (or near-optimal) hub network for a single firm to
serve a given set of demand specified as flows between many origins
and destinations. However, real-world hub-based transportation sys-
tems typically operate in a competitive environment where several
n
Corresponding author.
E-mail address: campbell@umsl.edu (J.F. Campbell).
likely to influence the optimal hub locations and hub network design.
With competition, the customers (e.g., freight shippers or individual
travelers) must decide which competing carrier(s) to use, and this is
typically based on the relative level of service provided and the cost
(or fare) charged. Thus, competitive hub models require designing hub
networks for each competitor and allocating the demand among the
competitors. The objective is usually to maximize the market share
captured, where market share may be measured in terms of the
percentage of the passengers, freight, revenue, or profit captured.
Reviews of competitive location research for general (non-hub) net-
works include [26,27,52].
Although a variety of hub location models have been studied in
the last two decades, studies on competitive hub location pro-
blems are not very common. The earliest work is Marianov et al.
[43], which formulated a sequential competitive hub median
problem on a network. This model assumes that one firm locates
p hubs optimally (as in a multiple allocation p-hub median
problem [13]) and then the second firm locates p hubs, given the
locations of the first firm's hubs, to maximize the flow captured.
Wagner [57] provides improved formulations and results for the
problems presented in [43] with optimal solutions for up to 50
nodes and 5 hubs. These works allocate customers between the
two firms based on the relative costs of the OD paths and include
both a binary “all-or-nothing” allocation, where all passengers for
each OD pair are allocated to whichever firm provides the lowest
cost OD path (with ties being allocated to the first firm), and a five-
level fractional allocation, where each firm captures 0%, 25%, 50%,

http://dx.doi.org/10.1016/j.cor.2014.01.009
0305-0548 & 2014 Elsevier Ltd. All rights reserved.
28 M. Sasaki et al. / Computers & Operations Research 47 (2014) 27–41
75% or 100% of the passengers, depending on the relative OD path
costs for the two competitors. Eiselt and Marianov [28] extend this
line of research by replacing the discrete passenger allocation
mechanism with a continuous proportional allocation based on
the relative travel time and travel cost of the OD paths of the
competitors. They provide solutions with up to five hubs for each
competitor using heuristic procedures where the first of the two
competing firms locates its hubs either at random or to provide
the optimal p-hub median network. More recently, Luer-Villagra
and Marianov [39] use a logit function to model competition in
hub networks, where the incumbent firm has an optimal network
to serve all demand. The new entrant designs a network to
maximize its profit, but the incumbent does not act preemptively
or react. Allocation of customers is based on price alone using a
logit model and all demand is assumed to be served. Solutions are
found using an “ad-hoc metaheuristic” based on a genetic algo-
rithm. This work incorporates both price determination and net-
work design; however it uses a heuristic solution procedure and is
a sequential location model (as are [28,43,57]) in considering hub
location and network design for the new entrant alone.
In contrast to the research mentioned above, we consider the
Stackelberg problem and employ an optimal algorithm. Further, we
employ the hub arc model rather than assuming the hub-level
network is fully connected. The Stackelberg problem is relevant
when competitors are aware of each other and one firm (the leader)
locates its hubs in anticipation of another firm (the follower)
optimally locating its hubs, based on the known locations of the
leader's hubs. Thus, the leader seeks to locate its hubs so that its
objective is optimized after the follower best locates its hubs. The
Stackelberg hub location problem is analogous to the Stackelberg
location problem on a network introduced by Hakimi [36] and used
in several other studies of non-hub facility location (e.g., [47,51]).
Sasaki and Fukushima [50] present a continuous Stackelberg hub
location model where passenger allocations are determined by a
logit function. The results showed that the leader firm may suffer
heavy losses if it neglects to consider the competitor's strategies.
Sasaki [46] considers a discrete Stackelberg hub location model
with flow threshold constraints to ensure that a firm does not carry
an unrealistically low level of flow for any OD pair. The model is
formulated as a bilevel programming problem where the upper
(leader) and lower (follower) problems are binary integer programs.
Both [48] and [50] limit OD paths to a single hub stop, so there is no
discounted inter-hub travel (i.e., no hub arcs).
Another relevant stream of research is on competitive airline
network design models that also consider hub location. The airline
research generally employs logit functions to determine market
shares for the competitors in idealized hub networks. Adler [1,2]
uses a common market share model based on a utility function
that incorporates flight frequencies on the least frequent leg of a
trip, airfares, whether the trip is direct or not, passenger type
(business or non-business), and the frequency elasticity of demand
for each type of traveler. Adler and Smilowitz [3] use a logit
function with a slightly different form of utility that includes the
number of flight legs in a trip. These models incorporate detailed
cost functions (see [55]), and sometimes consider profit maximi-
zation, so that not all demand is necessarily served. These works
differ from the hub location research that focus on locating hubs
and designing networks for more generic transportation systems
with a simple distance based cost function. In contrast, the airline
specific research may address locating a single hub (e.g., an
intercontinental gateway airport), with a focus on issues such as
airline alliances and mergers [3] and intercontinental service
frequencies [41,42], more than hub network design. An alternate
game theoretic approach is used by Lin and Lee [38] to solve small
competitive problems with up to three candidate hubs by enu-
merating all hub combinations.
In this paper, we present a general discrete Stackelberg hub
location problem using the multiple allocation hub arc location
model [14,15] that locates hub arcs whose endpoints are hub
nodes rather than the hub median model that locates fully
connected hub nodes as in [43,57]. The hub arc model allows OD
paths with one or two stops at hubs and helps concentrate flows
on the discounted hub arcs, by relaxing the restriction in hub
median models that every flow between two hubs is discounted,
whether or not it is warranted by the level of flow. This is
important to better match the model for economies of scale with
actual flows in the solutions [12]. Furthermore, in our model each
firm seeks to maximize the revenue (not traffic) captured and we
employ a customer allocation mechanism based on the relative
utility of the OD paths for each competitor, introduced in [43], to
model the different customer behaviors in selecting between the
competing carriers. Unlike other models that implicitly treat all
customers (e.g., shippers) equivalently by maximizing the traffic
captured, in our model customers who generate higher revenues
are more valuable. We assume the same revenue applies to each
competitor for each OD pair, which may be realistic in the long run
due to competitive pressures. This allows us to focus on competi-
tion from the different levels of service offered by the OD paths
through the hub networks rather than having the conflicting
interactions from price competition. This can be a strong assump-
tion in some contexts, such as for passenger airlines which use
extremely complex and dynamic pricing policies [18]. However,
our goal is not to model a particular network or particular pricing
strategies in great detail, but rather to analyze a more general
model for hub-based transportation, so as to develop general
insights regarding how competition affects the optimal hub loca-
tions and network design.
The remainder of this paper is organized as follows. In Section 2,
we provide some background and a formulation of the model.
Section 3 describes the solution algorithm and Section 4 includes
computational results. In Section 5, we give concluding remarks and
mention some areas for future work.
2. Model description
In this paper we consider the case when the firms' hub sets are
disjoint, so they do not share any hubs or hub arcs (as in [41,42]).
This seems most realistic for situations where capacity at a hub
(e.g., airport) limits operations to a single firm, as for air express
(e.g., FedEx and UPS). It can also be a factor for air passenger
transportation as the scale of operations at a large airline's hub
(e.g., Atlanta for Delta) tends to limit opportunities for other large
airlines to have the high number of operations needed at a hub.
However, for trucking it is entirely reasonable to have truck
terminals for competitors located in the same city, as the public
infrastructure (roads) is usually not a constraint. Note that when
competitors are allowed to share hubs, then the follower can
capture at least 50% of the market by using the same hub arcs as
the leader. Therefore, the more interesting cases are likely to be
when the number of hub arcs and/or hubs differ between the
leader and follower. Our model and solution algorithm could easily
by modified to handle co-located hubs and hub arcs, but we leave
these situations for future research. OD paths for our hub arc
location model are limited to three arcs, where the first arc is
for collection from the origin to a hub, the second one is a central
hub arc for transfer between two hubs, and the last arc is for
distribution from a hub to the destination. Each of these may be a
degenerate arc (from a node to itself) if the node is a hub.
Fig. 1 shows three possible multiple allocation hub networks
that serve flows among seven origin/destination nodes. Fig. 1
(a) shows a 3-hub median solution with hubs at nodes 1, 2 and
 M. Sasaki et al. / Computers & Operations Research 47 (2014) 27–41
Fig. 1. Hub arc networks: (a) three hub median model, (b) non-competitive model with two hub arcs, and (c) competitive model with one hub arc for each firm.
4 (shown as squares) and three hub arcs (in bold) connecting the
hubs. This is the type of model analyzed in previous competitive
hub location research [28,43,57] where each firm uses a fully
connected network of hub arcs. Fig. 1(b) shows a 2-hub arc
solution (for one firm) with hubs at nodes 1, 2, 3 and 4 connected
by two hub arcs between nodes 1 and 4 and nodes 2 and 3.
Because each OD pair is joined by a path with at most three arcs
where the central arc, if it exists, is a hub arc, the path from node
6 to node 3 is 6–1–4–3, not 6–1–2–3. Fig. 1(c) shows a competitive
hub arc solution where Firms A and B each locate one hub arc.
Each firm provides service for all OD pairs via one-stop or two-
stop routes through its own hubs. Thus, the path from node 6 to
node 3 for Firm A is 6–1–4–3 and for Firm B is 6–2–3. Similarly, the
path from node 4 to node 5 for Firm A is 4–5 and for Firm B is 4–3–
2–5, since each firm routes every flow via at least one of their
hubs, and thus the direct connection 4–5 is not allowed for Firm B.
2.1. Notation
Let Firm A be the leader firm and Firm B be the follower firm.
We employ the following notation:
V, the set of demand nodes, |V | ¼n.
Wij, the trip demand between nodes iA V and jA V (j4 i).
Fij, the revenue (e.g., airfare) per unit demand between nodes
iA V and j AV (j4 i).
dij, the distance between nodes iA V and jA V,j4 i; dii ¼0 (iA V).
α, the discount factor for transfer on hub arcs to reflect
economies of scale.
χ, the unit transportation cost for collection (from origin
to hub).
δ, the unit transportation cost for distribution (from hub to
destination).
Cijkl, the unit cost for a path from an origin iA V to destination
jA V (j4i) through hubs k A V and l A V, in this order.
qA, the number of Firm A's hub arcs.
qB, the number of Firm B's hub arcs.
A, a set of qA hub arcs used by Firm A.
ℬ, a set of qB hub arcs used by Firm B, where Firms A and B
share no hubs.
ℋA , the set of combinations of qA hub arcs that can be used by
Firm A.
ℋB , the set of combinations of qB hub arcs that can be used by
Firm B.
We assume that the trip demand is symmetric. Therefore, we
define Wij and Fij for all j4i for each i. Note that C ijkl ¼
χ dik þ αdkl þ δdlj and C ijkk ¼ χ dik þ δdkj for one-stop paths as dkk ¼ 0
for all k A V. By introducing C Aijkl and C Bijkl for each firm with specific
α,χ,δ, we can easily employ a different cost function for each firm.
However, in this paper we assume that both firms employ the
same cost function. Note that hub arcs are assumed to be
29
undirected and no degenerate hub arc, i.e., (k,k), is allowed to be
established. Because each firm provides service for each OD pair
via a single (lowest cost) path, the competition is between two
paths.
2.2. Customer allocation function
Hub networks provide an cost incentive to route flows along
circuitous paths using lower cost hub arcs, especially for small
values of α. While these circuitous paths are often desirable from
the carrier's perspective to reduce their own costs via consolida-
tion opportunities at hubs, they are often undesirable from the
customer's perspective due to the longer OD travel times. In our
competitive model, the given demand for each OD pair is divided
between Firms A and B using a customer allocation function.
Previous competitive hub location models have allocated custo-
mers between competitors based on the relative path costs [43,57],
path costs and travel times [28], or price alone [39]. The allocation
function should reflect the customers' preferences and will likely
differ among the many applications. For freight transportation hub
networks, the customer is the freight shipper, who is not particu-
larly concerned with the length of the OD path the freight follows,
as long as the freight is delivered when scheduled (e.g., next-day
by 10:30 am). On-time delivery could occur via a variety of OD
paths, and a freight carrier has flexibility to route freight via a
circuitous, but lower cost, path that exploits the reduced travel
costs on hub arcs. Because in our model the customer pays the
same cost (i.e., carrier revenue) with both competing carriers for a
given shipment (i.e., OD pair), the freight carrier can choose to
route freight along the minimum cost path via the hub network, as
long as this provides adequate service.
For transportation of high service level freight (on the ground
or through the air) there is a need for more direct routings due to
the high value of time. Airline passengers have a similar desire for
more direct routings, though many factors can be important when
selecting among competing carriers, including flight frequencies,
flight paths, departure and arrival times, loyalty programs, and
ease of access to airports; furthermore, the importance of different
factors varies among customers (see for example, [4,8,9,30,37,44]).
In the airline competition literature, there is a great deal of
research on mechanisms for allocating demand among competing
airlines, as well as on modeling costs. Logit functions are popular
in airline analyses [1–3,5], though other approaches are used. For
example, Refs. [34, 35] consider allocation of demand using a
multinomial logit model, an artificial neural network model, and a
custom model based on the ratio of attraction functions (them-
selves based on linear combinations of trip attributes). Results
showed that their custom model outperformed the other two
approaches. In competitive hub network-oriented research, mar-
ket share has been determined with a gravity model [28] and a
step function [32]. For some further discussion on complexities
30 M. Sasaki et al. / Computers & Operations Research 47 (2014) 27–41

