Anda di halaman 1dari 9

Omega 57 (2015) 203–211

Contents lists available at ScienceDirect

Omega
journal homepage: www.elsevier.com/locate/omega

An improved MIP heuristic for the intermodal hub location problem$


Yan He a,n, Tao Wu b, Canrong Zhang c, Zhe Liang d
a
State Key Laboratory of Mechanical Transmission, Chongqing University, Chongqing 400030, China
b
Business Analytics and Optimization, University of Phoenix, Apollo Group Inc., Phoenix 85040, AZ, USA
c
Logistics Engineering and Simulation Laboratory, Graduate School at Shenzhen, Tsinghua University, Shenzhen 518055, China
d
School of Economics & Management, Tongji University, Shanghai 200092, China

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

Article history: Optimization methods have been commonly developed for the intermodal hub location problem
Received 3 November 2014 because it has a broad range of practical applications. These methods include exact methods (limited
Accepted 28 April 2015 on solving large-size problems) and heuristics (no guarantee on solution quality). In order to avoid their
Available online 6 May 2015
weakness but to leverage their strength, we develop an improved MIP heuristic combining branch-and-
Keywords: bound, Lagrangian relaxation, and linear programming relaxation. In the heuristic, we generate a
Heuristics population of initial feasible solutions using the branch-and-bound and Lagrangian relaxation methods
Optimization and create a linear-relaxed solution using the linear programming relaxation method. We combine these
Mixed integer programming feasible and linear-relaxed solutions to fix a portion of hub location variables so as to create a number of
Lagrangian relaxation
restricted hub location subproblems. We then combine the branch-and-bound method to solve these
Hub location
restricted subproblems for iteratively improving solution quality. We discuss in detail the application of
the method to the intermodal hub location problem. The discussion is followed by extensive statistical
analysis and computational tests, where the analysis shows statistical significance of solutions for
guiding the heuristic search and comparisons with other methods indicate that the proposed approach is
computationally tractable and is able to obtain competitive results.
& 2015 Elsevier Ltd. All rights reserved.

1. Introduction a branch-and-bound scheme for the multiple-allocation case of the


p-hub median problem and showed how a similar method can be
The hub location problem has important applications at prac- adapted for solving the more difficult single-allocation case. Mayer
tice, such as telecommunication systems [29,33], transportation and Wagner [36] proposed a branch-and-bound procedure called
[45,5] and parcel delivery [19,22]. The problem is challenging and hubLocator for the uncapacitated multiple allocation hub location
computationally intractable by explicit enumeration due to a problem. Labbe et al. [35] proposed a branch-and-cut algorithm for
combinatorially increasing number of potential solution alterna- the hub location problem with single assignment. Camargo et al.
tives. This attracts an extensive amount of research being con- [12] developed a Benders decomposition algorithm for the uncapa-
ducted for finding effective solution methodologies [13,14,38], citated hub location problem with multiple assignments in tele-
such as exact methods and heuristics. communication and transportation systems. Contreras et al. [17]
Exact methods are guaranteed to achieve proven optimal solu- proposed a branch-and-price algorithm for the capacitated hub
tions under sufficient computation time. In the literature, Klince- location problem with single assignment, in which Lagrangean
wicz [32] proposed a dual ascent based method within a branch- relaxation is used to obtain tight lower bounds of the restricted
and-bound scheme for the uncapacitated hub location problem. master problem. Contreras et al. [16] presented an exact algorithm
Ernst and Krishnamoorthy [20] proposed a linear programming (LP) capable of solving large-scale instances of the uncapacitated hub
based method as well as an explicit enumeration algorithm for the location problem with multiple assignments. The algorithm applies
uncapacitated multiple allocation p-hub median problem. Ernst and Benders decomposition to a strong path-based formulation of the
Krishnamoorthy [21] developed a novel exact-solution approach in problem. Camargo et al. [11] developed a specially tailored Benders
decomposition algorithm for the vehicle routing problem and the

hub location problem with single assignment. Sa et al. [42]
This manuscript was processed by Associate Editor Yagiura.
n proposed a Benders decomposition method for the tree of hub
Corresponding author.
E-mail addresses: heyan@cqu.edu.cn (Y. He), danielwu9999@gmail.com (T. Wu), location problems. The exact methods are efficient at solving
crzhang@sz.tsinghua.edu.cn (C. Zhang), liangzhe@tongji.edu.cn (Z. Liang). relatively small problems. However, they are limited on thoroughly

http://dx.doi.org/10.1016/j.omega.2015.04.016
0305-0483/& 2015 Elsevier Ltd. All rights reserved.
204 Y. He et al. / Omega 57 (2015) 203–211

