You are on page 1of 23

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/279585374

Hub-and-spoke networks in air transportation: An


analytical review

Article in Journal of Regional Science · May 1999

CITATIONS
READS
186
772

2 authors, including:

Morton E. O'Kelly
The Ohio State University
144 PUBLICATIONS 6,574 CITATIONS

Some of the authors of this publication are also working on these related projects:

Fuel Burn by Aircraft: Network Implications View project

Retail trade areas from a spatial persective View project

All content following this page was uploaded by Morton E. O'Kelly on 11 December 2017.
The user has requested enhancement of the downloaded file.
JOURNAL OF REGIONAL SCIENCE, VOL. 39, NO. 2, 1999, pp. 275–295

HUB-AND-SPOKE NETWORKS IN AIR TRANSPORTATION: AN


ANALYTICAL REVIEW

Deborah L. Bryan
Department of Geography and Urban Planning, Wayne State University, Detroit, MI
48202, U.S.A. E-mail: ae8759@wayne.edu

Morton E. O’Kelly
Department of Geography, The Ohio State University, 1036 Derby Hall, 154 N.
Oval Mall, Columbus, OH 43210, U.S.A. E-mail: okelly+@osu.edu

ABSTRACT. In this review we survey advances in analysis of the hub location problem and its
variants. In the course of the review opportunities for enhanced analysis become apparent. We
emphasize the most pressing areas for further work. We found that first, research needs to be
devoted to developing more reliable heuristics for the multiple assignment model and its extensions
and second, that additional research is needed to understand the conditions under which the model
will tend to have integer solutions. Research in this area will contribute to the solution of a
longstanding puzzle in economics about the allocation of indivisible resources.

1. INTRODUCTION
Analytical research on the hub location problem began when O’Kelly (1987) devised a
mathematical formulation of the problem. This work attracted the attention of researchers
from a wide variety of fields and since 1987 the hub location problem has been a widely-
researched area. As evidence of this, a recent special issue of Location Science is devoted to
the topic. Entire conference sessions (for example at INFORMS) have been allocated to
papers related to hubs. An introductory overview on hub networks is provided by Daskin
(1995; Chapter 8.7). In an effort to organize the growing number of papers on hub-and- spoke
networks, Campbell (1994a), O’Kelly and Miller (1994), and Skorin-Kapov and Skorin-
Kapov (1995) have written literature reviews on this topic. However, ongoing research has
already made these recent reviews out of date. Almost half of the analytical research on hub
networks has been published since the last review. In addition, the latter two reviews are
limited in scope, omitting many papers discussed here. This paper will review analytical
research as it applies to discrete hub-and-spoke networks for passenger airlines and package
delivery systems (an excellent overview of research on communication hub-and-spoke
networks is included in Campbell, 1994a).

Received October 1997; revised and accepted April 1998.


© Blackwell Publishers 1999.
Blackwell Publishers, 350 Main Street, Malden, MA 02148, USA and 108 Cowley Road, Oxford, OX4 1JF, UK.
27 JOURNAL OF REGIONAL SCIENCE, VOL. 39, NO. 2, 1999

The analysis of hub-and-spoke systems is clearly of great interest in spatial analysis (see
O’Kelly, 1998a), however, the majority of publications on the hub-and-spoke problem
appear in operations research and location or transpor- tation–specialist journals. Regional
scientists and geographers have much to contribute to the problem because of their ability
to integrate location and transportation theory, and because of the centrality of spatial
interaction to their research interests. We believe that a literature review that highlights
substan- tive questions in hub-and-spoke network analysis will serve to increase regional
scientists’ awareness of the model and perhaps stimulate further research.
Hub-and-spoke networks are utilized by many distribution systems includ- ing airline
passenger carriers and express package delivery networks. Hubs serve as transshipment
points and allow for the replacement of direct connec- tions between all nodes with
fewer, indirect connections. Figure 1 shows a completely-interconnected network and a
hub-and-spoke network. In a com- pletely-interconnected network (Figure 1a) flow
travels directly from the origi- nating city to the destination with no intermediate stops. Of
all network designs, this type results in the lowest travel times. Yet, as shown in the figure,
it also results in a large number of links, many of which are underutilized. A reduction in the
total number of links would reduce costs by bundling flows and by concentrating
equipment and sorting at specific locations.
In a hub-and-spoke network (Figure 1b), the reduction in the number of links is made
possible by the establishment of hubs or transshipment points. In a pure hub-and-spoke
network all links must either begin or end at a hub; no other connections are allowed. A
characteristic feature of such networks is the bundling of flows on the interhub links. By
bundling flows carriers can use larger aircraft and thus reduce passenger-mile costs.
Hubbing results in lower total network costs but it increases individual travel miles
because most routes are no longer direct flights and instead are one- or two-stop routes.
There are two basic types of hub-and-spoke networks, differing in how non-hub
cities are connected to the hubs. In the single assignment model each city is connected to
a single hub. There is no sorting at the origin because all flow must travel to the same
hub. The multiple assignment model allows each city to be connected to more than one
hub. Sorting must occur at each origin that interacts with more than one hub. This model
increases the number of links in the network but at the same time decreases individual
travel times.
The scope of the research on hubs is very broad but one main body has focused on
the hub location problem and the development of models to represent it. This is also the
focus of this review. The hub location problem is concerned with locating hubs and
assigning non-hub nodes to the hubs to meet a predefined objective (usually the
minimization of total network costs). A challenge in the development of these models is
to incorporate important characteristics of real hub-and-spoke networks without increasing
the complexity of the model to such an extent that it cannot be solved. Even simple hub
location models have proven difficult, though not impossible, to solve.

© Blackwell Publishers 1999.