of discrete choice modeling for airlines, including a multinomial for all OD pairs, so all demand is transported. This can be a strong
probit model, see [33]. assumption and alternative models, especially to maximize profit,
Due to the strategic nature of hub location models and our often allow some demand to not be served (when not profitable).
desire to illuminate the role of competition on hub location and Note that the allocation functions φAij and φBij depend on the
network design, we do not address detailed pricing or scheduling
distance or cost ratio (DRA;ℬ and CRA;ℬ ), which depends on the
issues, such as flight frequencies or departure and arrival times, ij ij
A ℬ
and instead we focus on the cost and service provided by the OD path distances or costs ðDA ℬ
ij and Dij or C ij and C ij Þ, which in turn
paths in the hub network. The OD path length can be viewed as a depends on the decision variables xAijkl and xBijkl .
proxy for the OD travel time, given that in our model both firms
To link the cost or distance ratios CRA;ℬ and DRA;ℬ to the
have the same revenues for each OD pair. To include both service- ij ij

focused and cost-focused customers, we consider one customer fraction of flow captured, we adopt a five-level step function
allocation function based on the relative differences in OD travel (introduced in [43]) as shown in Table 1. Alternative allocation
distances, and a second allocation function based on the relative mechanisms could be used, such as a logit function (see for
differences in OD travel costs. Further refinements to include the example [1–3,39]), and our step function can be viewed as a
number of hub stops or total travel time including airport time, as discrete approximation of a logit function. (While a logit function
in [28], are an area of future research. has some appealing properties, its nonlinear structure can be
Because customers for a given OD pair do not in general all challenging to incorporate in mathematical programming solu-
select the same service, we model customer allocation as a tions approaches, as discussed below). The allocation step function
function of the relative utility of the OD paths in Firm A's and assigns a fraction of the demand for each OD pair to each
Firm B's hub network. We first define C A competitor based on the values r1 and r2, where r1 Zr2 Z0. This
ij as the minimum cost for
a trip from origin i to destination j using hub arc set A for Firm A: step function essentially allocates the customers equally (50% to
each firm) when the path costs or distances are similar (between
CA
ij ¼ min min fC ijkl ; C ijlk ; C ijkk ; C ijll g: ð1Þ  r2 and r2), allocates three times as many customers to the firm
ðk;lÞ A A
with better performance when the absolute value of the cost or
We define C ℬ
ij analogously for Firm B:
distance ratio is between r2 and r1, and allocates all customers to
the firm with better performance when the absolute value of the
Cℬ
ij ¼ min min fC ijkl ; C ijlk ; C ijkk ; C ijll g: ð2Þ cost or distance ratio exceeds r1. By adjusting the values of r1 and
ðk;lÞ A ℬ
r2, a range of different customer allocation schemes can be
Similarly, let DA ℬ
ij and Dij be the distance of the path that evaluated. In the extreme case of r1 ¼r2 ¼0 the firm offering the
corresponds to the minimum cost path from i to j for Firms A better (lower distance or cost) OD path will capture all the
and B, respectively. Thus, if the minimum cost path from i to j for demand. If the paths for Firm A and Firm B have equal lengths
Firm A with hub arc set A is i–k–l–j, then C A ij ¼ χ dik þ αdkl þ (or costs), then DRA;ℬ ¼ 0 (or CRA;ℬ ¼ 0) and we assume the
δdlj and DAij ¼ dik þ dkl þ dlj . We now define the customer allocation ij ij
demand is split equally between the two firms. (This is slightly
distance ratio DRA;ℬij as
different than in [43,55] where in the event of equal paths, all
DA ℬ
ij  Dij
demand is allocated to a single competitor.)
DRA;ℬ ¼ : ð3Þ The Stackelberg hub arc location problem can be approached as
DA ℬ
ij
ij þ Dij
a bilevel programming problem [19] in which each firm locates a
We define an analogous customer allocation cost ratio CRA;ℬ
ij as given number of its own hub arcs to maximize the revenue it
captures. In this view, the upper level problem is to locate the
CA ℬ
ij  C ij leader's hub arcs to maximize the revenue it captures knowing the
CRA;ℬ ¼ : ð4Þ
CA ℬ
ij follower will subsequently locate its hub arcs. The lower level
ij þ C ij
problem is to locate the follower's hub arcs to maximize the
Each of these is the ratio of the difference in path distances (or revenue it captures, given the locations of the leader's hub arcs
costs) to the sum of the path distances (or costs). Both and the condition that the follower cannot use any hubs also used
CRA;ℬ
ij and DRA;ℬ
ij range from  1 to þ1 and represent the relative by the leader. Appendix A provides a bilevel programming
advantage of Firm B over Firm A. A ratio of zero indicates formulation for the Stackelberg hub arc location problem with
equivalent distances or costs for the paths with Firm A and Firm the customer allocation cost ratio. Given the difficulty of solving
B. A large ratio (close to þ1) indicates that the best path via Firm single firm (non-competitive) hub arc location problems, and the
B's hub network is much preferred (shorter or less expensive) over interdependent nature of the leader and follower's problems, we
the best path via Firm A's hub network. do not solve the bilevel programming formulation directly. Rather,
To allocate customers among the competitors, we first define we develop an efficient optimal algorithm that uses two bounding
the path decision variables for Firms A and B are as follows: procedures within an enumeration scheme.

xAijkl ¼ 1 if the lowest cost path for OD pair i,j with Firm A uses
hub arc (k,l) from k to l, kA V and lA V, and 0 otherwise.
xBijkl ¼ 1 if the lowest cost path for OD pair i,j with Firm B uses
hub arc (k,l) from k to l, kA V and lA V, and 0 otherwise.
Table 1
The fraction of flow captured by Firm A.
Now we can define φAij ð½xA ; ½xB Þ and φBij ð½xA ; ½xB Þ as the fraction
CRA;ℬ
ij or DRA;ℬ
ij
φAij ð½xA ; ½xB Þ (%)
of demand between origin i and destination j captured by Firm A
and Firm B, respectively. We use the [  ] notation throughout the r  r1 100
paper to indicate vectors, so [xA] and [xB] are vectors whose  r1 to  r2 75
elements are xAijkl and xBijkl for all i, j, k, l, respectively. We set  r2 to r2 50
r2 to r1 25
φAij ð½xA ; ½xB Þþ φBij ð½xA ; ½xB Þ ¼ 1 ð5Þ 4r1 0
 M. Sasaki et al. / Computers & Operations Research 47 (2014) 27–41 31