searching all possible branches and finding an optimal solution solution of superior quality. The IMMIP heuristic has similarity with
under a reasonable amount of computational time when the some neighborhood search heuristics in the literature, including
problem size is particularly large. iterative production estimate (IPE) heuristic [41], feasibility pump
Heuristics are not guaranteed to achieve proven optimal solu- [24,2,8,26,10], relaxation induced neighborhood search (RINS) [18],
tions but aim to find good feasible solutions in a reasonable pivot-and-complement [6], local branching [25], relaxation enforced
amount of computational time. Different from exact methods, neighborhood search (RENS) [9], scatter search [27], and restrict and
heuristics intuitively search for a subset of all solution alternatives relax [28].
when the exhaustive whole search is impractical. In the literature, The remainder of this paper is organized as follows: In Section 2,
Pirkul and Schilling [40] presented a subgradient optimization we compare the IMMIP heuristic with other similar heuristics in the
method based on Lagrangian relaxation for the hub location literature. In Section 3, we describe the mathematical formulation
problem. Abdinnour-Helm [1] developed a hybrid heuristic that for the IHL problem. In Section 4, we present the IMMIP heuristic
integrates genetic algorithms and tabu search for the uncapaci- for the IHL problem. Section 5 presents statistical analysis of the
tated hub location problem. Klincewicz [34] proposed search application. Section 6 shows extensive computational results and
heuristics that are based on tabu search and greedy random comparisons of the IMMIP heuristic with other approaches. Finally,
adaptive search procedures for the hub location problem with we conclude with future directions in Section 7.
economies of scale. Arnold et al. [4] proposed a heuristic approach
for the rail/road transportation system in the Iberian Peninsula.
Topcuoglu et al. [46] presented a genetic search heuristic for the 2. The improved MIP heuristic and other similar methods
uncapacitated hub location problem with single allocation. Chen
[15] proposed a hybrid heuristic based on simulated annealing, This section will focus on comparing the improved MIP
tabu list, and improvement procedures for the uncapacitated hub heuristic with other similar methods, such as IPE, feasibility pump,
location problem with single allocation. Pi et al. [39] developed a RENS, RINS, and relax-and-fix.
new hybrid nested partitions and mathematical programming IPE is a heuristic that is similar with using the idea of “round-
approach for the hub location problem. Ishfaq and Sox [30] ing” the LP-relaxation solution to an integer solution. However,
developed a tabu search meta-heuristic for the multiple allocation claiming the value of the linear relaxation solution is not as such a
p-hub median problem. Meng and Wang [37] presented a hybrid good indicator of the integer optimal solution, Pochet and Vyve
genetic algorithm embedded with a diagonalization method for [41] proposed an idea which is to augment the model by adding
the intermodal hub-and-spoke network design problem with iteratively updated upper bounds for continuous variables so that
multiple stakeholders and multi-type containers. Sorensen et al. the new LP-relaxation solution is a better indicator of a good
[43] developed two different meta-heuristic procedures that integer solution. The heuristic continuously fixes integer variables
consist of two phases: a solution construction phase and a solution that have integer values in the LP-relaxation solution. For the
improvement phase based on local search for the intermodal remained integer variables that have fractional values, the relative
terminal location problem. Farahani et al. [23] provided a review continuous variables are updated with a smaller upper bound.
for the most recent advances of solution methodologies for the Such procedures are performed iteratively until all integer vari-
hub location problem. Other similar research includes [3,31]. ables are fixed to integer values.
In this paper, we study the intermodal hub location (IHL) Feasibility pump [24] is a heuristic scheme for finding a
problem that involves locating a set of fully interconnected feasible solution to generic mixed integer programming (MIP)
facilities called intermodal hubs. The hubs serve as transshipment problems of the form minfcT x : Ax Z b; xj integer 8 jA Jg. A feasi-
and switching points for moving flows between specified origins ble solution is defined as a point xn A P≔fx : Ax Z bg that is equal
and destinations. The transportation via intermodal hubs is gen- to its rounding x, ~ whereby the rounded point x~ is defined by
erally more efficient than the direct origin-to-destination trans- x~ j ≔½xnj  if jA J and x~ j ≔xnj otherwise, and ½ represents scalar
portation, as intermodal hubs have equipment suitable for rounding to the nearest integer. Feasibility pump starts from
transferring containers and truck trailers between different modes any xn A P, and defines its rounding x. ~ At each iteration, feasibility
that are connected via rails, while rail transportation becomes pump looks for a point xn A P that is as close as possible to the
more efficient than truck transportation when distances are over a current x~ by solving the distance problem minf△ðx; xÞ ~ : x A Pg. If
certain threshold and results in savings in time, operating costs △ðx; xÞ~ ¼ 0, then xn is a feasible MIP solution and feasibility pump
and labor force. The IHL problem consists of three connected is done. Otherwise, feasibility pump replaces x~ by the rounding of
movements: two local ones that involve truck movements and an xn, and repeats.
intermediate one that involves rail movement. The overall objec- RENS [9] is a neighborhood search heuristic for generic MIP
tive is to minimize the total costs of the system, which balance the problems. The method constructs a sub-MIP problem which
operating costs of intermodal hubs, transportation costs of flows represents the space of all feasible roundings of some fractional
and outsourcing costs when some of the flows are preferred to be point – normally the optimum of the LP-relaxation of the original
transported by other resources. MIP problem. Considering that some integer variables may
We propose an improved MIP (IMMIP) heuristic combining already take an integral value in the optimum of the LP-relaxa-
branch-and-bound, Lagrangian relaxation, and linear program- tion, the idea of the method is to fix these variables and perform
ming relaxation. The proposed IMMIP heuristic is designed to first a neighborhood search on the remaining variables, which are
generate an initial incumbent feasible solution and a linear rebounded to the nearest integers. Summarized, the sub-MIP
programming relaxation solution using branch-and-bound, problem is created by changing the bounds of all integer variables
Lagrangian relaxation, and linear programming relaxation. Then, xj to lj ¼ ⌊x j c and uj ¼ ⌈x j ⌉, where x is the optimum of the LP-
the heuristic utilizes domain knowledge derived from these relaxation.
solutions to iteratively develop subproblems through a neighbor- RINS [18] is also a neighborhood search heuristic for generic
hood search strategy. These subproblems have a number of hub- MIP problems. The method constructs a promising neighborhood
open variables fixed and therefore can be solved with relatively using information contained in the LP-relaxation and a feasible
effective solutions. The goal of the IMMIP heuristic is that, by incumbent solution point of the MIP model. While some variables
solving these subproblems, objective values of the original clearly take different values in the LP-relaxation and incumbent
problem can be iteratively improved and eventually reach to a solution point, it is important to note that many may take the
Y. He et al. / Omega 57 (2015) 203–211 205

same values. RINS is based on the intuition that the instantiation discount factor α for the inter-hub transportation. To present the
of these variables forms a partial solution that is likely to be problem formulation, we define the following notations:
extended towards a complete solution that achieves both inte- Sets:
grality and a good objective value. Therefore, RINS focuses att-
L set of origins and destinations, indexed by l
ention on those variables that differ in the LP-relaxation and in
H set of intermodal hubs, indexed by h. H ¼ ½1; j Hj 
the incumbent. RINS performs the following steps at a node of
F set of flows (demands), indexed by f
the global branch-and-cut tree: (i) fix the variables that have
A set of arcs, h1 h2 , lh, hl and l1 l2 A A
the same values in the incumbent and in the current LP-relaxa-
tion; (ii) set an objective cutoff based on the objective value of
the current incumbent; (iii) solve a sub-MIP on the remaining Parameters:
variables.
Relax-and-fix is a decomposition-oriented heuristic that usu- Of origin of flow f
ally constructs smaller subproblems by partitioning the set of Df destination of flow f
integer variables into three subsets. One subset is a window that wf amount of flow f
contains a number of indices for which all integer variables are tchl transportation cost of per-amount of flow between
remained as integer variables; another subset contains a number intermodal hub hand origin/destination l
of indices for which all integer variables have been fixed based on α discount factor for the inter-hub transportation
values of the previous iteration(s); and the last subset consists of hch1 h2 transportation cost of per-amount of flow between
indices for which all integer variables are relaxed as continuous intermodal hubs h1 and h2
variables. Such smaller restricted problems are solved by mixed och open cost of intermodal hub h
integer programming solver and the resulting solution value is vcf transportation cost of per-amount of flow fif the flow is
used to fix integer variables in the window, which will then roll outsourced to other operations such as pure truck
forward to the next few indices. The same manner is iterativ- movement
ely performed until the window rolls to the end and an integer cap limit of the amount of a flow moved through an inter-
value has been achieved for all integer variables. Relax-and-fix hub route
has been widely applied to solve production planning problems
[7,44,49,51,47].
The proposed IMMIP heuristic has similarity with the above Variables:
heuristics in that all methods create sub-MIP problems where a zfh1 h2 percentage of flow f moved through intermodal hubs h1
part of integer variables are fixed and solve these sub-MIP and h2
problems to obtain a feasible solution to the original MIP vf percentage of flow f outsourced to other operations
problem. However, the IMMIP heuristic is different from those yh binary hub-open decision variables: yh ¼ 1 if intermodal
heuristics in the way how solutions are used to determine the hub h is open, zero, otherwise
probability function for fixing integer variables to integer values,
and in the way how the solutions are used to build sub-MIP
problems. The IMMIP heuristic combines branch-and-bound, IHL:
Lagrangian relaxation, and linear programming relaxation to X X l1 X
h1 A A h1X
h2 A A h2X
l2 A A
obtain initial incumbent feasible solution and LP-relaxation min ðtcl1 h1 þ α
solution at first, and then use these solutions to build a prob- f A F l1 ¼ Of h1 A H h2 A H l2 ¼ Df
X X
ability function for fixing integer variables. The RINS method is hch1 h2 þ tch2 l2 Þ  wf  zfh1 h2 þ och  yh þ vcf  wf  vf ð1Þ
most similar with the IMMIP heuristic in terms of both using hAH f AF
incumbent feasible solution and LP-relaxation solution. However,
the RINS method always fixes integer variables that have the X h1X
h2 A A
same values on these two solutions. When the total number of Subject to : vf þ zfh1 h2 ¼ 1 8 f A F: ð2Þ
integer variables having the same values is small, the RINS h1 A H h2 A H