© BR
Bl Y
ac A
kw N
ell &
Pu
bli O’
she KE
rs LL
19 Y:
98. H
U
B-
A
N
D-
SP
O
KE
NE
T
W
O
R
KS
IN
AI
R
TR
A
NS
PO
RT
AT
Note the smaller number of links required: in this case (a) has 36 undirected links while (b) has only 9 undirected links. IO

FIGURE 1: A Completely-Interconnected 9-Node Network (Panel a) and a Hub-and-Spoke Network (Panel b).
27
7
27 JOURNAL OF REGIONAL SCIENCE, VOL. 39, NO. 2, 1999

We organize this review as follows (see Table 1). First, we introduce the single
assignment model, including a discussion of heuristics designed to solve the problem and
the development of linearizations of the model which allow for optimal solutions to be
found. This is followed by a discussion of the multiple assignment model. Recognizing that
the complexity of the hub location problem prevents many important characteristics of
actual hub-and-spoke networks from being modeled, some researchers have simplified
the problem by holding the hub locations fixed. The remaining problem, then, is to
determine the allocation of non-hub nodes to the hubs. This allows for the examination of
the effect of complicated characteristics on network design, given a set of hub locations.
We discuss research of this type in Section 4. As computing capabilities improved and
knowledge of the hub location problem grew some of these extensions have been
incorporated into the endogenous hub location model. These new developments are
discussed in Section 5. Finally, we discuss research exploring the sensitivity of the hub
location model to changes in the parameters and conclude with a summary of the current
state of hub location models.

2. THE SINGLE ASSIGNMENT HUB LOCATION MODEL


Consider a set of cities (nodes) in a region that must interact with each other within a
hub-and-spoke network. All flow must travel from the origin city to the destination city via at
least one hub and, in doing so, will incur a travel cost. How should the network be
designed such that these interaction costs are minimized? This is the question posed by
O’Kelly (1987) who modeled the hub-and-spoke network as a quadratic integer program
that minimizes total network cost. Realizing that (due to the bundling of flows on the
interhub links) scale economies are one of the primary advantages to hubbing, an interhub
discount factor is included to discount the travel costs for the interhub links (relative to
the cost of traveling along the spokes). The discount factor is determined exogenously
and travel costs on the interhub links are assumed to be independent of the amount of
flow traveling across the link. O’Kelly (1987) formulated a single assignment model
where every node is assigned to a single hub. See Appendix A for a mathematical
formulation of the model. The formu- lation implies that all non-hub nodes have single
assignment whereas the hubs have multiple assignment. In addition, all interhub links are
completely inter- connected.
The model is very similar to the p-median problem with the exception of the quadratic
term in the objective function. When the cost of traveling across the interhub link is free
the quadratic term drops out and the hub location problem reduces to the p-median
(O’Kelly, 1987). In the p-median, once facility locations are determined the remaining
nodes are simply assigned to the nearest facility which minimizes total network cost.
Travel costs then consist of a single component: the cost of traveling from the origin to
the facility. However, in the hub location model nearest-center assignment is not
guaranteed to be optimal.

© Blackwell Publishers 1999.


TABLE 1: Analytical Research on Hub-and-Spoke Networks
Single Assignment Model Formulation Multiple Assignment Model Formulation

O’Kelly (1987) Campbell (1994b)


Heuristics Solution methods
O’Kelly (1987) Campbell (1994b)
Klincewicz (1991) Ernst & Krishnamoorthy (1998)
Abdinnour-Helm & Venkataramanan (1992) Klincewicz (1996) Klincewicz
(1992) Skorin-Kapov, Skorin-Kapov,
O’Kelly (1992) and O’Kelly (1996)
Abdinnour-Helm & Venkataramanan (1993)
Skorin-Kapov & Skorin-Kapov (1994) Exact network based methods
Aykin (1995) for multiple and single allocation models
O’Kelly, Skorin-Kapov, and Skorin-Kapov (1995) Ernst & Krishnamoorthy (Forthcoming b)
Campbell (1996)
Ernst and Krishnamoorthy (1996)
Smith, Krishnamoorthy, and Palaniswami (1996)
Linearizations
Campbell (1994b)
Ernst & Krishnamoorthy (1996)
O’Kelly et al. (1996)
Skorin-Kapov, Skorin-Kapov, and O’Kelly (1996)

Extensions with Fixed Hub Location Extensions with Endogenous Hub Location

Grove and O’Kelly (1986) O’Kelly (1986)


Jeng (1987) Chou (1990)
Flynn and Ratick (1988) O’Kelly (1992)
Daskin and Panayotopoulos (1989) Campbell (1993)
Hall (1989) Aykin (1994)
Miller (1990) Campbell (1994b)
O’Kelly and Lao (1991) Aykin (1995)
Dobson and Lederer (1993) Ernst and Krishnamoorthy (Forthcoming a)
Kuby and Gray (1993) Jaillet et al. (1996) O’Kelly
(1998a, 1998b) Ebery et al.
(1998) O’Kelly and Bryan
(1998)

Sensitivity Analysis

O’Kelly et al. (1996)