3. Algorithm (d) If W  g A o g B ; then go to Step 2(f). Otherwise, go to

Step 2(c).
Single firm non-competitive hub arc location problems have (e) If g B 4g B ; then set g B ≔g B :
n
been solved in [15] using an optimal algorithm that employs upper (f) If g A o g A ; then set g A ≔g A : If g A o f ; go to Step 1. Otherwise
and lower bounds to eliminate poor solutions (i.e., sets of hub arcs) go to Step 2(a).
n n
quickly. In this paper, we extend that algorithm to competitive hub Step 3: If f og A ; then set f ≔g A ; An ≔A; ℬn ≔ℬ. Go to Step 1.
arc location problems. The key to the algorithm is to use bounding Step 4: The optimal hub arcs for Firms A and B are An and ℬn ,
techniques to avoid enumerating all sets of Firm B's hub arc respectively. The optimal objective value of Firms A and B are
combinations and to avoid evaluating the cost for all OD pairs. fn and W  fn, respectively.
The remainder of this section defines our optimal solution algo-
rithm for the Stackelberg hub arc location problem. In Step 1, we select a set of qA hub arcs for Firm A (the leader)
Recall that ℋA and ℋB are the sets of all possible combinations and then remove this from the set of feasible hub arc combinations
A
of qA and qB hub arcs for Firms A and B, respectively. Let f ðA; ℬÞ for Firm A. Then for Firm B (the follower), we create the set of all
A B
denote the revenue of Firm A, where A A ℋ and ℬ A ℋ are the combinations of qB hub arcs that do not involve any of the hubs
hub arcs employed by Firm A and Firm B, respectively. In the same used in the qA hub arcs for the leader. (To allow use of the same
B
manner, let f ðA; ℬÞ denotes Firm B's revenue with given A and ℬ. hub by Firm A and B, we would not need to restrict the set ℋBA to
n
Let ℬA be the optimal set of hub arcs for Firm B for a given set of exclude Firm A's hubs.) The goal of Step 2 is to find quickly the best
B
Firm A's hub arcs: ℬnA ¼ argmaxℬ f ðA; ℬÞ. Note that solution for Firm B for a given set of hub arcs for Firm A. This is
B A done by considering each set of qB hub arcs for the follower, one at
f ðA; ℬnA Þ ¼ W  max f ðA; ℬÞ; ð6Þ
ℬ A ℋB a time and aggregating the revenue for each firm for the OD pairs,
one at time (in Step 2(c)). However, we may not need to consider
where W is the total revenue for both firms: ∑i A V ∑j 4 i W ij F ij . The all OD pairs because of the bounding test in Step 2(d). Note that
idea behind the upper bounding procedure is as follows: For a W gA in Step 2(d) is an upper bound on Firm B's revenue. Also, we
given set of Firm A's hub arcs, denoted A,^ if we find some set of
may not need to enumerate all sets of qB hub arcs for Firm B
Firm B's hub arcs, say denoted ℬ, ^ ℬÞ
^ such that f A ðA; ^ is worse (i.e.,
because of the bounding test in Step 2(f). Step 2(f) is very
smaller) than a known solution for Firm A, then we need not important in recognizing that as soon as Firm B finds a solution
consider any other sets of Firm B's hub arcs with A.^ More formally
that limits Firm A to a solution worse than Firm A's current
as follows: solution, no further analysis for this set of qA hub arcs is needed.
Step 3 is reached only in the situation when the bounding test in
Proposition. Let fn be the best know value for the revenue captured
^ A ℋA . Step 2(f) was not able to end the enumeration of Firm B's hub arcs
by Firm A. Consider a given set of Firm A's hub arcs, denoted A
A ^ n B A ^ n n early. This step updates the new best known solution for Firm A
If f ðA; ℬÞ o f ; for some ℬ A ℋ , then f ðA; ℬA^ Þ o f .
and returns to Step 1 to consider the next set of qA hub arcs for the
Proof. Assume that ℬ ^ A ℋB satisfies the above inequality, leader (Firm A).
A ^ ^ n A ^ ^ B ^ ^ To make the SHAL algorithm efficient, we wish to select the hub
i.e., f ðA; ℬÞ o f , holds. Since f ðA; ℬÞ þ f ðA; ℬÞ ¼ W , then
B ^ ^
arc combinations in Steps 1 and 2(a) in a manner that is likely to
W  f ðA;
n
ℬÞ o f . Hence,
A ^ ℬn Þ ¼ W  f ðA;
f ðA; ^ ℬn Þ oW f
B B
^
A ^
A produce good (high revenue) solutions quickly. Based on experi-
^ ℬÞ
ðA; ^ o f . □
n
mentation over a wide range of problems and prior results for a
similar algorithm in [15], we select the hub arc combinations
To encourage early cutting in the algorithm, we seek to A and ℬ in Steps 1 and 2(a) by first sorting the potential hub arcs
enumerate hub arc combinations for Firm B in a sequence that is in decreasing order of OD demand  revenue (Wij  Fij), and then
likely to quickly find very good (high revenue) solutions for Firm B. selecting A and ℬ from the sorted lists as described in [15]. To
Similarly, we seek to enumerate the hub arc combinations for Firm further increase efficiency via the bounding test in Step (d), we
A in a sequence that is likely to quickly find very good (high select the OD pairs in Step 2(c) in a manner that is likely to
revenue) solutions for Firm A. The algorithm also uses upper produce good (high revenue) solutions quickly. Again, based on
bounds to avoid enumerating all OD pairs when evaluating a prior results in [15], we select the OD pairs in decreasing order of
candidate set of hub arcs for Firms A and B, as shown below in Step OD demand  distance (Wij  dij).
2. In the algorithm, g A is an upper bound for Firm A's revenue and
g B is used as the best known solution for Firm B with a given A.
The competitive Stackelberg hub arc location optimal algorithm, 4. Results
denoted SHAL, is as follows.
This section reports computational results from the optimal
SHAL algorithm. solutions for 360 problem instances using the standard CAB hub
location data set of flows between 25 major cities in the US. As is
n
Step 0: Set f ≔1; A ¼ ∅ and ℬn ¼ ∅. common with the CAB data, we set χ ¼ δ ¼1 and vary α from 0.2 to
Step 1: If ℋA is empty, go to Step 4. Select A A ℋA . Set 1.0. These problems were solved using the optimal SHAL algorithm
ℋA ≔ℋA n fAg and set g A ≔1. Create ℋBA as the set of all hub coded in Cþ þ on a DELL OPTIPLEX GX620 computer with a
arc combinations from ℋA that do not include a hub node 3.4 GHz Intel Pentium 4 processor operated under Windows XP
used in A. Professional with 2.0 GB DDR2-SDRAM memory. We set the
(a) candidate sets of hub arcs for Firms A and B to be the entire set
Step 2: of arcs connecting all pairs of the 25 cities, so there were 300
(a) If ℋBA is empty, then go to Step 3. Select ℬ A ℋBA . Set candidate hub arcs for each firm.
ℋBA ≔ℋBA n fℬg:
(b) g A ≔0; g B ≔0; g B ≔0; and Π is the set of all OD pairs {i,j}, 4.1. Problem scenarios
where j4i.
(c) If Π ¼∅, then go to Step 2(e). Select OD pair {i,j}A Π, Π: ¼ Π To provide general results for a range of transportation systems,
\{i,j}. gA ≔g A þ F ij W ij φAij ð½xA ; ½xB Þ: gB ≔g B þ F ij W ij φBij ð½xA ; ½xB Þ: we utilize a variety of scenarios with different OD revenues and
32 M. Sasaki et al. / Computers & Operations Research 47 (2014) 27–41

customer behaviors. For the OD revenues, we prepared two data Demand  Price, and Demand  Distance, originating and terminating
sets, denoted Price and Distance. Revenue set Price is based on IATA at each city. Column 1 is the rank of a city in terms of total demand
Y class standard airfares between the city pairs collected by the and column 2 is the identification number of the city (alphabetical).
lead author in 1998 from 〈http://www.airfare.com/〉. For each For example, the “largest” city is 17 (New York) has 17% of the total
OD pair, the lowest Y class airfare was selected from the many originating and terminating demand, 16.9% of the originating and
fares provided. The Price data set represents realistic revenues at a terminating demand weighted by the Price revenues, and 16.3% of
particular point in time. (Of course, the long run patterns of the originating and terminating demand weighted by the Distance
airfares may differ due to the dynamic nature of the industry revenues. One effect of the Distance revenue set is to effectively
and its pricing structures.) Details on the characteristics of the increase the importance of the west coast cities due to the longer
Price revenue set are in [49]. A second revenue set, denoted average travel distances from the west coast to the majority of
Distance, is the direct OD distance for each city pair. This can demand in the eastern US; for example, Los Angeles (city 12) and San
represent a situation where revenues are strongly correlated with Francisco (city 22) are the third and seventh ranked cities, respec-
OD distance, as is the case for aircraft operating costs [3,53]. Using tively, in terms of demand, but they rank second and third for
the Price and Distance revenue sets, our objective is to maximize Demand  Distance. Correspondingly, centrally located Chicago (city
the total revenue captured. 4) is ranked second in terms of demand, but fourth for Demand  -
Each OD pair has a total revenue of Wij  Fij to be allocated. Thus, Distance. The Price revenue set does not provide such large changes,
the importance of a node is influenced by both its originating and except for city 4, so the Price and Distance revenue sets provide rather
terminating demand, and the associated revenues. Table 2 lists different incentives for the optimal networks.
the 25 CAB cities with their percentages of the total Demand, Our computational results include three different settings for r1
and r2 in the customer allocation function to explore a range
of customer behaviors in selecting between competing firms.
Table 2 Our base case, denoted low customer selectivity, corresponds to
Cities in the CAB data set. r1 ¼ 0.75 and r2 ¼0.25 and reflects customers that are rather
insensitive to the differences between the competitors. With these
Rank ID City % of Total % of Total % of Total
large values of r1 and r2, a very large relative difference in distance
# Demand Demand  Price Demand  Distance
(or cost) is needed to strongly prefer one firm over the other. For
1 17 New York 17.0 16.9 16.3 example, for Firm B to capture all revenues, we need DRA;ℬ ij 4 0:75
2 4 Chicago 10.0 8.0 8.5 or ðCRA;ℬ
ij 4 0:75Þ which implies from Eq. (3) (Eq. (4)) that Firm B
3 12 Los Angeles 7.3 7.2 12.5 would need to provide a travel distance (or cost) of less than one-
4 3 Boston 6.1 5.3 4.5
5 25 Washington 5.7 4.7 4.1
seventh of Firm A's value. Similar analysis with r2 ¼0.25 using
DC Eq. (3) shows that to capture 75% of the demand, Firm B must
6 14 Miami 5.5 5.2 6.6 provide a distance (or cost) of 14.3–60% of Firm A's value; to share
7 22 San 5.1 4.9 8.7 the demand equally (capture 50%), Firm B must provide a distance
Francisco
(or cost) of 60–166.7% of Firm A's value; to capture 25% of the
8 9 Detroit 4.3 5.0 3.3
9 18 Philadelphia 3.6 3.6 3.3 demand, it must provide a distance (or cost) of 166.7–700% of Firm
10 7 Dallas 3.1 3.8 3.0 A's value; and it will capture none of the demand if it provides a
11 6 Cleveland 3.0 3.0 2.1 distance (or cost) of over 700% of Firm A's value.
12 21 St. Louis 2.9 2.6 2.2 Fig. 2(a) shows the distribution of the percentage of OD pairs
13 1 Atlanta 2.8 2.9 2.3
14 15 Minneapolis 2.5 3.5 2.3
captured by Firm A in each level of the five-level step function for
15 8 Denver 2.4 3.2 2.7 an instance with qA ¼qB ¼2 and α ¼0.2 with low selectivity. In this
16 10 Houston 2.4 2.9 2.4 figure the vertical axis shows the percentage of the OD pairs that
17 11 Kansas City 2.0 1.8 1.7 are captured by Firm A. Thus, with low selectivity 88.33% of the OD
18 23 Seattle 1.9 2.1 3.1
pairs (the middle bar) split the revenues equally between Firm A
19 24 Tampa 1.9 1.8 1.8
20 16 New Orleans 1.8 2.0 1.7 and Firm B; 8% of the OD pairs have Firm A capturing 25% of the
21 2 Baltimore 1.7 1.5 1.4 revenues (with Firm B capturing 75%); and the remaining 3.67% of
22 5 Cincinnati 1.6 2.1 1.0 the OD pairs have Firm A capturing 75% of the revenues (with Firm
23 19 Phoenix 1.5 1.3 1.8 B capturing 25%). Neither firm is able to capture 100% of the
24 20 Phoenix 1.5 1.3 1.8
25 13 Memphis 1.2 1.4 0.9
revenue for any OD pair with low selectivity, so the left- and right-
most bars in Fig. 2(a) are absent.