method has a risk of ending up with a sub-MIP problem that is


h1X
h2 A A
still large. The IMMIP heuristic provides flexibility of assigning
zfh1 h2 r yh1 8 f A F; h1 A H: ð3Þ
the number of integer variables to fix. The IMMIP heuristic is h2 A H
much more different from IPE, feasibility pump, and RENS in that
the three methods only utilize LP-relaxation solution to fix h1X
h2 A A
integer variables. The difference from the relax-and-fix method zfh1 h2 r yh2 8 f A F; h2 A H: ð4Þ
is that relax-and-fix is a constructive heuristic not utilizing h1 A H

LP-relaxation to fix integer variables.


wf  zfh1 h2 r cap 8 h1 A H; h2 A H: ð5Þ

zfh1 h2 ¼ 0 8 l1 A Of ; l1 h1 2
= A: ð6Þ
3. Mathematical formulation for the intermodal hub location
zfh1 h2 ¼ 0 8 l2 A Df ; h2 l2 2
= A: ð7Þ
problem

0 r zfh1 h2 r 1; vf Z 0 8 h1 ; h2 A H; f A F: ð8Þ
The IHL problem can be viewed as embedded in an undirected
network N ¼(L, H, A), where the set of nodes of the network L
yh A f0; 1g 8 h A H: ð9Þ
corresponds to origins and destinations and H corresponds to
intermodal hubs. Associated with link (a, b) A A, which connects Constraints (2) assure that all flows must be shipped either via
nodes a and b, is a non-negative weight dða; bÞ ¼ dðb; aÞ represent- inter-hub transportation or via other outsourcing operations.
ing the travel distance that is associated with the cost for move- Constraints (3) and (4) ensure that an intermodal hub must be
ment from node a to node b. Due to economies of scale, there is a open if there are any flows moving through the hub. Constraints
206 Y. He et al. / Omega 57 (2015) 203–211

(5) require that the amount of a flow moved through an inter-hub ϵ the iteration number
route cannot exceed its capacity that is limited by volumes of UBLag the current best upper bound for the original
trucks or trains assigned to the flow on the inter-hub route in problem (1)–(9)
industrial practice. Constraints (6) and (7) enforce that flow y the solution point of hub-open decisions
transportation is disallowed between two unconnected intermo- corresponding to the original problem (1)–(9)
dal hubs. Constraints (8) enforce integrality and non-negativity with the
requirements for various variables. current best upper bound
ψ ϵfh the dual multipliers at iteration ϵ
λ step size of movement at the subgradient
4. The improved MIP heuristic for the intermodal hub location procedure
problem ðzϵfh h2 ; vϵf ; yϵh Þ the optimal solution for the Lagrangian problem
1
at iteration ϵ
To solve the IHL problem, the IMMIP heuristic initially generates an OPTðψ ϵfh Þ the optimal objective value (10) for the
incumbent feasible solution using branch-and-bound and Lagrangian Lagrangian problem at iteration ϵ
relaxation-based heuristic within a limited amount of computational ω a predefined parameter that is initially set to two
time and generates a relaxation solution using LP relaxation. To
facilitate the explanation, we make the below definitions: The Lagrangian relaxation-based heuristic:
Step 0: Set UBLag ¼ 1; ϵ ¼1; ω ¼ 2; ψ ϵfh ¼ 0. Go to Step 1.
y the incumbent feasible solution point of y that leads to an
Step 1: The Lagrangian problem is solved to obtain the optimal
upper bound of the IHL problem
y the relaxation solution point of y that leads to a lower bound objective value (10) and solution ðzϵfh ; vϵf ; yϵh Þ at iteration ϵ.
1 h2

of the IHL problem ω is decreased whenever the Lagrangian solution has failed
to improve in a specified number of steps. At the first
iteration, λ is calculated by (11). Otherwise, it is calculated
We next describe the methods for getting y and y and the
by (12). Go to Step 2.
algorithm of the improved MIP heuristic in details.

w  OPTðψ ϵfh Þ
λ¼P P Ph 1 h 2 A A ϵ P P P 2 ð11Þ
f AF h1 A H ð h2 A H
zfh h2
 yϵh1 Þ2 þ f A F h2 A H h1 h2 A A ϵ
h1 A H
zfh  yϵh2
1 1 h2

w  ðUBLag  OPTðψ ϵfh ÞÞ


λ¼P P P h1 h2 A A ϵ P P P 2 ð12Þ
f AF h1 A H ð h2 A H
zfh h2
yϵh1 Þ2 þ f A F h2 A H h1 h2 A A ϵ
h1 A H
zfh yϵh2
1 1 h2