Unlike the p-median, each individual travel cost consists of three components:
(1) the travel cost from the origin to the hub; (2) the cost of traveling across the interhub link
(if necessary); and (3) the travel cost from the hub to the destina- tion. For each node, these
three components must be summed for every one of the node’s interactions in order to
determine the hub that will result in the lowest travel cost—making the problem nontrivial
even when hub locations are fixed (Sohn and Park, 1997).
Heuristics
Due to the quadratic nature of the hub location problem, research soon centered on
the development of heuristics to solve this problem. O’Kelly (1987) was the first to
develop two heuristics that computed upper bounds to the optimal objective function
value for the single assignment model. As time progressed the upper bounds became
tighter, bringing us closer to the true optimal solution. Several researchers were involved in
this, including Campbell (1996) with specialized heuristics; Klincewicz (1991, 1992) with
clustering, exchange, and tabu search heuristics; Abdinnour-Helm and Venkataramanan
(1992, 1993), Aykin (1995), Ernst and Krishnamoorthy (1996), and Smith,
Krishnamoorthy, and Palaniswami (1996) with heuristics borrowed from the physical
sciences; and Skorin-Kapov and Skorin-Kapov (1994) with tabu search. At the same time
lower bounds were being found. As improvements were made lower bounds became tighter,
closing in on the true optimal solution. Most notable in this line of research is the work of
O’Kelly (1992) and O’Kelly, Skorin-Kapov, and Skorin-Kapov (1995). A numerical
comparison of many of these heuristics is provided by O’Kelly et al. (1996).
The first heuristics developed were able to determine optimal hub locations given a
possibly suboptimal allocation pattern (O’Kelly, 1987). This was accom- plished by
considering all locational patterns. One of these, HEUR1, assigns all non-hub nodes to their
nearest hub, whereas the other, HEUR2, evaluates the total nodal cost of both the nearest
hub and the second-nearest hub before assigning non-hub nodes to hubs. O’Kelly (1987)
found that as the interhub discount increases, the difference in the allocation pattern and
objective- function value between the two heuristics decreases. In a discussion of O’Kelly’s
heuristics, Aykin (1990) suggested the use of a flow-based assignment approach rather than
the nearest hub approach utilized by O’Kelly (1987).
In an attempt to improve on the upper bounds obtained by O’Kelly (1987),
Klincewicz (1991) developed two sets of heuristics that use a multi-criteria distance and
flow-based allocation procedure to determine the allocation of nodes to hubs rather than
relying on distance alone. This enhanced allocation rule recognizes the importance of
flows in determining allocations in the single assignment model. Given an initial set of
good hub locations, the exchange heuristics first assign the non-hub nodes to hubs, and then
make improvements to the initial solution by exchanging hubs with non-hub nodes. The
best set of hub locations is replaced if the value of the objective function is lowered by
the exchange. On the other hand, the clustering heuristics first cluster the nodes into p
groups and then assign a hub to each group. Further improvements were made by Klincewicz
(1992) with the use of tabu search and a greedy randomized adaptive search procedure
(GRASP), although allocations are determined using the nearest hub rule. The advantage of
this heuristic over the exchange method is that local optima are more easily avoided.
By relaxing the assumption that a non-hub node must interact via a single hub,
Campbell (1996) used a greedy-exchange heuristic to find good hub locations
for the relaxed model. Given the hub locations found for the relaxed model, allocations
for the single assignment model are then determined in one of two ways. The first method
assigns a node to the hub through which it has its maximum flow, whereas the second
allocates a node to a hub so that total transportation costs are minimized. The latter
method consistently provides a tighter bound than the first. As shown by O’Kelly et al.
(1996) these two heuristics perform better (lower objective-function value) than both
O’Kelly’s and Klincewicz’s, except when the interhub discount is small.
Adapting heuristics from the physical sciences, Abdinnour-Helm and
Venkataramanan utilized both simulated annealing (1992) and genetic algo- rithms
(1993) to find good solutions to the single assignment hub location problem. An exchange
method is used to determine both good hub locations and allocations for the simulated
annealing heuristic. A distance-based assignment rule determines allocations for the genetic
algorithm. These heuristics are capable of designing large hub-and-spoke networks very
quickly. Unfortunately, initial results were disappointing with both methods because there
was no improvement in the upper bounds found by other heuristics. Using a clustering
method, Ernst and Krishnamoorthy’s (1996) simulated annealing heuristic improved on
the bounds obtained by Abdinnour-Helm and Venkataramanan (1992). Aykin (1995) also
developed a simulated annealing-based interchange heuristic to solve this problem.
Another borrowed technique from the physical sciences, neural networks, was used by
Smith, Krishnamoorthy, and Palan- iswami (1996) with disappointing results. The
simulated annealing heuristic proved superior to neural networks in almost all cases.
Perhaps more signifi- cantly, neural networks are currently capable of solving only very
small prob- lems (up to 15 nodes).
To date, the best results obtained for the single assignment model through the use of
heuristics were found by Skorin-Kapov and Skorin-Kapov (1994) with tabu search. Their
heuristic improved on that of Klincewicz (1992) by utilizing two exchange heuristics—
one for the location phase and one for the allocation phase. As shown by Skorin-Kapov,
Skorin-Kapov, and O’Kelly (1996), their tabu search heuristic found the optimal solution in
all problems tested. The effective- ness of these heuristics can be evaluated by tight lower
bounds. O’Kelly (1992) showed that a tight lower bound can be found by using a linear
approximation to the quadratic term in the objective function, which underestimates the costs
for interhub flows. The resultant objective function is a revised p-median problem with
easily calculated cost terms. Tighter lower bounds were later found by incorporating
information from good feasible solutions (O’Kelly, Skorin- Kapov, and Skorin-Kapov,
1995).

Linearizations of the Quadratic Hub Location Problem


The development of a linearized version of the quadratic model was a great advance
in hub research. This was initially accomplished by Campbell (1994b). This development
allows the use of linear programming to find optimal solutions.
Ideally, a linearization will provide integer solutions (in the case of the hub location
problem, the locations of the hubs must be integer), as the size of the problem restricts the
use of integer programming to very small networks. However, Campbell’s model resulted
in fractional solutions for all problems tested. By modifying this model, Skorin-Kapov,
Skorin-Kapov, and O’Kelly (1996) obtained a tight linearized version of the hub location
problem. A tight lineari- zation will result in integer solutions (for the hub locations)
without forcing integrality through the use of integer programming. Exact solution values
are now known for small-sized problems (up to 25 nodes).
By eliminating redundant and impractical routes and by exploiting the symmetry of
the available test data, O’Kelly et al. (1996) further modified this linearization, reducing
computation time and the number of variables without sacrificing integrality. This
reduction in problem size was made possible by exploiting the geographic nature of the
problem. For example, circuitous paths would never be found in the optimal solution and
can be eliminated from consideration a priori. A concise mathematical explanation of the
eliminated routes can be found in O’Kelly et al. (1996).
Once a linearization was found that resulted in the attainment of exact solutions to
small problems, some researchers began focusing on different linearizations of the
problem. Treating interhub transfers as a multi-commodity flow problem, Ernst and
Krishnamoorthy (1996) devised a linearized variation of O’Kelly’s (1987) original
model. The immediate result was to reduce the number of variables and computation
time, but at the cost of integrality. However, integer programming must be used to obtain
optimal solutions so this model did not result in an increase in the size of the problem that
could be solved to optimality.
Despite the advances made by the linearized models and their ability to find exact
solutions to some problems, heuristics still have several advantages. Known
linearizations involve a tremendous number of variables that grow rapidly as the size of
the problem increases, restricting their use to small problems. On the other hand,
heuristics can generally handle much larger problems. In addition, they are more flexible in
the sense that additional aspects of hub networks may be incorporated into the problem while
still achieving near optimal solutions in a reasonable amount of time.