Fig. 2. The distribution of the percentage of OD pairs captured by Firm A with low, medium and high selectivity (using customer allocation based on the distance ratio, the
Price revenue set, and qA ¼ qB ¼ 2 and α¼ 0.2): (a) low selectivity, (b) medium selectivity, and (c) high selectivity.
 M. Sasaki et al. / Computers & Operations Research 47 (2014) 27–41 33

For our second customer allocation pattern, we sought a more
even distribution of OD pairs among the five different levels of
allocation (100%, 75%, 50%, etc.). Based on our experimentation
over a large number of instances (i.e., number of hub arcs, revenue
sets, values of α, etc.), we denote r1 ¼ 0.083 and r2 ¼0.015 as
medium customer selectivity. Fig. 2(b) shows the more even
distribution of the percentage of OD pairs captured by Firm A
with medium customer selectivity (ranging from 11% to 31%) for
the same instance shown in Fig. 2(a).
For our third customer allocation pattern, denoted high custo-
mer selectivity, we set r1 ¼ r2 ¼ 0. In this case, the firm providing
better service captures all customers and the other firm captures
none. When the two hub networks provide identical service (i.e.,
identical path distances when using DRA;ℬ ij or identical path costs
when using CRA;ℬ ij ), we split the demand equally between the
competitors. Fig. 2(c) shows the percentage of OD pairs captured
by Firm A with high customer selectivity for the same instance
shown in Fig. 2(a). Although Firm A captures 100% of the revenues
for only about one-third of the OD pairs in this instance, these OD
pairs provide 52.6% of the total revenue, as Firm A captures more
higher revenue customers. Fig. 2 shows the flexibility of our
customer allocation function with its ability to model different
levels of customer selectivity. This is just one illustration, but
similar patterns were observed in other problem instances.
Our results use the low, medium and high levels of customer
selectivity to explore how the optimal networks and optimal revenue
captured depend on the customer allocation. If a carrier knows the
customer response pattern for a particular market, then appropriate
values of r1 and r2 should be selected to reflect the customer
behavior. To simplify the notation in presenting and discussing
results, we will adopt the notation HALCE(RS,AL,SL,qA,qB,α) where RS
indicates the revenue set (RP¼Price, RD¼Distance), AL indicates the
allocation type (PD¼ path distance, PC¼path cost), and SL indicates
the level of customer selectivity (low, medium or high).
The remainder of this section includes results from 360 instances
using the two revenue sets (Price and Distance), three levels of
customer selectivity (low, medium and high), five values of α (0.2,
0.4, 0.6, 0.8, 1.0) and nine 〈qA,qB 〉combinations with up to three
hub arcs for each firm: 〈qA,qB〉¼ 〈1,1〉,〈1,2〉,〈1,3〉,〈2,1〉,
〈2,2〉,〈2,3〉,〈3,1〉,〈3,2〉,〈3,3〉. The notation “〈U〉” is
used to indicate the combinations of qA hub arcs for Firm A and qB
hubs arcs for Firm B. Each problem is solved to optimality using the
SHAL algorithm from Section 3. We first consider customer alloca-
tion based on the distance ratio DRA;ℬ
ij . This represents transporta-
tion in systems where minimizing travel distance is a prime concern,
as may occur with high service level freight or passengers.
4.2. Base case
As a base case, we present detailed results using low customer
selectivity in Table 3 for the Price revenue set and in Table 4 for the
Distance revenue set. Because results for α ¼1.0 were identical to
those for α ¼0.8, we include only the results for α ¼ 0.8. (Note also
that with α ¼ 1, the hub arcs provide no advantage, so hub arcs that
connect any combination of the hub nodes shown are optimal.)
The results show that Firms A and B capture similar revenue
shares in all problems with low customer selectivity, even when
one firm has three hub arcs and the other has only one. Firm A's
share ranges from 46.12% to 53.75% in Table 3, with a similar, but

Table 3
Optimal results for HALCE(RP,PD,low,qA,qB, α).

qA qB α CPU sec. A's hub arcs B's hub arcs A's share(%) B's share(%)

1 1 0.2 0.11 2–21 4–17 49.77 50.23

0.4 0.08 4–17 8–25 50.33 49.67
0.6 0.06 4–17 8–25 50.55 49.45
0.8 0.08 4–17 8–25 50.57 49.43
1 2 0.2 2.70 4–20 1–18 5–8 47.12 52.88
0.4 1.12 4–17 5–8 10–25 48.14 51.86
0.6 1.14 4–17 8–9 10–25 48.40 51.60
0.8 1.12 4–17 8–10 9–25 48.41 51.59
1 3 0.2 295.72 4–20 1–18 5–8 7–15 46.12 53.88
0.4 109.06 4–17 1–3 5–7 8–25 46.91 53.09
0.6 107.31 4–17 1–12 3–9 7–25 47.14 52.86
0.8 107.48 4–17 1–12 3–9 7–25 47.15 52.85
2 1 0.2 4.36 4–8 17–20 11–25 52.64 47.36
0.4 4.42 4–7 12–17 11–25 52.52 47.48
0.6 4.64 1–4 12–17 11–25 52.59 47.41
0.8 4.30 1–17 4–12 7–25 52.57 47.43
2 2 0.2 23.91 4–8 12–17 7–22 9–18 50.64 49.36
0.4 16.50 4–8 17–20 5–25 10–12 50.67 49.33
0.6 15.56 4–7 17–22 9–25 11–12 50.69 49.31
0.8 17.84 4–22 7–17 9–25 11–12 50.69 49.31
2 3 0.2 1189.12 4–8 17–20 1–3 5–25 7–22 49.32 50.68
0.4 636.25 4–7 12–17 3–22 9–11 16–25 49.46 50.54
0.6 767.64 4–7 17–22 3–9 11–25 12–16 49.60 50.40
0.8 907.18 4–22 7–17 3–9 11–25 12–16 49.60 50.40
3 1 0.2 482.28 1–4 7–12 17–20 2–21 53.75 46.25
0.4 498.92 1–12 4–7 17–20 2–13 53.64 46.36
0.6 492.65 1–17 4–7 12–20 2–13 53.60 46.40
0.8 504.31 1–17 4–7 12–20 13–25 53.60 46.40
3 2 0.2 847.58 4–8 7–22 17–20 9–25 11–19 51.74 48.26
0.4 755.32 4–8 16–22 17–20 5–25 7–12 51.73 48.27
0.6 657.08 1–17 4–7 20–22 3–8 5–25 51.73 48.27
0.8 640.69 1–17 4–7 20–22 3–8 5–25 51.73 48.27
3 3 0.2 4809.90 3–17 4–20 7–22 5–8 12–18 13–15 50.72 49.28
0.4 4516.08 3–17 4–7 20–22 1–18 9–12 11–15 50.64 49.36
0.6 3813.45 1–17 4–7 20–22 3–9 12–13 15–25 50.67 49.33
0.8 3901.92 1–17 4–7 20–22 3–9 12–13 15–25 50.67 49.33
34 M. Sasaki et al. / Computers & Operations Research 47 (2014) 27–41

Table 4
Optimal results for HALCE(RD,PD,low,qA,qB,α).

qA qB α CPU sec. A's hub arcs B's hub arcs A's share(%) B's share(%)

1 1 0.2 0.16 7–25 5–19 49.70 50.30

0.4 0.12 8–20 7–18 49.74 50.26
0.6 0.14 6–22 4–18 49.79 50.21
0.8 0.12 6–22 2–12 49.84 50.16
1 2 0.2 9.42 5–8 4–11 17–22 47.92 52.08
0.4 2.06 4–17 5–7 8–25 47.99 52.01
0.6 4.08 13–17 4–7 12–25 48.05 51.95
0.8 4.09 13–17 4–25 7–12 48.07 51.93
1 3 0.2 580.65 5–8 4–11 13–19 17–23 47.28 52.72
0.4 466.26 2–11 1–17 4–7 12–25 47.36 52.64
0.6 524.34 13–17 3–25 4–24 7–12 47.49 52.51
0.8 524.26 13–17 3–25 4–24 7–12 47.51 52.49
2 1 0.2 5.95 4–8 12–17 13–25 51.97 48.03
0.4 5.78 1–12 4–18 2–11 51.98 48.02
0.6 4.67 4–7 12–17 9–22 51.92 48.08
0.8 4.72 1–17 4–22 9–12 51.86 48.14
2 2 0.2 26.47 4–8 12–17 7–22 21–25 50.19 49.81
0.4 19.25 4–7 12–17 9–11 22–25 50.25 49.75
0.6 16.22 4–7 17–22 3–25 12–21 50.43 49.57
0.8 18.23 4–22 7–17 3–25 12–21 50.43 49.57
2 3 0.2 804.91 4–8 12–17 1–25 7–22 19–21 49.49 50.51
0.4 618.11 4–7 12–17 3–22 9–11 16–25 49.64 50.36
0.6 550.78 4–7 17–22 3–9 11–25 12–16 49.87 50.13
0.8 763.33 4–22 7–17 3–9 11–25 12–16 49.87 50.13
3 1 0.2 644.18 4–8 7–12 17–25 2–21 52.69 47.31
0.4 666.97 1–12 4–17 7–20 2–13 52.54 47.46
0.6 544.43 1–4 7–17 12–20 9–22 52.51 47.49
0.8 562.97 4–8 7–22 17–24 11–25 52.41 47.59
3 2 0.2 1355.77 1–7 4–8 12–17 19–22 21–25 50.96 49.04
0.4 1249.52 1–12 4–7 17–22 3–25 19–21 50.97 49.03
0.6 983.57 1–17 4–22 7–12 3–21 19–25 50.97 49.03
0.8 938.79 1–17 4–22 7–12 3–21 19–25 50.97 49.03
3 3 0.2 5887.58 4–8 7–22 17–20 3–21 9–25 10–12 50.37 49.63
0.4 4077.64 1–12 4–7 17–22 2–16 3–19 9–11 50.42 49.58
0.6 3212.09 1–17 4–22 7–12 2–16 3–11 9–19 50.43 49.57
0.8 3568.35 1–17 4–22 7–12 2–16 3–11 9–19 50.43 49.57

slightly smaller range in Table 4. Not surprisingly, the firm with

more hub arcs always captures more than 50% of the revenue, and
adding a hub arc for either firm will increase its share, though the
improvements are marginally decreasing. The results also show
the share captured is rather insensitive to α for each particular
number of hub arcs, with the greater differences for smaller values
of α due to the relatively greater changes between small α values.
Another interesting finding is that the “first entry paradox”
[27], in which Firm A does not take advantage of being the leader
(it has less than 50% of the revenue), occurs in some instances with
qA ¼qB ¼1. Note that for any given optimal solution, if the leader
was to instead select the hub arcs used by the follower, then the
optimal hub arcs for the follower are not necessarily those that
were selected by the leader in the optimal solution. Consider for Fig. 3. Firm A's share with different levels of customer selectivity for HALCE (RP,PD,
example, the case where qA ¼qB ¼1 with α ¼0.2 (i.e., the first row low,qA, qB,α).
in Table 4). If Firm A was to locate its hub arc according to Firm B's
optimal location in Table 4, i.e., 5–19, then Firm B's best location is
hub arc 9–12, not 7–25, which decreases Firm A's share to 48.94%, lines and the dashed lines show the results with α ¼0.8 and α ¼ 0.2,
instead of the 49.70% shown in Table 4. This lack of symmetry respectively. The results for other values of α, and for the Distance
highlights the complexity of the model and interactions in revenue set are similar. Figs. 2 and 3 show some interesting relations
competitive location. between captured share and the distribution of demand for different
levels of selectivity. When the selectivity level is low, both firms share
4.3. Customer selectivity the customers almost equally regardless of the number of hub arcs
selected, since the demand for the vast majority of OD pairs is split
To explore how customers' preferences influence the optimal equally (see Fig. 2). At the other extreme of high selectivity, a firm can
networks, we compare results with the low, medium and high levels gain a considerable advantage when it has more hub arcs than its
of customer selectivity. Complete results for these problems are competitor. For example, with high selectivity Firm A captures over
available in [49]. Fig. 3 shows Firm A's share (% of total revenue) 77% of the revenue (for all α values) when 〈qA,qB〉¼〈3,1〉, and
for the problems with the Price revenues and 〈qA,qB〉¼〈3,1〉, Firm B captures over 65% of the revenue when 〈qA,qB〉¼〈1,3〉.
〈2,2〉 and 〈1,3〉 with high, medium and low selectivity. The solid Fig. 3 also reveals a small advantage for the leader, as the captured
 M. Sasaki et al. / Computers & Operations Research 47 (2014) 27–41 35

revenue shares are not symmetric about 50%, but rather are skewed to Table 6
favor Firm A. Ranking of hub arcs used in optimal solutions.