4.1. The combined methods for getting y and y Step 2: The Lagrangian dual multipliers ψfh are updated by
standard subgradient optimization schemes (13) and (14).
The IMMIP heuristic combines branch-and-bound (using Cplex Go to Step 3.
12.6 as the solver) and a Lagrangian relaxation-based heuristic to
h1X
h2 A A
obtain the initial incumbent feasible solution. The purpose of using ψ ϵfhþ1 1 ¼ maxf0; ψ ϵfh1 þ λ  ð zϵfh1 h2  yϵh1 Þg
these two techniques to generate the initial incumbent solution is h2 A H (13)
to reduce the chance of the solution being infeasible. During our h1X
h2 A A

computational tests, we have seen that a number of computational ψ ϵfhþ2 1 ¼ maxf0; ψ ϵfh2 þ λ  ð zϵfh1 h2  yϵh2 Þg
h1 A H (14)
tests generate infeasible initial solutions if the Lagrangian
relaxation-based heuristic is not combined. We next start to Step 3: Fix yh to the solution value zϵfh for the original problem
1 h2
describe the combined Lagrangian relaxation-based heuristic, in
(1)–(9). Then, solve this restricted problem with a limited amount
which constraints (3) and (4) are dualized into the objective
of computational time. Update UBLag to the optimal objective value
function (1) with non-negative dual multipliers ψfh. The Lagran-
and y to the corresponding of the solution of y if the optimal
gian problem becomes
objective value of the problem is less than UBLag. Set ϵ ¼ ϵ þ 1 and
X X l1 X
h1 A A h1X
h2 A A h2X
l2 A A
go to Step 1 if the computational time consumed is less than Tal.
min ðtcl1 h1 þ α  hch1 h2 þ tch2 l2 Þ
Otherwise, stop the algorithm.
f A F l1 ¼ Of h1 A H h2 A H l2 ¼ Df
X X
wf  zfh1 h2 þ och  yh þ vcf  vf  wf
hAH f AF
0 1
XX h1X
h2 A A XX
þ ψ fh1  @ zfh1 h2  yh1 A þ ψ fh2 We next describe the combined branch-and-bound method, for
f A F h1 A H h2 A H f A F h2 A H which the original IHL problem is reduced to a smaller restricted
0 1 P
h1X
h2 A A problem by adding a constraint h A H yh ¼ m. The constraint
@ zfh1 h2  yh2 A ð10Þ enforces there are an exact number of m hubs being open where
h1 A H m is a parameter smaller than H. By adding the constraint, the
branch-and-bound method can quickly get various initial incum-
Subject to : ð2Þ; ð5Þ–ð9Þ bent feasible solutions for hub-open decision variable y by setting
Before the description of the Lagrangian relaxation-based m to different values.
heuristic, we provide the following notations: In the improved MIP heuristic, the Lagrangian relaxation-based
heuristic and the branch-and-bound method are applied to obtain a
Y. He et al. / Omega 57 (2015) 203–211 207

number of initial feasible solutions in a limited amount of time. The follows:


best solution among these initial feasible solutions is selected as y.
ρ1ρ2 ð1  φh Þ
The IMMIP heuristic combines the LP relaxation to obtain the Ph ¼ P ρ ð1  φj Þ 8 hA H binary : ð15Þ
relaxation solution y of the hub-open decision variables y. Theore- ρ2
j A H binary 1
tically, a lower bound provided by the Lagrangian relaxation method
Here, parameters ρ1 and ρ2 (ρ1 Z 1 and ρ2 Z 1) are used to
is always at least as good as the LP relaxation bound. However, the
adjust sampling weights. A larger value of ρ1 and/or ρ2 leads to
reason we still use the LP relaxation to achieve the lower bound is
a higher likelihood of selecting a hub-open variable that has a
that we cannot obtain a better lower bound by the Lagrangian
smaller gap between y and y. The intuition is that our
relaxation than the LP relaxation in a limited amount of computa-
computational tests show that the smaller the 1-norm distance
tional time, due to that the convergence of the sub-gradient
between the variable's values y h and y h , the higher the
optimization process of the Lagrangian relaxation is hard to achieve.
probability of the hub-open variable's value y h is the same
with its corresponding optimal value of yh of the IHL problem.
4.2. The algorithm of the improved MIP heuristic The analysis of its statistical significance on the application for
the IHL problem will be given in Section 5. We therefore define
Given y and y, the intuition behind the IMMIP heuristic is to the selection probability function (15) for selecting Hsample that
find an efficient way to fix a subset of hub-open variables y that favors the variables that have smaller 1-norm distances
leads to a restricted version of the IHL problem that is easier to between hub-open variable's values y h and y h . Go to Step III.
solve and hence enables a quick solution procedure. To define the
 Step III – subproblem creation: With the individual selection of
restricted problem (called subproblem), three subsets for the set of Hsample and Hbinary for each promising subproblem and fixing
intermodal hubs H are introduced. Subset Hfix defines an index set the hub-open variables with indices in Hsample to the corre-
p
of hub-open variables that have been fixed based on results of the sponding values in y, a number of N d promising subproblems
previous iteration(s); Hsample indicates an index set of hub-open can be easily created. Besides promising subproblems, a num-
r
variables that are individually selected and fixed to the corre- ber of N d random subproblems are also created using a random
sponding values of y for a specific subproblem; and Hbinary defines sampling scheme where the size of fixed variables is defined as
the remaining index set of free hub-open variables. The subpro- j H fix j þ nd. Different from the strategy of fixing variables to y
blem is defined as follows with the above definitions: for promising subproblems, the binary values, which the
selected variables are fixed to, are randomly sampled with a
Minimize ð1Þ constant probability of being 0 or 1. Go to Step IV.
Subject to : ð2Þ–ð8Þ
 p
Step IV – subproblem evaluation: A total number of N d þ N d
r