3. THE MULTIPLE ASSIGNMENT MODEL


The multiple assignment model allows non-hub nodes to interact with more than one
hub. Given the hub locations, every interacting pair selects the path that minimizes their
own total travel costs (this property is formally shown by Sohn and Park, 1995). This
model was first formulated by Campbell (1994b). A tighter linearization was later provided
by Skorin-Kapov, Skorin-Kapov, and O’Kelly (1996). See Appendix 2 for a mathematical
formulation of the linearized multiple assignment model.
The single assignment model may be seen as a special case of the more general
multiple assignment model. That is, the optimal solution to a multiple assignment model
may result in single allocations for all nodes. For example, when travel across the
interhub links is free, both the single and the multiple assignment models generate the
same network design (O’Kelly et al., 1996).
Variations on this linearization have been proposed. For example, Ernst and
Krishnamoorthy (1998) have shown how the multiple assignment problem may be
modeled as a multi-commodity flow problem. Whereas their linearization results in
fractional solutions when relaxed, it does significantly reduce the number of variables,
allowing the use of effective heuristics on large-size problems. The authors proposed one
such heuristic, based on the least-cost path property of the multiple assignment model,
which results in good (though not generally optimal) solutions.1
Klincewicz (1996) developed a heuristic for the multiple assignment model based on
the dual ascent and dual adjustment techniques frequently used for uncapacitated facility
location problems. When applied to conventional facility location problems these
techniques are preferred over most other heuristics because they provide optimal or near-
optimal solutions to most problems (Klincewicz, 1996). However, when applied to the
hub location problem these heuristics are not as successful. Part of the reason for the
disappointing and unexpected results may be due to the particular formulation used.
Klincewicz used Campbell’s (1994b) linearization of the hub location problem. When
com- paring results from the hub location problem to those from facility location
problems, Klincewicz notes that the relaxation of the facility location problem is quite
tight, frequently resulting in integer solutions whereas Campbell’s formulation does not.
He suggests that the lack of a tight linearization may have contributed to the disappointing
results. Perhaps more reliable results may be obtained by using a tighter linearization of the
hub location problem. However, use of these heuristics is further limited for hub location
problems because they can only solve very small problems—smaller than those that may be
solved using exact techniques.

4. EXTENSIONS WITH FIXED HUB LOCATION


In the models discussed so far, the basic multiple and single assignment models are
relatively unchanged in their ability to portray actual hub-and-spoke networks. However,
some researchers have devoted their attention to incorpo- rating more realistic
characteristics of hub networks into the model. Due to its complexity, the model is often
simplified by assuming fixed hub locations before the extensions are introduced. This
allows an examination of the effects of

1
Note that as this paper entered its final draft, the authors received a preprint of a paper (Ernst and
Krishnamoorthy) which makes improvements to the size of problem that can be tackled, using a novel
imbedding of a shortest path problem to obtain lower bounds, and then goes on to work out an efficient branch
and bound approach.
complex characteristics on allocation patterns. In the following paragraphs we discuss
several such extensions including the use of direct links between non-hub nodes, the
adaptation of the hub location problem for express package delivery systems, network
design when profits are maximized, and a study of congestion at the hubs.
Variations on the pure hub-and-spoke network allow some interacting pairs to interact
directly with each other even though they are not hubs. Other variants specify that hubs
cannot be completely interconnected (thereby reduc- ing the number of interhub links).
O’Kelly and Miller (1994) present a classifi- cation scheme for hub-and-spoke networks
based on these variations. Their scheme reflects the reality of the differences between
types of distribution systems. Different systems require different network designs based
upon their particular characteristics (O’Kelly, 1998a). For example, express package deliv-
ery systems guarantee delivery within a narrow time window—generally one or two
days. Time limits place a constraint on the distance traveled by each package and must be
factored into network design. The design of networks for passenger airlines produces
different obstacles; consumers travel across the actual routes and if the route deviates too
much from a direct path the consumers may opt to use a different airline.
Jeng (1987) considered hub-and-spoke networks in which some non-hub nodes are
permitted to bypass the hubs and interact directly with each other. He studied the
relationship between airline network routing strategies and key network parameters—
demand level, the size of the region served, and the number of nodes in the network—for
a single fixed hub network. Considering both one-stop and nonstop (direct path between
two non-hub nodes) routing strategies, he found that the number of nonstop routes
increases as demand increases. The number of nodes in the network and the size of the
region were not found to have a significant effect on network design. Consideration of these
key parameters shows that a mixed routing strategy (both one-stop and nonstop routes) is
generally more cost-efficient than a pure hub-and-spoke network (one-stop routes only).
Miller (1990) was concerned that the restriction embedded in the definition of a hub
(which prohibits any non-hub node from acting as a transshipment point) could be
violated when direct routes between non-hub nodes are permit- ted. He developed an
algorithm, based on the traditional Floyd routing algo- rithm, which prevents non-hub
nodes from receiving through traffic. The algorithm allocates flows to routes in a hub-
and-spoke network with fixed hub locations. Direct non-hub to non-hub links are
allowed, though stopovers or feeders are not. Four different costs are considered in this
model: (a) travel costs,
(b) arc construction costs, (c) hub construction costs, and (d) hub congestion costs.
The relaxation of the assumption that all non-hub nodes must interact via a hub is
also examined by Flynn and Ratick (1988). They evaluated options for the Essential Air
Services program which provides continued air service to small communities within a
modified hub-and-spoke network. Stopovers, where a
route will include one or more stops at small cities before reaching a hub, as well as direct
flights to the hubs are considered. Recognizing the need to provide maximum access to
the air transportation system for these small communities while simultaneously minimizing
the airlines’ costs, the Flynn and Ratick (1988) model includes both objectives.
Kuby and Gray (1993) expanded the work of Flynn and Ratick (1988) by allowing
the network to include both stopovers and feeders. Feeders typically serve small cities.
On such routes, small planes transfer their loads to larger planes at intermediate cities and
the larger planes continue on to the hubs. Within the context of express package delivery
systems, Kuby and Gray (1993) simultaneously considered the trade-offs between load
factors, economies of scale, time constraints, and distance. A mixed integer program with the
objective of minimizing total network costs was applied to the western U.S. portion of the
Federal Express package collection system. They designed a single hub network with the hub
location determined exogenously. They found that by allowing both stopovers and feeders,
network costs are lowered, load factors increase, and fewer miles are flown with fewer
aircraft than with the pure hub-and-spoke network originally designed by O’Kelly (1987).
Express package delivery systems, such as Federal Express, typically utilize mini or
regional hubs in addition to a single main or major hub. This type of hub-and-spoke
network was modeled by Hall (1989) and O’Kelly and Lao (1991). Hall (1989) examined the
effect of express package delivery time restrictions on network design in the United States.
He found that when the service area is large, time constraints prevent the use of two-hub
routing. Based on his analysis of routing strategies Hall suggested the use of a major hub to
handle the flow between regions and smaller hubs to serve as regional terminals for local
flows. Significant in the design of a major and mini hub model is the absence of the
interhub links; time constraints prevent the use of such links. O’Kelly and Lao (1991)
extended this work by also incorporating mode choice into the decision process. Given
fixed hub locations, they designed a network with one major hub and one regional hub for
several cities in the United States taking both time constraints and modal choice into
consideration. In order to minimize sorting requirements all cities connected to the
regional hub must interact with each other via that hub (this is known as the ‘club-
membership rule’). Allocations are determined such that total network costs are minimized
within the constraints of the model. They found that the allocations are sensitive to the
delivery schedule and that the integration of air with trucking services lowers total
network cost.
Daskin and Panayotopoulos (1989) addressed the problem of assigning aircraft to
routes with the objective of maximizing profits. They implemented a heuristic based on
Lagrangian relaxation and found an upper bound on the profits. In this model each route
originates at a single hub, visits one or more other cities (usually one), and then returns to
the hub. Dobson and Lederer (1993) extended the work of Daskin and Panayotopoulos
(1989) by including
consumer choice in the model. They explored the competitive choice of flight schedules
and prices in a hub-and-spoke network with the objective of profit maximization. Their
model assumes that the airline operates within a pure hub-and-spoke network with a
single fixed hub. For each route, consumer demand was a function of service quality and
prices of all routes, where service quality depends on the time of the flight and its length. A
heuristic was used to determine an airline’s schedule and prices in a competitive
environment with the goal of maximizing the airline’s profits. They illustrated their
method with a small example.
Grove and O’Kelly (1986) analyzed the relationship between hub-and-spoke networks
and congestion, or schedule delay. In this analysis they assumed a single assignment hub-
and-spoke network with fixed hub locations. Simulation of daily operations showed that the
amount of the locally originating demand is a crucial determinant of schedule delay. As
local demand increases, both the required total number of connections and the potential
for schedule delay decreases. Interestingly, eastern hubs appeared to have less schedule delay
than western hubs, due, according to the authors, to the configuration of U.S. time zones.
They also found less schedule delay with multiple hub networks than with single hub
systems.