Hub arc Total usage Usage by A Usage by B Usage ratio Affinity

(%) (%) (%) (A%/B%)
4.4. Hub and hub arc use
1–4 26.7 23.0 3.7 6.2 A
We are especially interested in how competition influences the 4–8 16.3 16.3 0.0 1 A
optimal hub locations and hub network designs. The results suggest 1–17 13.3 13.0 0.4 35.3 A
3–17 11.5 9.6 1.9 5.2 A
that rather different network strategies are employed by the leader 4–17 11.5 10.4 1.1 9.4 A
(Firm A) and follower (Firm B), even though the shares of revenue 4–7 10.4 9.3 1.1 8.4 A
captured may be very similar. A pairwise comparison of the optimal 12–17 10.4 10.4 0.0 1 A
networks for the two revenue sets (from Tables 3 and 4) for the 4–22 8.9 8.2 0.7 11.1 A
same instances (i.e., values of qA, qB, and α) shows that Firm A uses 12–22 7.8 5.9 1.9 3.2 A
8–17 6.7 6.3 0.4 17.2 A
the same optimal hub arcs with both revenue sets in 27 of the 72 12–23 6.7 3.7 3.0 1.3 –
instances (37.5%); in contrast, Firm B uses the same optimal hub 3–9 6.3 0.0 6.3 1 B
arcs with the two revenue sets in only 12 of the 72 instances 14–17 6.3 6.3 0.0 1 A
(16.7%). Similar results hold for the medium and high levels of 6–11 5.6 0.0 5.6 1 B
7–12 5.6 3.3 2.2 1.5 –
customers selectivity, and Tables 5 and 6 summarize the hub nodes 8–14 4.8 2.6 2.2 1.2 –
and hub arc use, respectively, in 270 instances using customer 17–20 4.8 3.7 1.1 3.4 A
allocation based on the distance ratio DRA;ℬij (from 45 instances, as 2–13 4.1 1.1 3.0 2.7 B
in Tables 3 and 4, for each of the six problem scenarios using the 4–6 3.0 2.2 0.7 3.0 A
7–8 3.0 3.0 0.0 1 A
Price or Distance revenue set and low, medium or high customer
20–22 3.0 2.6 0.4 7.1 A
selectivity). Results for each problem scenario are in Appendix B
and complete results are in [49]. In Tables 5 and 6, the first column
identifies the hub city or hub arc, respectively. These are sorted in instances overall and is commonly used by both firms, though not
decreasing order of the total usage shown in column 2 (i.e., the in the same instance (since hubs are not shared). The next most
percentage of times a city or hub arc appears in the optimal frequently used cities, 8¼ Denver and 1 ¼Atlanta, show similar
network for all 270 problems). The third and fourth columns aggregate usage of about 50%, primarily by Firm A, although this
provide the percentage of times a city or hub arc appears in the derives from different patterns: Denver's usage as a hub varies
optimal network for Firm A and Firm B, respectively. To show across the six problem scenarios from 15.6% to 84.4%, while
intensity of usage, the fifth column is the ratio of the maximum Atlanta is used more evenly across the scenarios. Collectively, the
usage by A or B to the minimum usage and the final column shows results in Table 5 show that Firm A utilizes a small set of hub cities
the affinity for the hub or hub arc, with a usage ratio of 2 or greater intensively, especially 4, 17, 12, 8, 1 and 7, while Firm B uses a
indicating affinity to the heavier user and a ratio of less than wider range of hub cities, and not as intensively (as shown in the
2 indicating more equal usage by the competing firms. “Usage by B” column. Firm B most heavily uses cities 25, 3, 11, 21,
Table 5 shows that Chicago (city 4 at 91.5%) and New York (city and 9, though none are used as intensively as are the hubs for
17 at 85.2%) are optimal hub locations in nearly every problem, Firm A. Only Los Angeles (city 12) is consistently and commonly
especially for Firm A. There is a significant break before the next used as a hub by both firms, though 22, 20 and 14 are also used by
city, Los Angeles (city 12), which is an optimal hub in 57.0% of both firms to a lesser degree. The variation in hub usage across the
six different problem scenarios (see Appendix B) highlights the
Table 5 importance of the customer selectivity in modeling different
Ranking of cities used as hubs in optimal solutions.
transport systems. For example, Miami (city 14) is heavily used
City Total usage Usage by A Usage by B Usage ratio Affinity with high selectivity, but not at all with low selectivity. Conversely,
(%) (%) (%) (A%/B%) Dallas (city 7) and Kansas City (city 11) are used predominantly
with low and medium selectivity, respectively.
4 91.5 79.3 12.2 6.5 A The frequent appearance of certain cities as hubs (Chicago,
17 85.2 76.7 8.5 9.0 A
12 57.0 35.6 21.5 1.7 –
New York, Los Angeles, Denver and Atlanta are the top five cities in
8 50.0 35.2 14.8 2.4 A Table 5) suggests that some cities are attractive hub locations regard-
1 49.3 38.2 11.1 3.4 A less of the particular problem (customer selectivity, revenue set and α
25 47.8 4.5 29.6 6.7 B value). These robust hub locations may be due to a city's large
7 41.5 27.8 13.7 2.0 A
demand, as for Chicago, New York, Los Angeles which have the three
3 40.7 10.8 30.0 2.8 B
22 37.0 23.4 13.7 1.7 – largest demands; or to a city's geographic position, as for Denver and
20 35.2 18.9 16.3 1.2 – Atlanta which have the 13th and 15th largest demands (see Table 2).
11 33.7 4.5 29.3 6.6 B Thus, more centrally located Atlanta (city 1) and Denver (city 8) are
21 30.4 3.3 27.0 8.2 B favored as optimal hubs over very similar sized Houston (city 10) and
9 28.2 1.5 26.7 18.0 B
14 27.4 14.2 18.7 1.3 –
Minneapolis (city 15) (which have the 14th and 16th largest demands).
6 25.6 5.9 19.6 3.3 B Similarly, peripheral cities Boston (city 3) and Miami (city 14) seem
2 21.5 7.0 14.5 2.1 B disadvantaged, as they rank fourth and sixth in terms of demand (see
18 15.9 1.9 14.1 7.6 B Table 2), but 8th and 15th in usage as a hub (Table 5). These results
5 14.8 1.5 13.3 9.1 B
highlight the complementary roles of demand at a city (i.e., originating
13 13.7 5.6 8.2 1.5 B
15 13.0 0.4 12.6 34.4 B and terminating traffic) and geography in determining optimal hub
19 12.2 0.0 12.2 1 B locations.
16 9.6 1.1 8.5 7.7 B Table 6 summarizes the hub arc usage for the 270 instances,
23 9.3 3.7 5.6 1.5 B similar to that for hub usage in Table 5. The 21 hub arcs included in
24 7.4 0.7 6.7 9.1 B
10 6.7 1.1 5.6 5.0 B
this table are the five most frequently used hub arcs in the optimal
solutions for each of the six problem scenarios. Not surprisingly, the
36 M. Sasaki et al. / Computers & Operations Research 47 (2014) 27–41
top eight hub arcs in Table 6 use either Chicago or New York (city
4 or 17) or both, reflecting their importance as hubs (as shown in
Table 5). The most frequently used hub arc (1,4) connects the first
and fifth most used hub cities, and is used in about one-quarter of
the problems overall. Reflecting Firm A's intensive use of a few hubs
(Table 5), the top nine hub arcs for Firm A in Table 6 connect only
eight different hubs, while the top nine hub arcs for Firm B connect
17 different hubs. As with its intensive use of few hub cities, Firm A
uses a small fraction of the hub arcs intensively, with five hub arcs
used at a level over 10% and 15 of the 21 hub arcs in Table 6 being
used predominantly by Firm A. In contrast, Firm B uses only two
hub arcs at a level over 5% in Table 6 (maximum¼ 6.3%). Interest-
ingly, while Firm B uses a wider range of hub arcs at less intensity
compared to Firm A, it does manage to capture nearly the same
revenue as Firm A does when using the same number of hub arcs.
To illustrate the complex patterns of hub arc use, Fig. 4 shows
the optimal hub arcs with the Price and Distance revenue sets with
low, medium and high selectivity for qA ¼ qB ¼ 2 and α ¼0.6. The
solid lines show Firm A's hub arcs and the dashed lines show Firm
B's hub arcs. (Note that Firm B uses hub arc (3,25) in (b) and (f),
and hub arc (5,9) in (e).) With low selectivity, the pattern of hub
arcs is quite similar with both revenue sets, but with medium and
high selectivity the patterns are quite different. Of the 24 hub arcs
in these six optimal solutions, there are 20 different hub arcs using
20 different hub cities, with Chicago and New York being optimal
hubs in all six solutions (mostly for Firm A), Los Angeles being a
hub twice for Firm A and twice for Firm B, and Atlanta and Dallas
being hubs for Firm A alone, in three of the six solutions. This is an
illustrative example of how the level of customer selectivity can
produce dramatically different optimal hub networks, even though
Fig. 4. Optimal solutions for qA ¼ qB ¼ 2 and α¼ 0.6: (a) HALCE(RP,PD,low,2,2,0.6), (b) HALCE(RD,PD,low,2,2,0.6), (c) HALCE(RP,PD,medium,2,2,0.6), (d) HALCE(RD,PD,
medium,2,2,0.6), (e) HALCE(RP,PD,high,2,2,0.6), and (f) HALCE(RD,PD,high,2,2,0.6).
the revenue captured and some of the optimal hub nodes may be
very similar between the competitors.
4.5. Customer allocation using the cost ratio
The results above use the customer allocation function based on
the ratio of travel distances ðDRA;ℬ
ij Þ, which is most appropriate for
customers that place a premium on shorter travel times. In this
section, we provide results for 90 instances using the customer
allocation function based on the ratio of travel costs ðCRA;ℬ
ij Þ. These
problems are most appropriate for customers that place a premium on
low costs, which is achieved by carriers routing flows more circui-
tously to minimize cost. In these results we use the same two revenue
sets, five α values, and nine 〈qA,qB〉 combinations as before. We
present an illustration of the results here using the high level of
customer selectivity, which corresponds to the situation in [43]. For
complete results with the customer allocation function based on the
cost ratio, see [49].
As an example of the results, Table 7 provides the optimal solutions
with 〈qA,qB〉¼〈3,3〉, and for comparison purposes the corre-
sponding results with customer allocation based on the distance ratio
DRA;ℬ
ij (which were summarized as part of Tables 5 and 6). The first
column in this table indicates the revenue set and the customer
allocation type. A comparison of the first ten rows and the last ten
rows shows that the cost ratio CRA;ℬij gives Firm B some advantages to
capture more revenue, especially for small values of α. (Note that for
α ¼1.0 the cost ratio and distance ratio become equivalent.) The results
also show that although many optimal hubs are common between the
solutions using the cost and distance ratios (Chicago, Los Angeles and
 M. Sasaki et al. / Computers & Operations Research 47 (2014) 27–41 37

Table 7
Optimal results for HALCE(RS,AL,high,3,3,α).