subproblems are created in Step III. These subproblems are


yh ¼ y h ; hA H fix [ H sample solved by the branch-and-bound method. As all created sub-
problems are solved by the branch-and-bound method, possi-
yh A f0; 1g; hA H binary ðIHLPSÞ
bly obtaining a better objective value than the current best
PS
Subproblems (IHL ) are called promising subproblems. In order to upper bound BestUB. To provide better guidance for subsequent
have a global perspective, the method also generates a number of iterations, BestUB and y are therefore updated whenever a
random subproblems by fixing some of the variables randomly. We better objective value is found. Go to Step V.
next start to describe the IMMIP heuristic that mainly contains
 Step V – update: The subproblem with the smallest upper
initialization, neighborhood search, subproblem creation, subpro- bound is selected as the best subproblem at the current
blem evaluation, update and stop-criterion check. Before the iteration. If the best subproblem is one of the promising
description of the method, we provide the following notations: subproblems, we add a subset of Hsample of the best subproblem
into Hfix, set Hbinary ¼ Hbinary - Hfix, Hsample ¼ ∅, and d ¼ d þ 1.
BestUB the current best upper bound of the IHL problem Otherwise, we update Hfix ¼∅, Hbinary ¼H, Hsample ¼∅, and d ¼1.
d depth of iteration Go to Step VI.
N pd Number of promising subproblems created at the d-th  Step VI – stop-criterion check: The algorithm stops when the
iteration computational time limit (Tlim) is met. Otherwise, the algorithm
N rd Number of random subproblems created at the d-th goes to Step II.
iteration
nd number of hub-open variables selected to be fixed on 5. Statistical analysis of the IMMIP heuristic for the
the promising subproblem at the d-th iteration application
Ts computational time limit for solving subproblem
(IHLPS) In Section 4.2, we made two claims without verification:
T lim computational time limit for the whole method
 Step I – initialization: Set d ¼1, BestUB ¼ 1, H fix ¼ ∅, Hsample ¼ ∅  Claim 1: Potential subproblems yield better upper bounds than
and H binary
¼ H. Solve the IHL problem using the branch-and- random subproblems.
 Claim 2: The smaller the distance between y h and y ( 8 h A H),
bound method and the Lagrangian relaxation-based heuristic h
to obtain y and solve the LP relaxation of the IHL problem to the higher the probability that y h is the same with the
obtain y. Go to Step II. corresponding optimal value of the IHL problem.
 Step II – neighborhood search: Neighborhood search is to use a
domain-knowledge-guided selection procedure to determine In this section, we will perform statistical analysis to show that
subsets Hsample and Hbinary for each promising subproblem. All the two claims are statistically significant for the application of the
selected hub-open variables with indices in Hsample are fixed to IMMIP heuristic to the IHL problem. We generated two groups of
the corresponding values in y. test instances (SetA and SetB) using different parameter settings in
We define parameter φh, indicating the distance between y h order to perform statistical analysis. SetA has 100 origins/destina-
and y h (that is, φh ¼ j y h  y h j ). The probability (Ph) of selecting tions, 80 hubs and 100 flows and SetB has 120 origins/destina-
variable yh to be fixed in a promising subproblem is defined as tions, 90 hubs and 130 flows. We assume origins/destinations and
208 Y. He et al. / Omega 57 (2015) 203–211

Fig. 1. Boxplots and finite mixtures for promising and random subproblems.

Of the origin of flow f is randomly generated


Df the destination of flow f is randomly generated
wf there are five such settings denoted by low, low-
medium, medium, medium-high and high that indicate
the amount of flow is uniformly distributed in
f0:34; 0:5; 1; 2; 2:96g  ½1000; 3000, respectively

tchl there are five such settings denoted by low, low-


medium, medium, medium-high and high that indicate
the transportation cost per-amount of flow between
intermodal hub h and origin/destination l is uniformly
distributed in f0:34; 0:5; 1; 2; 2:96g  ½0:002; 0:003  dist hl ,
respectively, where disthl is the Euclidean distance
between intermodal hub h and origin/destination l.
α there are five such settings denoted by low, low-
medium, medium, medium-high and high that indicate
the discount factor for the inter-hub transportation is
uniformly distributed in f0:47; 0:7; 1; 1:2; 1:78g  ½0:3; 0:5,
respectively
Fig. 2. Boxplots and normal quantiles.

hch1 h2 there are five such settings denoted by low, low-


medium, medium, medium-high and high that indicate
hubs are located in a square area that can be described by the transportation cost per-amount of flow between
Euclidean Coordinate System with four corners' coordinates given intermodal hubs h1 and h2 is uniformly distributed in
as (0, 0), (0, 800), (800, 0) and (800, 800); and we also assume the f0:34; 0:5; 1; 2; 2:96g  α  ½0:002; 0:003  dist h1 h2 ,
square area is divided into a center area and a surrounding area respectively. Note that coefficient {0.34, 0.5, 1, 2, 2.96}
where the center area is an area with four corners' coordinates used for setting tchl and hch1 h2 is the same for a test
given as (200, 200), (200, 600), (600, 200) and (600, 600) and the instance
surrounding area is the remaining area. With the coordinate och there are five such settings denoted by low, low-
system, all hub locations were randomly generated in the center medium, medium, medium-high and high that indicate
area but all origins/destinations were randomly generated in the the open cost of intermodal hubs is uniformly
surrounding area. For generating test instances, we applied the distributed in f0:34; 0:5; 1; 2; 2:96g  ½20  temp; 30  temp
Central Composite Design (CCD) method to capture a wide range where the parameter temp is defined as
of problem structures. In design-of-experiment (DOE) terminol- P P P P P 
w hch1 h2 tc
f AF f h1 A H h2 A H
ogy, a CCD consists of a two-level factorial design, i.e., cube points j Fj  jHj2 þ hAH l A L hl
j H j j Lj .
xi ¼ 71, superposed by specifically located star (xi ¼ 7 β  δij vcf The outsource cost per-unit amount and per-unit
where βrepresents the positions of the star points, while δij is distance is set to 30. The level is high such that the
the Kronecker delta which equals 1 if i¼j and 0 elsewise) and outsource operation should not be used if not necessary
center points xi ¼0. Relatively, a CCD design is more compact than cap The limit of the amount of a flow moved through an
a peering regular three-level factorial DOE. In the CCD design, five inter-hub route is set to 2000
levels were applied to four parameters (flow amount, shipping
cost, discount factor and hub-open cost) to increase varieties of
test instances so as to ensure validness of the statistical analysis. The IMMIP heuristic was programmed using GAMS 24.3 with a
There are 52 test instances in each group. The detailed design for computing capacity (Intel Pentium 4, 1.7 GHz processor and 4.0 GB
parameter settings is given as follows: RAM). The parameter settings for the IMMIP heuristic was given as
Y. He et al. / Omega 57 (2015) 203–211 209