5. EXTENSIONS WITH ENDOGENOUS HUB LOCATION


In the previous section, the basic model was extended by incorporating additional
characteristics of actual hub-and-spoke networks, while holding hub locations fixed. However,
one must keep in mind that optimal allocations depend on the hub locations, and the optimal
location of the hubs depend on allocation decisions. Results obtained when hub locations
are fixed, then, are not necessar- ily upheld when hub locations must also be determined.
Recognizing this, several researchers have developed models that incorporate important
additional char- acteristics of hub networks within the location-allocation framework. The
use of direct links between non-hub nodes and the adaptation of the hub location problem
for express package delivery systems developed for fixed hub location models have been
extended to the endogenous location case. In addition, models have been developed that
include fixed costs for establishing hubs, capacity constraints to reduce congestion at the
hubs, minimum threshold constraints to prevent links from being underutilized, and an
improved cost function for the interhub links. Models with various objectives have also
been formulated. A different approach to the hub location problem has been undertaken
by Jaillet, Song, and Yu (1996) who designed a network in which no a priori hub-and-spoke
structure is assumed.
In the basic hub location model, the number of hubs (p) is determined exogenously.
By accounting for the fixed costs that are incurred when estab- lishing hubs, p may be
made endogenous. The model then determines how many hubs to open, the number being
directly dependent on the fixed cost of opening hubs. As the fixed cost increases the
number of hubs in the network will decrease.
Several researchers have incorporated fixed costs into the model, including O’Kelly
(1992), Campbell (1994b), and Aykin (1995).
Another common extension was prompted by the amount of flow arriving at the
hubs. Many researchers noted that some links were transporting ex- tremely large amounts
of flow, resulting in congestion problems at the hubs. In practice, this is more of a problem for
passenger airlines than package delivery systems (Kuby and Gray, 1993). In an effort to curb
this congestion, a constraint that prohibits the flow entering a hub from exceeding a certain
level, its capacity, is introduced into the model. Aykin (1994) was first to formulate the
capacitated network model. Unfortunately, with endogenous hub locations the capacity
constraint is nonlinear. His efforts were followed by Ernst and Krishnamoorthy (Forthcoming
a), who developed a linear capacity constraint for the single assignment endogenous hub
location model. More recent work has developed a capacitated multiple allocation hub
location problem (Ebery et al., 1998).
Although solutions for these models were found using heuristics, neither Aykin
(1994) or Ernst and Krishnamoorthy (Forthcoming a) fully examined the resultant network
design. However, it would be expected that as the capacity for a hub decreases, the number
of multiple allocations will increase in order to spread out the flow among all hubs more
equitably. An alternative to this capacity constraint would be one where the amount of
flow traveling across the interhub links (where the largest bundling of flows occurs) is
constrained to be less than a certain amount. This type of capacity constraint would recognize
the limits of a network to ship extremely large amounts of flow, all destined for the second
hub, across the same link.
Further examination of the amount of flow traveling across the links in the multiple
assignment model reveals that some spokes carry very little flow. Campbell (1994b)
suggested that the multiple assignment model should be extended to include flow
thresholds and fixed costs for the spokes. Careful choice of these parameters would alleviate
the problem of underutilized spokes. The flow threshold would require a minimum flow
across the spoke for it to be utilized. The fixed costs for the spokes recognizes that the
utilization of a link incurs a cost and this cost should be accounted for in the model.
Although Campbell (1994b) formulated this model, no solutions were provided.
In the basic hub location model demand is assumed to occur only at discrete points. That
is, it is assumed that flows originate only at selected points in the network (the nodes).
Campbell (1993) relaxed this assumption by assuming demand to be uniformly
distributed over the service region and developing an approximation formula to calculate
transportation costs. He compared his results to those obtained with discrete demand and
found that the approxima- tion formula provides a good indicator of optimal
transportation costs.
Chou (1990) relaxed the assumption that the hubs are completely intercon- nected. He
designed a hub location model, called HIERHUB, that results in a minimum-cost
spanning tree. His model determines the number of hubs endo- genously. Recognizing that
the HIERHUB model can be seen as an intermediate
step towards a multiple assignment network, Chou suggested ways to upgrade HIERHUB to
a multiple assignment network by placing capacities on the links to encourage multiple
allocations.
Hub location models have been designed for express package delivery systems when
hub locations are fixed (Kuby and Gray, 1993; Hall, 1989; O’Kelly and Lao, 1991). O’Kelly
(1998b) also considered the use of major and mini hub networks for package delivery
systems when hub locations must also be deter- mined within the model. His work
improved on previous efforts by: (1) making the choice of hub locations endogenous to the
model; (2) relaxing the restriction that all cities assigned to the mini-hub must interact with
each other only via the mini-hub (the ‘club membership rule’); (3) designing a network
with more than one mini-hub; and (4) exploring the geometric properties of the model.
Modal choice is not explicitly considered. Instead, one extension to the model restricts
the mini-hubs’ service area to a maximum radius that can be serviced within the
allowable time span. O’Kelly found that time constraints prevent certain interacting pairs
from using their least-cost path. That is, by guarantee- ing quick delivery, express package
delivery systems are forced to incur larger transportation costs than if no such guarantees
were made. However, these transportation costs decrease when the club membership rule
is relaxed.
Despite empirical evidence that travel costs in hub networks are a function of the
amount of flow traveling across individual links, the basic hub location model uses an
exogenously-determined interhub discount factor to account for the scale economies—
giving the same discount to all interhub links regardless of flow volume. O’Kelly and
Bryan (1998) presented a model in which costs are dependent on flows. They showed that, by
simplifying interhub travel costs and assuming that these costs are independent of flows,
the basic model not only miscalculates total network cost but may also erroneously select
optimal hub locations and allocations. Through the use of a piecewise linear cost function
their formulation more explicitly models the scale economies that are generated on the
interhub links, and in doing so provides a more reliable model repre- sentation of the
reality of hub-and-spoke networks.
An examination of different routing models and their implications was conducted by
O’Kelly (1998a). When interhub costs are a function of flow, total network cost is often
minimized by forcing some individual interacting pairs to use their non–least-cost path.
Comparing the resultant network designs for the basic multiple assignment model and one
with a flow-dependent cost function, he found that passenger inconvenience (in terms of
individual travel times) made the network design with the flow-dependent model
inappropriate for passenger airlines. However, this model was found to be appropriate for
air freight as the opportunity to maximize load factors regardless of routing gives the
carrier every incentive to make allocation decisions based on the ability of flows to
capture scale economies (O’Kelly, 1998a). This suggests that the model developed by
O’Kelly and Bryan (1998) should be modified when designing
passenger airlines. A compromise must be made between the minimization of passenger
inconvenience and the minimization of total network cost.
Aykin (1995) considered the design of a hub-and-spoke network that allows non-hub to
non-hub connections as well as one- and two-stop routes. This is an extension of similar
research for fixed hub locations (Jeng, 1987; Miller, 1990). Recognizing that bundling of
flows occurs on the spokes as well as on the interhub links, Aykin (1995) included an
exogenously-determined discount factor for the spokes to reflect the economies of scale
earned on those links but not for direct non-hub to non-hub routes. A heuristic based on
Lagrangian relaxation was developed to solve this model. Comparing the results to those for
the single assignment model (with discounted spokes) it was found that allowing direct routes
between non-hub nodes means that total network cost is always less than or equal to the
total network cost for the single assignment model.
Extending the work of Aykin (1995), Jaillet, Song, and Yu (1996) presented a unique
approach to hub-and-spoke network design. They formulated a model in which no a priori
hub-and-spoke structure is assumed. If cost efficient the resulting network will suggest
the presence of hubs, defined here as “strong connecting cities.” In addition, the type and
number of aircraft used on each link is determined by the model. Using heuristics to find
good solutions to their model, they found that the ‘hubs’ tend to be located in the center
of the region even when more than one ‘hub’ emerges. As the demand increases the
number of nonstop flights also increases resulting in a network design even more
dissimilar to networks designed using the basic model. Caution must be ex- tended when
comparing results from this analysis to ones using the basic models (or their extensions) as
the definition of a ‘hub’ is not the same. Most studies require a hub to serve as a
transshipment point and only allow hubs to serve in this capacity. Jaillet, Song, and Yu did
not place this restriction on their model. Rather ‘hubs’ are defined as any city that receives
a large amount of flow. Any city in the network is allowed to serve as a transshipment point,
even if it is not designated as a ‘hub.’
These extensions have increased our ability to represent actual hub-and- spoke
networks by relaxing some of the assumptions in the basic model. In these extended models
the objective remains the minimization of total network cost. However, other goals are also
important in network design, and research efforts have been directed towards designing
networks with various objectives. For example, O’Kelly (1986) considered two different
congestion-based objectives for network design that measure hub usage (number of arrivals
and departures). He found that by minimizing variability of hub usage the resultant
design provides an equitable allocation of nodes to hubs. When the objective is the
minimization of total hub usage, demand tends to be concentrated quite heavily in one
central facility.
The minimization of total network costs may not be appropriate for all applications.
Campbell (1994b) considered two alternative objectives and devel- oped integer
programming formulations of the models, but no solutions are
presented. In hub center problems the objective is to minimize the maximum distance
traveled by any interacting pair of nodes (either for the entire route or for individual
links). This type of network would be appropriate when the maximum travel time needs
to be as small as possible. Examples include the location of emergency vehicle stations
(when the network design would need to minimize the total travel time for the entire route)
or when there are time limits on the length of individual links (for example, when pilots are
restricted to a time limit on continuous service—the length of each individual link would
be the concern here).
Hub-covering problems were also considered by Campbell (1994b). Perhaps the most
promising application of hub-covering problems is in the express package delivery system
where shipments must be made within a specified time frame. Hub-covering problems
locate hubs to cover all demand such that the cost for the hubs is minimized. A node is
considered covered by a hub if the cost of the node’s path via that hub is less than a
specified amount or alternatively, if the flow between an origin and destination can travel
the path via the assigned hub within a specified period of time. In terms of package delivery
networks an interacting pair would be covered by a hub only if timely delivery of their
packages can be made.