Problem α CPU sec. A's hub arcs B's hub arcs A's share (%) B's share (%)

RP,PC 0.2 967.92 4–12 17–21 18–24 2–14 6–7 22–25 54.96 45.04
0.4 820.38 4–12 6–17 7–25 2–21 9–24 19–20 56.73 43.27
0.6 861.01 4–12 13–18 17–20 6–21 15–22 24–25 58.12 41.88
0.8 1043.79 4–17 7–25 9–12 2–18 6–8 14–21 58.81 41.19
1.0 709.49 4–9 7–17 12–25 2–22 3–21 6–14 63.09 36.91
RD,PC 0.2 467.36 4–22 12–18 17–21 1–20 8–25 13–19 62.37 37.63
0.4 654.88 4–17 9–12 19–22 1–18 6–25 8–21 59.49 40.51
0.6 929.15 4–17 5–12 18–22 19–21 20–24 21–25 58.95 41.05
0.8 502.90 4–12 7–17 18–22 8–11 9–20 24–25 63.42 36.58
1.0 341.00 4–22 12–25 14–17 1–8 3–18 9–21 71.47 28.53
RP,PD 0.2 1111.08 1–4 3–17 12–23 14–25 16–21 20–24 62.82 37.18
0.4 1051.97 1–4 3–17 12–22 5–9 8–14 10–15 60.52 39.48
0.6 1011.01 1–4 3–17 12–14 5–9 10–15 22–23 61.56 38.44
0.8 779.09 1–4 12–22 14–17 3–10 9–25 15–21 62.93 37.07
1.0 1087.78 4–9 7–17 12–25 2–22 3–21 6–14 63.09 36.91
RD,PD 0.2 648.97 1–4 3–17 12–23 14–25 16–21 20–24 67.80 32.20
0.4 634.64 1–4 12–22 14–17 3–25 15–21 19–23 69.22 30.78
0.6 629.72 1–4 12–22 14–17 5–9 7–15 24–25 69.34 30.66
0.8 700.72 3–17 4–14 12–22 1–9 7–11 23–25 72.18 27.82
1.0 827.12 4–22 12–25 14–17 1–18 3–21 8–9 71.47 28.53

Fig. 5. Optimal solutions for qA ¼ qB ¼ 3 and α ¼ 0.6: (a) HALCE(RD,PC,high,3,3,0.6) and (b) HALCE(RD,PD,high,3,3,0.6).

New York are optimal in all solutions in Table 7), different hub arcs are
being used.
Fig. 5 is an example of the optimal hub networks using the
customer allocation function based on the cost ratio and the
distance ratio for the Distance revenue set with qA ¼qB ¼3 and
α ¼0.6. These two maps show strikingly different patterns that
result from the quite different underlying motivations reflected in
the cost and distance ratios. With the distance ratio (Fig. 5(b)), the
customers prefer shorter travel distances, so one-stop paths
(origin-hub-destination) have high utility and use of the hub arcs
is discouraged. Thus, the hub arc networks in Fig. 5(b)) are
designed to create more one-stop trips and the locations of the
hub nodes are more important than the locations of the hub arcs.
This is reflected in the hub arcs utilization, where in Fig. 5(b) less
than 15% of the total revenues are derived from trips that utilize a
hub arc (Firms A and B derive only 9.1% and 5.2%, respectively, of
their total revenues from trips that utilize a hub arc). In contrast,
with the cost ratio (Fig. 5(a)), traffic is routed via the hub arcs to
take advantage of the cost savings. In this solution, two-thirds of
the total revenues result from trips that utilize a hub arc, with
Firms A and B deriving 40.9% and 24.7% of their total revenues,
respectively, from trips that utilize a hub arc. Thus, the hub arcs in
Fig. 5(a) are positioned to attract customers, so they tend to reflect
the underlying demand patterns (between east and west, and
between the northeast and Florida). These results in Table 7 and
Fig. 5 reflect some interesting behaviors that suggest strong
differences in service-oriented and cost-oriented optimal hub
networks. These results highlight how optimal network design
can be very sensitive to customer preferences, and the difficulties
of designing optimal competitive hub networks to handle
customers with different preferences in the same vehicle (e.g.,
expedited and deferred freight, or business and leisure
passengers).
4.6. Computation times
All the instances in this paper were solved to optimality with
the SHAL algorithm described in Section 3. Details on the cpu
times for all instances are available in [49]. For the largest
problems, with 〈3,3〉, the average cpu times across the different
revenue sets and customer selectivities were 3545 s and 915 s for
the distance ratio and cost ratio problems, respectively. The
maximum cpu time was 9110 s for 〈3,3〉 with α ¼ 0.6, medium
selectivity and the distance ratio. In general, the cpu times were
largest with medium selectivity and smallest with high selectivity.
As expected, the cpu times increase as the number of hub arcs
increases as shown in the results. With n cities, there are O(n2)
potential hub arcs connecting pairs of cities, and the number of
possible network configurations with 〈qBA,qB〉 hub arcs is of
order Oðn2ðqA þ qB Þ Þ. Thus, the number of potential competitive hub
networks increases polynomially with the number of nodes
(cities) and exponentially with the number of hub arcs. However,
the bounding procedures in the SHAL algorithm are effective in
reducing the computational effort to find optimal solutions, and the
cpu times increase at a considerably slower rate than does the
number of potential networks. For example, there are 88,804 times
as many possible networks (i.e., hub arc combinations) to consider
with 〈3,3〉, compared to with 〈2,2〉; yet the increase in the
average cpu times was only 212.4 times greater with the distance
ratio and 203.7 times greater with the cost ratio for these problems.
38 M. Sasaki et al. / Computers & Operations Research 47 (2014) 27–41

Similarly, there are 298 times as many possible networks with EU. Future research could also consider models that allow hubs to
〈3,3〉, compared to having one fewer hub arc with 〈3,2〉 or be shared by both firms and more complex models that transport
〈2,3〉, yet the average increase in cpu time was only 4.3 times several types of customers in the same vehicle (e.g., business
greater. This shows the effectiveness of the bounding procedures passengers, leisure passengers and belly cargo in the same air-
and that the computational effort and cpu time increase at a much craft). Extensions of our models include developing better bound-
slower rate than does the number of possible networks. This savings ing strategies in the SHAL algorithm and exploring use of the SHAL
in computational effort for the SHAL algorithm derives from both algorithm as a heuristic by limiting the cpu time. Research should
the bounding that prevents evaluating all possible hub arc combina- also be directed at extending models to include revenues and
tions and the bounding that prevents evaluating all OD pairs for costs, as in the airline modeling (e.g., [1,2]).
each hub arc combination. To examine one instance in detail,
consider the HALCE(RD,PD,high,3,3,0.6) problem shown in Fig. 5(b).
For this instance, the SHAL algorithm evaluated only 1.78  107 hub
arc combinations from the possible total of 4.28  1012 hub arc Acknowledgments
combinations required by complete enumeration. Furthermore, for
each hub arc combination the algorithm evaluated on average only The authors thank the anonymous referees for their valuable
132 of the 300 OD pairs due to the upper bounding procedure for comments and suggestions that have helped improve the paper
Firm B's revenue. The net effect is that the evaluation effort of the considerably. We also thank the editor for his understanding. The
SHAL algorithm was about 0.000183% of that required for complete work was supported by Grant-in-Aid for Scientific Research (C)
enumeration. (grant No. 22510162) from Japan Society for the Promotion of
Solving larger problems with more cities or more hub arcs can Science and Nanzan University Pache Subsidy I-A-1.
require much more computational effort and time due to the
limits of the solution approach. However, the SHAL algorithm may
generate good feasible solutions for larger problems in a limited
Appendix A
time. Alternative solution approaches (e.g., to attack a mathema-
tical programming formulation directly (as presented in Appendix
This appendix provides a formulation of the Stackelberg hub
A) or improvements in the SHAL algorithm (such as finding good
arc location problem as a bilevel programming problem where
orders for selecting hub arcs in Step 1 and Step 2(a) of the
each firm locates a given number of its own hub arcs to maximize
algorithm description) may be needed to find optimal solutions
the revenue it captures. The formulation uses the following
for larger problems.
decision variables (the flow variables that were defined earlier
are repeated here for completeness):
5. Conclusions
xAijkl ¼ 1 if the lowest cost path for OD pair i,j with Firm A uses
hub arc (k,l) from k to l, k A V and l A V, and 0 otherwise.
This paper provides a model and optimal results for Stackel-
xBijkl ¼ 1 if the lowest cost path for OD pair i,j with Firm B uses
berg hub arc location problems as a step towards the design of
hub arc (k,l) from k to l, k A V and l A V, and 0 otherwise.
competitive large-scale transportation systems. We examined
yAkl ¼ 1 if Firm A locates a hub arc between nodes kA V and lA V,
how the optimal solutions are affected by different customer
(l4 k) and 0 otherwise,
allocation functions, different revenue sets, the number of hub
yBkl ¼ 1 if Firm B locates a hub arc between nodes k A V and lA V,
arcs, and the degree of discount for hub arc travel, and we
(l4 k) and 0 otherwise,
highlighted some interesting differences between the leader's
zAk ¼ 1 if node kA V is a hub for Firm A, and 0 otherwise, and
and follower's optimal hub arcs and hubs. The results show a
zBk ¼ 1 if node k A V is a hub for Firm B, and 0 otherwise.
strong link in many cases between the level of demand at a city (i.
e., originating and terminating traffic) and its usage as a hub.
Note that because the hub arcs are assumed to be undirected, it is
However, the results also show how geography plays an impor-
sufficient to define variables yAkl and yBkl for all l 4k for each k.
tant role in designing hub networks, with centrally located cities
To model the step function for customer allocation described in
having a locational advantage over more peripheral cities.
the paper, we define the following notation:
Because city sizes and economic activity tend to be highly
concentrated (for example, [10]), we believe our results are likely
S, the index set of steps in the step function.
to be generalizable to other real-world hub location problems.
Ls, lower limit of step sA S.
These results also highlight how optimal network design can be
Us, upper limit of step sA S.
very sensitive to the details of the customer allocation mechan-
ϕs, fraction of flow captured by Firm A when the ratio is in
ism in competitive problems – even though the amount of
step s A S.
business captured in aggregate and many of the optimal hub
M, a large number.
locations are relatively robust.
wsij ¼ 1 if the cost ratio of the path for OD pair (i,j) served by
The optimal competitive hub networks with customer alloca-
Firm A is between Ls and Us, and 0 otherwise.
tion based on OD path distances reflect the customer's desire for
shorter routes with fewer stops at hubs. These networks may be
very different than the optimal hub networks designed to mini- For example, if we use the cost ratio function shown in Table 1,
mize transportation costs. Research on competitive hub location then S, Ls, Us, and ϕs would be defined as
problems is still a new area, so there are considerable opportu- S¼ {1,2,3,4,5}, Ls ¼ ð  1; r 1 ;  r 2 ; r 2 ; r 1 Þ; U s ¼ ð  r 1 ;  r 2 ; r 2 ;
nities for future research. A few important areas are to develop r 1 ; 1Þ; and ϕs ¼ ð1:0:0:75; 0:5; 0:25; 0Þ, respectively.
better solution algorithms; to explore different customer alloca- First we consider Firm B's problem in which Firm A's decision
tion functions, especially logit functions based on real world data variables ðxAijkl ; yAkl ; and zAk Þ are all given. Given Firm A's hub arcs,
[58]), to compare solutions with different revenue sets and Firm B will then establish its hub network (i.e., hub arcs) so as to
different levels of demand; and to extend and apply this approach maximize its total revenue captured. This is formulated as for the
in a detailed setting such as airline passenger traffic in the US or hub arc model in [11] based on [40]. Thus, Firm B's competitive
 M. Sasaki et al. / Computers & Operations Research 47 (2014) 27–41 39

hub arc location problem using the customer allocation cost ratio, zBk r 1  zAk ; k A V; ðA:3Þ
denoted HALCE-B, is as follows:
zBk r ∑ yBkl þ ∑ yBkl ; kAV; ðA:4Þ
l4k lok
HALCE-B
xBijkl r yBkl ; i; j; k; l A V; j 4 i; l 4 k ðA:5Þ
 
maximize U B ¼ ∑ ∑ F ij W ij 1  ∑ ϕs wsij ðA:1Þ xBijlk r yBkl ; i; j; k; l A V; j 4 i; l 4 k ðA:6Þ
½xB ;½yB ;½zB  i A Vj 4 i sAS

xBijkk þ ∑ ðxBijkm þ xBijmk Þ r zBk ; i; j; k A V; j 4i; ðA:7Þ

m A V nfkg
subject to

∑ ∑ yBkl ¼ qB ; ðA:2Þ ∑ ∑ xBijkl ¼ 1; i; j A V; j 4 i; ðA:8Þ

k A Vl 4 k k A Vl 4 k

Table B1
Ranking of cities used as hubs in optimal solutions.