follows: The total computational time of 1800 s was allocated for somewhat right-skewed. Although this is the case, the clustering
all test instances and computational time allocated for each results still show that the lower bounds obtained by the promising
subproblem is 30 s. Parameters ρ1 and ρ2 were both set to 2. and random subproblems can be well classified.
The numbers of promising and random subproblems at each To test Claim 2, we generated three groups of test instances
iteration were set to 5 and 2 and the size of subset Hsample was with the same parameter settings mentioned above but with
set to maxð j H2 j ; 0:60  j H binary j Þ. For the Lagrangian relaxation- smaller sizes to ensure the optimal solution be found for all test
based heuristic that is used to generate a number of initial instances. The test results are shown in Fig. 2. The top sub-figure
solutions, its total time is set to 80 s; parameter ω is reduced by shows results for the group with 30 origins/destinations, 25 hubs
half after 5 iterations without improvements of objectives; and, in and 30 flows; the middle sub-figure is for the one with 25 origins/
Step 3, the limited amount of computational time is set to 5 s to destinations, 30 hubs and 35 flows; and the bottom sub-figure is
solve the original problem with fixed hub-open decision variables. for the one with 35 origins/destinations, 25 hubs and 35 flows. MR
For the branch-and-bound method, m is set to {3, 4, 5,…, 10}, denotes the probability of the value y h be the same with the
respectively. Time allocated for each reduced problem is 30 s. The corresponding optimal solution value for each individual test
best solution obtained by the Lagrangian relaxation-based heur- instance. For each hub-open decision variable yh ( 8 h A H), distance
istic and the branch-and-bound method is selected as y. is calculated as j y h  y h j . The distance value is separated into two
To test Claim 1, we took out all generated promising and different bins in order to perform statistical comparisons shown
random subproblems for each test instance and performed statis- by boxplots and normal quantiles. From the experimental results,
tical analysis using different statistical techniques, such as boxplot, we can see that both statistical methods have shown a strong
Welch's test and finite mixture model. The computational results statistical relationship between MR and distance. That is, the
showed that all promising and random subproblems are feasible smaller the distance between y h and y h ( 8 h A H), the higher the
but Welch's test showed that upper bounds derived from promis- probability that y q is the same with the corresponding optimal
ing subproblems are statistically significantly better than those value of the IHL problem.
derived from random subproblems with a p-value less than
0.0001. Fig. 1 shows the results of boxplot and the fitting curves
of finite mixture model for three representative examples. The top 6. Computational results
three sub-figures illustrate boxplots showing that promising sub-
problems have significantly better upper bounds when compared We compared the IMMIP heuristic with the commercial MIP
with random subproblems. The bottom three sub-figures show the solver (Cplex 12.6), the Lagrangian relaxation-based heuristic
histograms of upper bounds of both promising and random described in Section 4.1, feasibility pump, RINS, RENS, and relax-
subproblems. The grey curves were derived by the finite mixture and-fix [50,52,48]. All seven approaches were programmed using
model method for which the distribution of the upper bounds is GAMS 24.3. The same total computational time (1800 s) was
assumed to be a mixture of two normal distributions. The results assigned to the seven approaches to ensure comparison fairness.
show that one component of the mixed distribution well follows a Computational tests were performed on two groups of test
normal distribution but the other component's distribution is instances (SetA and SetB) described in Section 5.

Table 1
Comparisons of computational results for SetA.

Factor LB (%) IMMIP-G (%) Cplex-G (%) LagH-G (%) FP-G (%) RINS-G (%) RENS-G (%) RF-G (%)

Flow amount
Low 161,392 16.35 42.76 20.01 42.76 17.13 63.27 n
Low-medium 301,465 25.46 29.97 36.97 30.38 26.30 57.65 n
Medium 550,017 33.63 50.00 47.31 51.37 58.51 60.62 n
Medium-high 1,200,556 28.93 37.12 39.22 35.88 43.05 42.94 n
High 1,400,738 39.90 38.46 38.46 38.46 38.46 38.46 n

Shipping cost
Low 159,246 29.76 47.09 48.28 47.09 65.32 64.35 n
Low-medium 303,383 33.08 40.77 49.62 39.58 46.40 65.73 n
Medium 555,468 33.15 49.43 44.69 50.79 52.46 60.52 n
Medium-high 1,198,638 21.30 26.32 26.58 26.67 22.95 34.85 n
High 1,359,276 30.28 38.71 31.19 38.71 38.71 38.24 n

Discount factor
Low 426,981 32.84 46.82 43.41 46.82 46.82 55.78 n
Low-medium 732,480 28.07 31.99 37.04 31.22 33.05 50.49 n
Medium 628,392 32.63 44.66 44.19 46.03 49.97 56.91 n
Medium-high 769,541 26.31 35.10 39.15 35.04 36.31 50.10 n
High 508,151 31.35 77.11 40.01 77.11 77.11 75.68 n

Hub-open cost
Low 362,164 27.04 37.34 42.44 37.34 47.11 46.88 n
Low-medium 637,419 24.86 30.67 42.35 29.48 35.86 46.73 n
Medium 622,186 30.90 47.96 42.37 49.33 52.03 59.67 n
Medium-high 864,602 29.52 36.42 33.84 36.77 33.50 53.85 n
High 622,613 51.01 60.21 55.59 60.21 60.33 62.49 n

Average 691,478 29.24 39.15 40.25 39.32 41.48 53.52 n

LB indicates lower bounds yielded by the LP relaxation of the IHL formulation. IMMIP-G, Cplex-G, LagH-G, FP-G, RINS-G, RENS-G, and RF-G indicate optimality gaps achieved
by the IMMIP, Cplex, Lagrangian relaxation-based, feasibility pump, RINS, RENS, and relax-and-fix methods, respectively. The optimality gaps are calculated as the difference
between lower and upper bounds divided by the upper bound. n indicates the solutions obtained by the corresponding methods are infeasible.
210 Y. He et al. / Omega 57 (2015) 203–211

Table 2
Comparisons of computational results for SetB.

Factor LB (%) IMMIP-G (%) Cplex-G (%) LagH-G (%) FP-G (%) RINS-G (%) RENS-G (%) RF-G (%)

Flow amount
Low 193,252 19.24 66.03 41.43 66.03 66.29 n n
Low-medium 365,924 27.64 31.87 44.66 31.68 32.83 n n
Medium 690,035 31.36 56.74 54.79 56.80 59.66 n n
Medium-high 1,458,978 36.13 39.12 37.39 40.46 43.97 n n
High 1,670,956 38.59 43.64 43.64 43.64 43.64 n n

Shipping cost
Low 199,750 28.80 68.42 57.24 68.42 68.42 n n
Low-medium 367,179 33.23 40.13 49.09 41.47 44.48 n n
Medium 675,643 31.06 56.79 53.16 56.85 59.75 n n
Medium-high 1,457,722 30.53 30.86 32.96 30.67 32.32 n n
High 1,779,596 31.38 40.84 40.84 40.84 40.84 n n

Discount factor
Low 543,430 30.63 49.64 49.15 49.64 49.64 n n
Low-medium 876,831 32.34 36.28 41.47 37.76 38.05 n n
Medium 777,861 30.71 54.37 52.41 54.42 57.32 n n
Medium-high 948,070 31.43 34.71 40.58 34.38 38.74 n n
High 618,173 32.40 79.01 54.93 79.01 79.01 n n

Hub-open cost
Low 477,160 24.68 24.83 55.95 24.28 48.22 n n
Low-medium 776,457 27.76 29.13 44.46 30.37 34.94 n n
Medium 770,328 29.89 58.45 50.13 58.58 58.48 n n
Medium-high 1,048,444 36.01 41.86 37.59 41.77 41.86 n n
High 744,709 44.85 71.14 66.32 71.14 71.14 n n

Average 845,527 31.49 43.52 45.37 43.89 46.22 n n

n indicates the solutions obtained by the corresponding methods are infeasible.