6. SENSITIVITY ANALYSIS OF HUB-AND-SPOKE NETWORKS


The development of a tight linearization and an exact solution method allows
sensitivity analysis of hub-and-spoke networks to be conducted. An evaluation of the
effects on network design of changes in key parameters is the focus of the research by
O’Kelly et al. (1996). Of primary concern is the effect of changes in the interhub discount
factor. They found that with the multiple assignment model an increase in the number of
multiple allocations occurs as the interhub discount decreases. The single assignment
model cannot employ this cost-saving strategy and may only respond to changes in the
interhub discount by (1) changing hub locations or (2) changing a node’s allocation
(requiring all of that node’s interactions to use the newly allocated hub). These responses
do decrease costs but they are not as cost efficient as the responses available to the
multiple assignment model. As a result, the difference in total cost between the two
models increases as the interhub discount increases. O’Kelly et al. (1996) also examined
the effect of changes in the interhub discount factor on the optimal number of hubs p to
open when fixed costs are taken into consideration. They determined the range (for a
given network) of fixed costs and interhub discounts for which a specific number of hubs
should be opened. They found that, holding fixed costs constant, as the interhub discount
increases the optimal number of hubs increases. Interestingly, they found it is possible
that a specific value of p is never optimal for a certain interhub discount regardless of the
size of the fixed cost for opening the hubs. This indicates that when designing a hub-and-
spoke network one must consider not only how many hubs to open for a given fixed cost and
interhub discount but also how the optimal
number of hubs changes as either the fixed cost or the interhub discount (or both)
changes.

7. SUMMARY
Since the hub location model was first formulated in 1987 a great deal of research
has been conducted on this interesting problem, and many advances have been made.
Researchers now have a tight linearization to solve small-size problems (up to 25 nodes) and
reliable heuristics to solve much larger, realisti- cally-sized problems. Ernst and
Krishnamoorthy (Forthcoming b) have also extended the size of problems that can be
solved exactly by recognizing an imbedded shortest path problem and completing the
analysis using branch and bound. Advances have also been made in relaxing the assumptions
of the original model, resulting in a more realistic abstraction of actual hub-and-spoke net-
works.
Despite the many advances, further work remains. Research efforts need to be devoted
to developing more reliable heuristics for the multiple assignment model. The most
reliable heuristics for the single assignment model, such as tabu search, have not been
modified to include many of the extensions currently found in the literature. Until reliable
heuristics are found that can solve more complicated models—those that include additional
characteristics of actual hub networks—the use of the models for designing actual
networks is limited. However, we anticipate that a number of energetic research teams
working on these and related problems will soon make further gains.
Several extensions to the basic model have been formulated. The purpose of these
extensions is to improve modeling capabilities by incorporating impor- tant characteristics
of actual networks into models. However, in almost all cases the extensions are only
relaxing one assumption at a time. Additional work should be focused on integrating the
extended models into a larger one that would include several different characteristics of
actual hub-and-spoke net- works within the same model.
Additional research needs to be conducted on understanding the fundamen- tals of the
hub location problem. As a result of the tight linearization it is now possible to solve this
problem to optimality. However, due to the relaxation an integer solution cannot be
guaranteed; fragmentation of hubs cannot always be prevented. These noninteger solutions
appear as the parameters are varied and although they are not common they are quite
persistent. One key avenue of research is to investigate the conditions under which the
model will tend to have integer or noninteger solutions. This may lead to an even tighter
formulation of the problem. Research in this area will increase understanding of the
model’s behavior and contribute to the solution of a longstanding puzzle about the
allocation of indivisible resources. Further understanding of the model’s behav- ior may also
be gained by more sensitivity analyses; to date, this has been a fairly neglected area in
hub research.
REFERENCES
Abdinnour-Helm, Sue and Munirpallam A. Venkataramanan. 1992. “Using Simulated Annealing to Solve the p-
Hub Location Problem,” Working Paper No. 9304, IRMIS.
———. 1993. “Solving the Hub Location Problem Using Genetic Research,” Working Paper No. 9303, IRMIS.
Aykin, Turgut. 1990. “On ‘A Quadratic Integer Program for the Location of Interacting Hub Facilities’,”
European Journal of Operational Research, 46, 409–411.
———. 1994. “Lagrangean Relaxation Based Approaches to Capacitated Hub-and-Spoke Network Design
Problems,” European Journal of Operational Research, 79, 501–523.
———. 1995. “Networking Policies for Hub-and-Spoke Systems with Applications to the Air Trans-
portation System,” Transportation Science, 29, 201–221.
Campbell, James F. 1993. “Continuous and Discrete Demand Hub Location Problems,” Transporta- tion
Research B, 27, 473–482.