# Hub % RP,PD,low RP,PD,medium RP,PD,high RD,PD,low RD,PD,medium RD,PD,high

Total A B Total A B Total A B Total A B Total A B Total A B

4 91.5 100.0 97.8 2.2 80.0 73.3 6.7 91.1 71.1 20.0 91.1 68.9 22.2 88.9 82.2 6.7 97.8 82.2 15.6
17 85.2 95.6 93.3 2.2 84.4 68.9 15.6 100.0 95.6 4.4 86.7 80.0 6.7 57.8 46.7 11.1 86.7 75.6 11.1
12 57.0 60.0 24.4 35.6 13.3 0 13.3 68.9 53.3 15.6 82.2 42.2 40.0 35.6 17.8 17.8 82.2 75.6 6.7
8 50.0 46.7 13.3 33.3 84.4 77.8 6.7 42.2 26.7 15.6 26.7 24.4 2.2 84.4 68.9 15.6 15.6 0 15.6
1 49.3 48.9 31.1 17.8 42.2 35.6 6.7 53.3 37.8 15.6 35.6 31.1 4.4 51.1 44.4 6.7 64.4 48.9 15.6
25 43.3 77.8 0 77.7 8.9 0 8.9 20.0 6.7 13.3 64.4 4.4 4.4 60.0 8.9 24.4 55.6 6.7 48.9
7 41.5 71.1 48.9 22.2 35.6 28.9 6.7 20.0 8.9 11.1 80.0 55.6 24.4 28.9 24.4 4.4 13.3 0 13.3
3 40.7 37.8 4.4 33.3 24.4 6.7 17.8 68.9 26.7 42.2 42.2 0 42.2 22.2 15.6 6.7 48.9 11.1 37.8
22 37.0 42.2 35.6 6.7 2.2 0 2.2 31.1 6.7 24.4 66.7 48.9 17.8 13.3 0 13.3 66.7 48.9 17.8
20 35.2 46.7 44.4 2.2 55.6 37.8 17.8 37.8 4.4 33.3 8.9 8.9 0 31.1 17.8 13.3 31.1 0 31.1
11 33.7 26.7 0 26.7 42.2 6.7 35.6 15.6 8.9 6.7 33.3 2.2 31.1 62.2 8.9 53.3 22.2 0 22.2
21 30.4 4.4 2.2 2.2 53.3 8.9 44.4 35.6 2.2 33.3 26.7 0 26.7 20.0 4.4 15.6 42.2 2.2 40.0
9 28.1 42.2 0 42.2 20.0 2.2 17.8 40.0 6.7 33.3 31.1 0 31.1 17.8 0 17.8 17.8 0 17.8
14 27.4 0 0 0 6.7 0 6.7 55.6 20.0 35.6 0 0 0 31.1 13.3 17.8 71.1 37.8 33.3
6 25.6 0 0 0 75.6 24.4 51.1 24.4 0 24.4 6.7 6.7 0 44.4 4.4 40.0 2.2 0 2.2
2 21.5 8.9 2.2 6.7 22.2 6.7 15.6 6.7 0 6.7 22.2 2.2 20.0 64.4 31.1 33.3 4.4 0 4.4
18 15.9 11.1 0 11.1 53.3 6.7 46.7 0 0 0 6.7 2.2 4.4 22.2 2.2 20.0 2.2 0 2.2
5 14.8 24.4 0 24.4 13.3 0 13.3 20.0 4.4 15.6 8.9 4.4 4.4 17.8 0 17.8 4.4 0 4.4
13 13.7 17.8 0 17.8 13.3 8.9 4.4 6.7 6.7 0 20.0 13.3 6.7 24.4 4.4 20.0 0 0 0
15 13.0 13.3 0 13.3 20.0 0 20.0 20.0 2.2 17.8 0 0 0 8.9 0 8.9 15.6 0 15.6
19 12.2 2.2 0 2.2 22.2 0 22.2 0 0 0 26.7 0 26.7 17.8 0 17.8 4.4 0 4.4
16 9.6 11.1 2.2 8.9 17.8 4.4 13.3 4.4 0 4.4 17.8 0 17.8 0 0 0 6.7 0 6.7
23 9.3 0 0 0 6.7 0 6.7 11.1 6.7 4.4 2.2 0 2.2 13.3 4.4 8.9 22.2 11.1 11.1
24 7.4 0 0 0 0 0 0 8.9 0 8.9 11.1 4.4 6.7 8.9 0 8.9 15.6 0 15.6
10 6.7 11.1 0 11.1 2.2 2.2 0 17.8 4.4 13.3 2.2 0 2.2 0 0 0 6.7 0 6.7

Table B2
Ranking of hub arcs used in optimal solutions.

Hub arc % RP,PD,low RP,PD,medium RP,PD,high RD,PD,low RD,PD,medium RD,PD,high

Total A B Total A B Total A B Total A B Total A B Total A B

1–4 26.7 4.4 4.4 0 26.7 22.2 4.4 44.4 33.3 11.1 2.2 2.2 0 26.7 26.7 0 55.6 48.9 6.7
4–8 16.3 13.3 13.3 0 26.7 26.7 0 4.4 4.4 0 17.8 17.8 0 35.6 35.6 0 0 0 0
1–17 13.3 24.4 24.4 0 13.3 13.3 0 4.4 4.4 0 20.0 17.8 2.2 17.8 17.8 0 0 0 0
3–17 11.5 4.4 4.4 0 11.1 6.7 4.4 26.7 26.7 0 0 0 0 8.9 8.9 0 17.8 11.1 6.7
4–17 11.5 28.9 26.7 2.2 8.9 8.9 0 8.9 6.7 2.2 4.4 4.4 0 8.9 6.7 2.2 8.9 8.9 0
4–7 10.4 33.3 33.3 0 2.2 2.2 0 2.2 0 2.2 20.0 15.6 4.4 4.4 4.4 0 0 0 0
12–17 10.4 8.9 8.9 0 0 0 0 20.0 20.0 0 15.6 15.6 0 0 0 0 17.8 17.8 0
4–22 8.9 8.9 8.9 0 0 0 0 2.2 0 2.2 26.7 26.7 0 0 0 0 15.6 13.3 2.2
12–22 7.8 0 0 0 0 0 0 13.3 6.7 6.7 0 0 0 2.2 0 2.2 31.1 28.9 2.2
8–17 6.7 0 0 0 20.0 20.0 0 17.8 15.6 2.2 0 0 0 2.2 2.2 0 0 0 0
12–23 6.7 0 0 0 0 0 0 4.4 0 4.4 8.9 6.7 2.2 13.3 4.4 8.9 13.3 11.1 2.2
3–9 6.3 20.0 0 20.0 0 0 0 8.9 0 8.9 6.7 0 6.7 0 0 0 2.2 0 2.2
14–17 6.3 0 0 0 0 0 0 6.7 6.7 0 0 0 0 0 0 0 31.1 31.1 0
6–11 5.6 0 0 0 15.6 0 15.6 0 0 0 0 0 0 17.8 0 17.8 0 0 0
7–12 5.6 4.4 2.2 2.2 0 0 0 0 0 0 26.7 15.6 11.1 2.2 2.2 0 0 0 0
8–14 4.8 0 0 0 0 0 0 15.6 4.4 11.1 0 0 0 13.3 11.1 2.2 0 0 0
17–20 4.8 15.6 15.6 0 4.4 2.2 2.2 0 0 0 2.2 2.2 0 6.7 2.2 4.4 0 0 0
2–13 4.1 4.4 0 4.4 4.4 2.2 2.2 0 0 0 2.2 0 2.2 13.3 4.4 8.9 0 0 0
4–6 3.0 0 0 0 13.3 11.1 2.2 2.2 0 2.2 0 0 0 2.2 2.2 0 0 0 0
7–8 3.0 0 0 0 13.3 13.3 0 0 0 0 0 0 0 4.4 4.4 0 0 0 0
20–22 3.0 15.6 15.6 0 0 0 0 2.2 0 2.2 0 0 0 0 0 0 0 0 0
40 M. Sasaki et al. / Computers & Operations Research 47 (2014) 27–41

Ls ∑ ∑ C ijkl ðxAijkl þ xBijkl Þ  Mð1  wsij Þ r ∑ ∑ C ijkl ðxAijkl  xBijkl Þ; ∑ ∑ xAijkl ¼ 1; i; j A V; j 4 i; ðA:23Þ
k A Vl A V k A Vl A V k A Vl A V

i A V; j 4 i; s A S ðA:9Þ
ð½xB ; ½yB ; ½zB ; ½wÞ A arg max fU B jð½xA ; ½yA ; ½zA Þ A Tg ðA:24Þ
∑ ∑ C ijkl ðxAijkl  xBijkl Þ r U s ∑ ∑ C ijkl ðxAijkl þ xBijkl Þ þ Mð1  wsij Þ;
k A Vl A V k A Vl A V xBijkl A f0; 1g; i; j; k; l A V; j 4 i; ðA:25Þ
i A V; j 4 i; s A S ðA:10Þ
yBkl A f0; 1g ; k; l A V; l 4 k; ðA:26Þ
∑ wsij ¼ 1 ; i A V ; j 4 i; ðA:11Þ
sAS
zBk A f0; 1g ; kAV; ðA:27Þ

xBijkl A f0; 1g ; i; j; k; l A V; j 4 i;
yBkl A f0; 1g ; k; l A V; l 4 k;
zBk A f0; 1g ; k A V;
wsij A f0; 1g ; s A S; i A V; j 4 i:
The objective function (A.1) is the total revenue captured by
Firm B. Constraint (A.2) ensures that exactly qB hub arcs are
selected. Constraints (A.3) ensure that Firm B cannot select a
node as a hub if it is a hub for Firm A. Note that constraints
(A.3), (A.9) and (A.10) and the objective function link HALCE-B
to the hub arc locations for Firm A. Constraints (A.4) ensure
that hub nodes are established only at the endpoints of hub
arcs. Constraints (A.5)–(A.7) ensure that hub arcs and hub
nodes are established for every active OD path. Constraints
(A.8) ensure there is a path with Firm B for all OD pairs.
Constraints (A.9)–(A.11) ensure that the proper fraction of flow
is captured for each firm. Note that the customer allocation cost
ratio can be expressed as
ðA:12Þ The objective function (A.17) is the total revenue for Firm A.
Constraint (A.18) ensures that exactly qA hub arcs are selected.
ðA:13Þ
Constraints (A.19) ensure that hub nodes are established only at
the endpoints of hub arcs. Constraints (A.20)–(A.22) ensure
ðA:14Þ
that hub arcs and hub nodes are established for every active
ðA:15Þ path. Constraints (A.23) ensure there is a possible path with
Firm A for all OD pairs. Constraint (A.24) indicates that HALCE-
B is solved optimally for any solution to HALCE-A (i.e.,
ð½xA ; ½yA ; ½zA Þ A T), where T is the feasible region formed by
constraints (A.18)–(A.23) and (A.25)–(A.27). Thus, Firm A solves
its own problem subject to the condition that ð½xB ; ½yB ; ½zB Þ is
an optimal solution of HALCE-B. Given the interdependent
nature of the two problems HALCE-A and HALCE-B, and their
difficulties individually, we do not solve the formulations
directly.
Appendix B
See Tables B1 and B2.