Detailed parameter settings for the IMMIP heuristic can be achieved for all hub-open variables. In the implementation, para-
referred to Section 5. For the commercial MIP solver, Cplex 12.6 meters α and βare set to 10 and 5, respectively; and the computer
with default setting is called by GAMS. For the Lagrangian time for solving subproblems is set to 120 s. The total computa-
relaxation-based heuristic, parameter ω is reduced by half after tional time for relax-and-fix is close to 1800 s as well for a fair
5 iterations without improvements of objectives. The computa- comparison with the IMMIP heuristic.
tional time for solving the original problem with fixed hub-open Computational results are given in Tables 1 and 2. From the
decision variables is set to 5 s. For the feasibility pump heuristic, results, we have that the IMMIP heuristic obviously offers solu-
Cplex 12.6 is called by GAMS, whereby option of feasibility pump tions superior to other six compared methods across various
“fpheur” is invoked. The value of fpheur is set to 1, indicating that parameter settings. Results show that optimality gaps achieved
the feasibility pump heuristic is invoked with an emphasis on by the other methods are significantly affected by problems'
finding a feasible solution. For the RINS heuristic, Cplex 12.6 is parameter settings and have a large variation, while optimality
called by GAMS, whereby option of RINS “rinsheur” is invoked. The gaps achieved by the IMMIP heuristic are comparatively more
value of rinsheur is set to 1, indicating that RINS is invoked every stable and have a small variation. Problems' parameter settings
another node in the tree. For the RENS heuristic, the SCIP solver is also might have considerable effect on problems' complexities, e.
called by GAMS. SCIP means solving constraint integer programs g., a strong pattern of the optimality gaps obtained by the seven
which can be used as a pure MIP solver. When calling the solver, compared methods shows that the larger hub-open costs are set
the option “heuristics/rens/freq” is invoked with the option value for a problem, the more difficultly the problem is solved
set to 1, indicating the frequency of invoking RENS.
For the relax-and-fix method, the set H of hubs is separated
into three subsets. One subset is a window that contains a number 7. Conclusions and future research
of hubs for which all hub-open variables are remained as binary
variables. The subset contains ½1; α at the first iteration. Another This paper proposed an improved MIP heuristic for the IHL
subset contains a number of hubs for which all hub-open variables problem. The merit of the heuristic is that it systematically
have been fixed based on values of the previous iteration(s). The combines exact methods with heuristics to derive problems'
subset is empty at the first iteration. The last subset consists of domain knowledge whose statistical significance for deriving the
hubs for which all hub-open variables are relaxed as continuous optimal solution value has been verified through statistical meth-
variables. The subset contains ½α þ 1; j Hj  at the first iteration. Such odologies, such as boxplots, normal quantiles, Welch's test, linear
smaller restricted problems are solved by MIP solver and the regression and finite mixture models. Computational results
resulting solution value is used to fix a number of β hub-open showed that the method is capable of deriving solutions superior
variables in the window. At the second iteration, the window roll to those obtained by other state-of-the-art methods.
forwards to the next α hubs, ½1 þ β ; α þ β . The hub-open variables One future work along this line of research is to investigate the
that have been fixed contain indices [1, β]. The remained hub-open possibilities of applying the heuristic to other MIP problems. An
variables with indices ½α þ β þ 1; j Hj  are relaxed to continuous interesting application of the heuristic is radiation treatment
variables. After the second iteration, another number of βhub- planning. The problem aims to select a set of angles of laser
open variables are fixed. The same manner is iteratively performed beams and to determine dose quantities of these beams targeted
until the window rolls to the end and a binary value has been on tumors with an objective of optimizing treatments. This
Y. He et al. / Omega 57 (2015) 203–211 211