———. 1994a. “A Survey of Network Hub Location,” Studies in Locational Analysis, 6, 31–49.
———. 1994b. “Integer Programming Formulations of Discrete Hub Location Problems,” European Journal of
Operational Research, 72, 387–405.
———. 1996. “Hub Location and the p-Hub Median Problem,” Operations Research, 44, 923–935. Chou, Yue-
Hong. 1990. “The Hierarchical-Hub Model for Airline Networks,” Transportation Plan-
ning and Technology, 14, 243–258.
Daskin, Mark S. and Nicholaos D. Panayotopoulos. 1989. “A Lagrangian Relaxation Approach to Assigning
Aircraft to Routes in Hub-and-Spoke Networks,” Transportation Science, 23, 91–99.
Daskin, Mark S. 1995. Network and Discrete Location. New York: Wiley Interscience.
Dobson, Gregory and Phillip J. Lederer. 1993. “Airline Scheduling and Routing in a Hub-and-Spoke System,”
Transportation Science, 27, 281–297.
Ebery, Jamie, Mohan Krishnamoorthy, Andreas Ernst, and Natashia Boland. 1998. “The Capacitated
Multiple Allocation Hub Location Problem: Formulations and Algorithms,” CSIRO Division of Mathematics
and Statistics, and the Department of Mathematics and Statistics, University of Melbourne (unpublished
manuscript).
Ernst, Andreas and Mohan Krishnamoorthy. 1996. “Efficient Algorithms for the Uncapacitated Single
Allocation p-Hub Median Problem,” Location Science, 4, 139–154.
———. 1998. “Exact and Heuristic Algorithms for the Uncapacitated Multiple Allocation p-Hub Median
Problem,” European Journal of Operational Research, 104, 100–112.
———. Forthcoming a. “Solution Algorithms for the Capacitated Single Allocation Hub Location Problem,”
Annals of Operations Research.
———. Forthcoming b. “An Exact Solution Approach Based on Shortest-Paths for p-Hub Median Problems,”
INFORMS Journal on Computing.
Flynn, John and Samuel Ratick. 1988. “A Multiobjective Hierarchical Covering Model for the Essential
Air Services Program,” Transportation Science, 22, 139–147.
Grove, Peter G. and Morton E. O’Kelly. 1986. “Hub Networks and Simulated Schedule Delay,” Papers of the
Regional Science Association, 59, 103–119.
Hall, Randolph W. 1989. “Configuration of an Overnight Package Air Network,” Transportation Research A, 23,
139–149.
Jaillet, Patrick, Mark Song, and Gang Yu. 1996. “Airline Network Design and Hub Location Problems,”
Location Science, 4, 195–212.
Jeng, Chawn-Yaw. 1987. “Routing Strategies for an Idealized Airline Network,” unpublished Ph.D.
dissertation, University of California–Berkeley.
Klincewicz, John G. 1991. “Heuristics for the p-Hub Location Problem,” European Journal of
Operational Research, 53, 25–37.
———. 1992. “Avoiding Local Optima in the p-Hub Location Problem Using Tabu Search and Grasp,”
Annals of Operations Research, 40, 283–302.
———. 1996. “A Dual Algorithm for the Uncapacitated Hub Location Problem,” Location Science, 4, 173–
184.
Kuby, Michael J. and Robert Gray. 1993. “The Hub Network Design Problem with Stopovers and Feeders: The
Case of Federal Express,” Transportation Research A, 27, 1–12.
Miller, Harvey J. 1990. “Route 6: Routing Algorithm for Hub Networks,” unpublished manuscript.
O’Kelly, Morton E. 1986. “Activity Levels at Hub Facilities in Interacting Networks,” Geographical Analysis, 18,
343–356.
———. 1987. “A Quadratic Integer Program for the Location of Interacting Hub Facilities,” European Journal
of Operational Research, 32, 393–404.
———. 1992. “Hub Facility Location with Fixed Costs,” Papers in Regional Science: The Journal of The
Regional Science Association International, 71, 293–306.
———. 1998a. “A Geographer’s Analysis of Hub-and-Spoke Networks,” Journal of Transport Geog- raphy, 6,
171–186.
———. 1998b. “On the Allocation of a Set of Nodes to a Minihub in a Package Delivery Network,” Papers in
Regional Science: The Journal of The Regional Science Association International, 77, 77–99.
O’Kelly, Morton E. and Yong Lao. 1991. “Mode Choice in a Hub-and-Spoke Network: A Zero–One Linear
Programming Approach,” Geographical Analysis, 23, 283–297.
O’Kelly, Morton E. and Harvey J. Miller. 1994. “The Hub Network Design Problem: A Review and
Synthesis,” Journal of Transport Geography, 2, 31–40.
O’Kelly, Morton E., Darko Skorin-Kapov, and Jadranka Skorin-Kapov. 1995. “Lower Bounds for the
Hub Location Problem,” Management Science, 41, 713–721.
O’Kelly, Morton E., Deborah Bryan, Darko Skorin-Kapov, and Jadranka Skorin-Kapov. 1996. “Hub Network
Design with Single and Multiple Allocation: A Computational Study,” Location Science, 4, 125–138.
O’Kelly, Morton E. and Deborah Bryan. 1998. “Hub Location with Flow Economies of Scale,”
Transportation Research B, 32, 605–616.
Skorin-Kapov, Darko and Jadranka Skorin-Kapov. 1994. “On Tabu Search for the Location of Interacting
Hub Facilities,” European Journal of Operational Research, 73, 502–509.
———. 1995. “On Hub Location Models,” Journal of Communication and Information Technology,
3, 183–192.
Skorin-Kapov, Darko, Jadranka Skorin-Kapov, and Morton E. O’Kelly. 1996. “Tight Linear Program- ming
Relaxations of Uncapacitated p-Hub Median Problems,” European Journal of Operational Research, 94,
582–593.
Smith, Kate, Mohan Krishnamoorthy, and M. Palaniswami. 1996. “Neural Versus Traditional Approaches to
the Location of Interacting Hub Facilities,” Location Science, 4, 155–171.
Sohn, Jinhyeon and Sungsoo Park. 1995. “On Uncapacitated p-Hub Location Problems,” Department of
Industrial Engineering, Korea Advanced Institute of Science and Technology (unpublished manuscript).
———. 1997. “The Single Allocation Problem in the Interacting Three-Hub Network,” Department of
Industrial Engineering, Korea Advanced Institue of Science and Technology (unpublished manuscript).