∑k A V ∑l A V C ijkl ðxAijkl  xBijkl Þ References

∑k A V ∑l A V C ijkl ðxAijkl þ xBijkl Þ
[1] Adler N. Competition in a deregulated air transportation market. Eur J Oper
Res 2001;129:337–45.
So that constraints (A.9) and (A.10) could equivalently be [2] Adler N. The effect of competition on the choice of an optimal network in a
written as liberalized aviation market with an application to western Europe. Transp Sci
2005;39:58–72.
[3] Adler N, Smilowitz K. Hub-and-spoke network alliances and mergers: price-
∑k A V ∑l A V C ijkl ðxAijkl  xBijkl Þ location competition in the airline industry. Transp Res Part B 2007;41:
Ls  Mð1  wsij Þ r r U s þ Mð1  wsij Þ; 394–409.
∑k A V ∑l A V C ijkl ðxAijkl þ xBijkl Þ
[4] Akartunalıa K, Boland N, Evans I, Wallace M, Waterer H. Airline planning
benchmark problems—part II: passenger groups, utility and demand alloca-
i A V; j 4 i; s A S: ðA:16Þ
tion. Comput Oper Res 2013;40:793–804.
[5] Alamdari FE, Black IG. Passengers' choice of airline under competition: the use
of the logit model. Transp Rev 1992;12:153–70.
Firm B's competitive hub arc location problem HALCE-B is [6] Alumur S, Kara B. Network hub location problems: the state of the art. Eur
defined by (A.1)–(A.15). Firm A's competitive hub arc location J Oper Res 2008;190:1–21.
problem, denoted HALCE-A, can be formulated similarly as the [7] Blanco V, Puerto J, Ramos AB. Expanding the Spanish high speed railway
network. Omega 2011;39(2):138–50.
following bilevel programming problem. [8] Boguslaski C, Ito H, Lee D. Entry patterns in the Southwest Airlines route
HALCE-A
system. Rev Ind Organ 2004;25:317–50.
[9] Borenstin S. Hubs and high fares: dominance and market power in the U.S.
maximize U A ¼ ∑ ∑ F ij W ij ∑ ϕs wsij ðA:17Þ airline industry. RAND J Econ 1989;20:344–65.
½xA ;½yA ;½zA  i A Vj 4 i sAS [10] Brakman S, Garretsen H, Van Marrewijk C, van den Berg M. The return of Zipf:
towards a further understanding of the rank-size distribution. J Reg Sci
1999;39:183–213.
[11] Campbell JF. Hub location for time definite transportation. Comput Oper Res
subject to 2009;36:3107–16.
[12] Campbell JF. Modeling economies of scale in transportation hub networks. In:
∑ ∑ yAkl ¼ qA ; ðA:18Þ Proceedings of the 46th annual Hawaii international conference on system
k A Vl 4 k
sciences. Computer Society Press; 2013. pp.1154–63.
[13] Campbell JF, Ernst AT, Krishnamoorthy M. Hub location problems. In: Facility
zAk r ∑ yAkl þ ∑ yAlk ; kAV; ðA:19Þ location: applications and theory. Berlin: Springer; 2002; 373–407.
l4k lok [14] Campbell JF, Ernst AT, Krishnamoorthy M. Hub arc location problems: part i–
introduction and results. Manag Sci 2005;51:1540–55.
xAijkl r yAkl ; i; j; k; l A V; j4 i; l 4 k ðA:20Þ [15] Campbell JF, Ernst AT, Krishnamoorthy M. Hub arc location problems: part ii–
formulations and optimal algorithms. Manag Sci 2005;51:1556–71.
[16] Campbell JF, O'Kelly ME. Twenty-five years of hub location research. Transp Sci
xAijlk r yAkl ; i; j; k; l A V; j4 i; l 4 k ðA:21Þ 2012;46(2):153–69.
[17] Chen H, Campbell AM. Network design for time-constrained delivery. Nav Res
Logist 2008;55:493–515.
xAijkk þ ∑ ðxAijkm þ xAijmk Þ rzAk ; i; j; k A V ; j 4 i; ðA:22Þ [18] Chi J, Koo WW. Carriers’ pricing behaviors in the United States airline industry.
m A V nfkg Transp Res E 2009;45:710–24.
 M. Sasaki et al. / Computers & Operations Research 47 (2014) 27–41
[19] Colson B, Marcotte P, Savard G. An overview of bilevel optimization. Ann Oper
Res 2007;153:235–56.
[20] Contreras I, Cordeau JF, Laporte G. The dynamic uncapacitated hub location
problem. Transp Sci 2011;45:18–32.
[21] Contreras I, Cordeau JF, Laporte G. Stochastic uncapacitated hub location. Eur J
Oper Res 2011;212(3):518–28.
[22] Contreras I, Cordeau JF, Laporte G. Benders decomposition for large-scale
uncapacitated hub location. Oper Res 2011;59:1477–90.
[23] Contreras I, Díaz JA, Fernández E. Branch and price for large-scale capacitated
hub location problems with single assignment. INFORMS J Comput 2011;23:
41–55.
[24] Correia I, Nickel S, Saldanha-da-Gama F. Single-assignment hub location
problems with multiple capacity levels. Transp Res B 2010;44:1047–66.
[25] Cunha CB, Silva MR. A genetic algorithm for the problem of configuring a hub-
and-spoke network for a LTL trucking company in Brazil. Eur J Oper Res
2007;179(3):747–58.
[26] Eiselt HA, Laporte G. Sequential location problems. Eur J Oper Res 1996;96:
217–31.
[27] Eiselt HA, Laporte G, Thisse JF. Competitive location models: A framework and
bibliography. Transp Sci 1993;27:44–54.
[28] Eiselt HA, Marianov V. A conditional p-hub location problem with attraction
functions. Comput Oper Res 2009;36:3128–35.
[29] Ernst AT, Krishnamoorthy M. Efficient algorithms for the uncapacitated single
allocation p-hub median problem. Locat Sci 1996;4(3):139–54.
[30] Evans WN, Kessides IN. Localized market power in the U.S. airline industry.
Rev Econ Stat 1993;75:66–75.
[31] Gelareh S, Pisinger D. Fleet deployment, network design and hub location of
liner shipping companies. Transp Res E 2011;47:947–64.
[32] Gelerah S, Nickel S, Pisinger D. Liner shipping network design in a competitive
environment. Transp Res E 2010;46:991–1004.
[33] Grammig J, Hujer R, Scheidler M. Discrete choice modelling in airline network
management. J Appl Econom 2005;20:467–86.
[34] Grosche T. Computational intelligence in integrated airline scheduling. Berlin:
Springer-Velag; 2009.
[35] Grosche T, Rothlauf F. Air travel itinerary market share estimation. Working
Paper 6/2007. Mannheim, Germany: University of Mannheim; 2007.
[36] Hakimi SL. On locating new facilities in a competitive environment. Eur J Oper
Res 1983;12:29–35.
[37] Lian JI, Ronnevik J. Airport competition – regional airports losing ground to
main airports. J Transp Geogr 2011;19:85–92.
[38] Lin CC, Lee SC. The competition game on hub network design. Transp Res B
2010;44:618–29.
41
[39] Lüer-Villagra A, Marianov V. A competitive hub location and pricing problem.
Eur J Oper Res 2013;231:734–44.
[40] Marin A, Canovas L, Landete M. New formulations for the uncapacitated
multiple allocation hub location problem. Eur J Oper Res 2006;172:274–92.
[41] Martín JC, Román C. Analyzing competition for hub location in interconti-
nental aviation markets. Transp Res E: Logist Transp Rev 2004;40:135–50.
[42] Martín JC, Román C. Hub location in the South-Atlantic airline market: a
spatial competition game. Transp Res A 2003;37:865–88.
[43] Marianov V, Serra D, ReVelle C. Location of hubs in a competitive environ-
ment. Eur J Oper Res 1999;114:363–71.
[44] Morrison SA, Winston C. The dynamics of airline pricing and competition. Am
Econ Rev 2004;80:389–93.
[45] O'Kelly ME. A quadratic integer program for the location of interacting hub
facilities. Eur J Oper Res 1987;32:393–404.
[46] Plastria F. Static competitive facility location: an overview of optimization
approaches. Eur J Oper Res 2001;129:461–70.
[47] Plastria F, Vanhaverbeke L. Discrete models for competitive location with
foresight. Comput Oper Res 2008;35:683–700.
[48] Sasaki M. Hub network design model in a competitive environment with flow
threshold. J Oper Res Soc Jpn 2005;48:158–71.
[49] Sasaki M, Campbell JF, Ernst AT, Krishnamoorthy M. Hub arc location with
competition. Technical Report of the Nanzan Academic Society – Information
Sciences and Engineering (NANZAN-TR-2009-02); 2009.
[50] Sasaki M, Fukushima M. Stackelberg hub location problem. J Oper Res Soc Jpn
2001;44:390–402.
[51] Serra D, ReVelle C. Market capture by two competitors: the pre-emptive
location problem. J Reg Sci 1994;34:549–61.
[52] Serra D, ReVelle C. Competitive location in discrete space. In: Facility location:
a survey of applications and methods. New York: Springer-Verlag; 1995;
367–386.
[53] Silva MR, Cunha CB. New simple and efficient heuristics for the uncapacitated
single allocation hub location problem. Comput Oper Res 2009;36:3152–65.
[54] Sim T, Lowe TJ, Thomas B. The stochastic p-hub center problem with service-
level constraints. Comput Oper Res 2009;36:3166–77.
[55] Swan W, Adler N. Aircraft trip cost parameters: a function of stage length and
seat capacity. Transp Res E 2006;42:105–15.
[56] Vergara HA, Root S. Mixed fleet dispatching in truckload relay network design
optimization. Transp Res E 2013;54:32–49.
[57] Wagner B. A note on “location of hubs in a competitive environment”. Eur J
Oper Res 2008;184:57–62.
[58] Warburg V, Bhat C, Adler TJ. Modeling demographic and unobserved hetero-
geneity in air passengers’ sensitivity to service attributes in itinerary choice.
Transp Res Rec 2006;1951:7–16.