problem is notoriously difficult to solve due to problem size and [22] Ernst AT, Krishnamoorthy M. Solution algorithms for the capacitated single
highly combinatorial structure present. The proposed heuristic has allocation hub location problem. Annals of Operations Research 1999;86
(0):141–59.
the potential to effectively solve this problem by fixing some of [23] Farahani RZ, Hekmatfar M, Arabani AB, Nikbakhsh E. Hub location problems: a
these decision variables in advance that reduces the large-size review of models, classification, solution techniques, and applications. Com-
original problem to more manageable, smaller subproblems. puters Industrial Engineering 2013;64(4):1096–109.
[24] Fischetti M, Glover F, Lodi A. The feasibility pump. Mathematical Programming
Other interesting future work includes exploration of theore- 2005;104(1):91–104.
tical foundations of the intelligent neighborhood search and [25] Fischetti M, Lodi A. Local branching. Mathematical Programming, Series B
investigation of relative effectiveness of different bounding solu- 2003;98:23–47.
[26] Fischetti M, Salvagnin D. Feasibility pump 2.0. Mathematical Programming
tions for deriving domain knowledge. An interesting bounding
Computation 2009;1:201–22.
technique is approximation algorithm. This class of algorithms are [27] Glover F, Lokketangen A, Woodruff DL. Scatter search to generate diverse MIP
known for deriving objective values that are guaranteed to be solutions. OR computing tools for modeling, optimization and simulation:
within a range between the optimal objective and σ (σ 41) times interfaces in computer science and operations research 2000(1).
[28] Guzelsoy M, Nemhauser G, Savelsbergh M. Restrict-and-relax search for 0-1
optimal objectives. We are interested in investigating whether the mixed-integer programs. EURO Journal on Computational Optimization
optimal objective of subproblems can follow similar metric but 2013;1(1–2):201–18.
under a smaller σ value if the approximation algorithm is applied [29] Hakimi SL. Optimum distribution of switching centres in a communication
network and some related graph theoretic problems. Operations Research
to generate the subproblems. 1965;13(3):462–75.
[30] Ishfaq R, Sox CR. Hub location-allocation in intermodal logistic networks.
European Journal of Operational Research 2011;210(2):213–30.
Acknowledgements [31] Karaoglan I, Altiparmakb F, Karac I, Dengizc B. The location-routing problem
with simultaneous pickup and delivery: formulations and a heuristic
approach. Omega 2012;40(4):465–77.
The authors would like to thank the support from the National [32] Klincewicz JG. A dual algorithm for the uncapacitated hub location problem.
Natural Science Foundation of China (Grant No. 51105394), and Location Science 1996;4(3):173–84.
[33] Klincewicz JG. Hub location in backbone/tributary network design: a review.
Fundamental Research Funds for the Central Universities (Grant
Location Science 1998;226(1–4):307–35.
No. 30920130111001). [34] Klincewicz JG. Enumeration and search procedures for a hub location problem
with economies of scale. Annals of Operations Research 2002;110(1–
4):107–22.
References [35] Labbe M, Yaman H, Gourdin E. A branch and cut algorithm for hub location
problems with single assignment. Mathematical Programming 2005;102
[1] Abdinnour-Helm S. A hybrid heuristic for the uncapacitated hub location (2):371–405.
problem. European Journal of Operational Research 1998;106(2-3):489–99. [36] Mayer G, Wagner B. Hublocator: an exact solution method for the multiple
[2] Achterberg T, Berthold T. Improving the feasibility pump. Discrete Optimiza- allocation hub location problem. Computer and Operations Research 2002;29
tion 2007;4(1):77–86. (6):715–39.
[3] Alumur SA, Kara BY, Karasan OE. Multimodal hub location and hub network [37] Meng Q, Wang X. Intermodal hub-and-spoke network design: incorporating
design. Omega 2012;40(6):927–39. multiple stakeholders and multi-type containers. Transportation Research
[4] Arnold P, Peeters D, Thomas I. Modelling a rail/road intermodal transportation Part B: Methodological 2011;45(4):724–42.
system. Transportation Research Part E: Logistics and Transportation Review [38] O'Kelly ME, Bryan DL. Hub location with flow economies of scale. Transporta-
2004;40(3):255–70. tion Research Part B: Methodological 1998;32(8):605–16.
[5] Aykin T. Network policies for hub-and-spoke systems with applications to the [39] Pi L, Pan Y, Shi L. Hybrid nested partitions and mathematical programming
air transportation system. European Journal of Operational Research 1995;29 approach and its applications. IEEE Transactions on Automation Science and
(3):201–21. Engineering 2008;5(4):573–86.
[6] Balas E, Martin CH. Pivot and complement—a heuristic for 0–1 programming. [40] Pirkul H, Schilling DA. An efficient procedure for designing single allocation
Management Science 1980;26(1):86–96. hub and spoke systems. Management Science 1998;44(12):235–42.
[7] Belvaux G, Wolsey LA. Bc-prod: a specialized branch-and-cut system for lot- [41] Pochet Y, Vyve MV. A general heuristic for production planning problems.
sizing problems. Management Science 2000;46(5):724–38. INFORMS Journal on Computing 2004;16(3):316–27.
[8] Bertacco L, Fischetti M, Lodi A. A feasibility pump heuristic for general mixed- [42] Sa EM, Camargo R, Miranda G. An improved benders decomposition algorithm
integer problems. Discrete Optimization 2007;4(1):77–86. for the tree of hubs location problem. European Journal of Operational
[9] Berthold T. RENS the optimal rounding. Mathematical programming computa- Research 2013;226(2):185–202.
tion 2013; Online. [43] Sorensen K, Vanovermeire C, Busschaert S. Efficient metaheuristics to solve
[10] Bonami P, Cornuejols G, Lodi A, Margot F. A feasibility pump for mixed integer the intermodal terminal location problem. Computers and Operations
nonlinear programs. Mathematical Programming 2009;119(2):331–52. Research 2012;39(9):2079–90.
[11] Camargo R, Miranda G, Lokketangen A. A new formulation and an exact [44] Stadtler H. Multilevel lot sizing with setup times and multiple constrained
approach for the many-to-many hub location-routing problem. Applied resources: internally rolling schedules with lot-sizing windows. Operations
Mathematical Modelling 2013;37(12–13):7465–80. Research 2003;51(3):487–502.
[12] Camargo R, Miranda G, Luna HP. Benders decomposition for the uncapacitated [45] Toh RS, Higgins R. The impact of hub-and-spoke network centralization and
multiple allocation hub location problem. Computers and Operations Research route monopoly on domestic airline profitability. Transportation Journal
2008;35(4):1047–64. 1985;24(4):16–27.
[13] Campbell JF. Integer programming formulations of discrete hub location [46] Topcuoglu H, Corut F, Ermis M, Yilmaz G. Solving the uncapacitated hub
problems. European Journal of Operational Research 1994;72(2):387–405. location problem using genetic algorithms. Computers and Operations
[14] Campbell JF. Hub location and the p-hub median problem. Operations Research 2005;32(4):967–84.
Research 1996;44(6):923–35. [47] Wu T, Akartunali K, Song J, Shi L. Mixed integer programming in production
[15] Chen J. A hybrid heuristic for the uncapacitated single allocation hub location planning with backlogging and setup carryover: modeling and algorithms.
problem. Omega 2007;35(2):211–20. Discrete Event Dynamic Systems 2013;23(2):211–39.
[16] Contreras I, Cordeau J, Laporte G. Benders decomposition for large-scale [48] Wu T, Shi L. Mathematical models for capacitated multi-level production
uncapacitated hub location. Mathematical Programming 2011;59(6):1477–90. planning problems with linked lot sizes. International Journal of Production
[17] Contreras I, Diaz JA, Fernandez E. Branch and price for large-scale capacitated Research 2011;49(20):6227–47.
hub location problems with single assignment. INFORMS Journal on Comput- [49] Wu T, Shi L, Geunes J, Akartunalı K. On the equivalence of strong formulations
ing 2010;23(1):41–55. for capacitated multi-level lot sizing problems with setup times. Journal of
[18] Danna E, Rothberg E, Pape CL. Exploring relaxation induced neighborhoods to Global Optimization 2011;53(4):615–39.
improve mip solutions. Mathematical Programming 2005;102(1):71–90. [50] Wu T, Shi L, Geunes J, Akartunalı K. An optimization framework for solving
[19] Ernst AT, Krishnamoorthy M. Efficient algorithms for the uncapacitated single capacitated multi-level lot-sizing problems with backlogging. European Jour-
allocation p-hub median problem. Location Science 1996;4(3):139–54. nal of Operational Research 2011;214(2):428–41.
[20] Ernst AT, Krishnamoorthy M. Exact and heuristic algorithms for the uncapa- [51] Wu T, Shi L, Song J. An mip-based interval heuristic for the capacitated multi-
citated multiple allocation p-hub median problem. European Journal of level lot-sizing problem with setup times. Annals of Operations Research
Operational Research 1998;104(1):100–12. 2012;196(1):635–50.
[21] Ernst AT, Krishnamoorthy M. An exact solution approach based on shortest- [52] Wu T, Zhang D, He Y. A novel mixed integer programming formulation and
paths for p-hub median problems. INFORMS Journal on Computing 1998;10 progressively stochastic search for capacitated lot sizing. Journal of Systems
(2):149–62. Science and Systems Engineering 2011;20(2):173–92.

Anda mungkin juga menyukai