APPENDIX 1
The Quadratic Single Assignment Model
F I
(1) Min W G  Z ik Cik  Z jm C jm   Z
Z C JH
ij ik

jm km i j k m k m
s.t.

(2)
bn  p  1gZkk   Zik  0 k
i

(3) (4)
where Z
k
ik 1 i

Z
k
kk  p

Zik  0,1 l q i, k

n = the number of nodes in a network


p = the number of hubs to be located
 = the interhub discount factor 0    1
Wij = the amount of flow traveling between i and j Cik =
the per unit cost of traveling between i and k Zik = 1 if
node i is allocated to hub k
0 otherwise
The objective function (1) minimizes total network cost. Constraint (2) requires a hub to be
open before a node is assigned to it. Constraint (3) constrains each node to be assigned to
a single hub. Constraint (4) requires that p hubs be open.

APPENDIX 2

The Linearized Multiple Assignment Model


(5) Min  W dC ij ik i
 Ckm  Cmj X ijkm
i j m
k
s.t.

(6)
Zp
k
k

(7)  k m
Xijkm  1 i, j

(8)
 Xijkm  Zk  0
m
i, j, k

(9)
X
k
ijkm  Zm  0 i, j, m
where
Xijkm = the fraction of flow from origin i to destination j that is routed via hubs
k and m in that order
Zk = 1 if k is a hub and 0 otherwise. All
other notation is as before.
The objective function (5) minimizes total network cost (as in the quadratic model).
Constraint (6) requires that p hubs be open. Constraint (7) ensures every interacting pair (i,j)
is allocated to a path via hubs k and m. Taken together, constraints (8) and (9) guarantee
that the O–D flow will not be routed via k and m unless both k and m are in fact hubs.

You might also like