Sei sulla pagina 1di 13

European Journal of Operational Research 240 (2015) 147159

Contents lists available at ScienceDirect

European Journal of Operational Research


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

Decision Support

Operations research models for coalition structure in collaborative


logistics
Mario Guajardo a,, Mikael Rnnqvist a,b
a
b

Department of Business and Management Science, NHH Norwegian School of Economics, N-5045 Bergen, Norway
Dpartement de gnie mcanique, Universit Laval, Qubec, G1V 0A6, Canada

a r t i c l e

i n f o

Article history:
Received 10 August 2013
Accepted 16 June 2014
Available online 2 July 2014
Keywords:
Game theory
Group decisions and negotiations
Collaborative logistics
Forest transportation
Inventory sharing

a b s t r a c t
Given a set of players and the cost of each possible coalition, the question we address is which coalitions
should be formed. We formulate mixed integer linear programming models for this problem, considering
core stability and strong equilibrium. The objective function looks for minimizing the total cost allocated
among the players. Concerned about the difculties of managing large coalitions in practice, we also
study the effect of a maximum cardinality constraint per coalition. We test the models in two applications. One is in collaborative forest transportation and the other one in inventory of spare parts for oil
operations. In these situations, collaboration opportunities involving signicant savings exist, but for several reasons, it may be better to group the players in different sub-coalitions rather than in the grand coalition. The models we propose are thus relevant for deciding how to partition the set of players. We also
prove that if the strong equilibrium model is feasible, its optimal cost is equal to the optimal cost of the
core stability model and, consequently, a coalition structure that solves one problem also solves the other
problem. We present results that illustrate this property. We also present results where the core stability
problem is feasible and the strong equilibrium problem is infeasible. Setting an upper bound on the maximum cardinality of the coalitions, allows us to study the marginal savings of enlarging the cardinality of
the coalitions. We nd that the marginal savings of allowing one more player signicantly decreases as
the bound increases.
2014 Elsevier B.V. All rights reserved.

1. Introduction
Collaboration among different agents is an effective way to
improve logistic operations. Evidence has been provided in many
industries. For instance, Nagarajan and Sosic (2009) and
Nagarajan and Bassok (2008) refer to a variety of contexts where
different suppliers create coalitions to achieve the benets of collaboration, including complementary component-suppliers for
car assembly systems, the health-care industry and various service
sectors. Other examples arise in inventory management (Chen,
2009; zen, Fransoo, Norde, & Slikker, 2008) and transport (Frisk,
Gthe-Lundgren, Jrnsten, & Rnnqvist, 2010; Lozano, Moreno,
Adenso-Daz, & Algaba, 2013). In order to implement collaboration,
there needs to be an agreement among the agents with respect to
how to share the benets. A growing body of literature in Operations Research and Management Science has applied principles of
cooperative game theory to resolve this sharing issue. One of the
main streams in this literature aims at dening allocation rules
Corresponding author. Tel.: +47 55959834; fax: +47 55959650.
E-mail addresses: mario.guajardo@nhh.no (M. Guajardo), mikael.ronnqvist
@gmc.ulaval.ca (M. Rnnqvist).
http://dx.doi.org/10.1016/j.ejor.2014.06.015
0377-2217/ 2014 Elsevier B.V. All rights reserved.

to predict or prescribe how rational players will distribute the


gains they obtain from cooperation, assuming in general that the
grand coalition will be formed, in line with the classical approach
to transferable utility games (Meca-Martnez, Snchez-Soriano,
Garca-Jurando, & Tijs, 1998). The formation of the grand coalition
may nd support in the assumption of superadditive games, i.e.,
where the value of a coalition is at least as good as the sum of
the values of its members acting separately. This would mean that
the more agents collaborate, the better is the outcome they achieve
(or at least not worse than if only some of them or none of them
collaborate). In practice, however, several reasons might exist for
the grand coalition not to be formed. Some of these reasons, for
example, are the managerial complexity of conforming large coalitions and the political issues that can affect the decision of changing from a non-collaborative to a collaborative solution. For
example, in transport operations, collaboration usually involves
just a few partners, because as the number of partners grows coordinating the cooperation becomes more problematic and/or costly
(Lozano et al., 2013). Then, given a set of players and the cost of
each possible coalition, the relevant question arises of which
coalitions should be formed. The formed coalitions dene a partition of the set of players, which in game theory is called a coalition

148

M. Guajardo, M. Rnnqvist / European Journal of Operational Research 240 (2015) 147159

structure (Aumann & Dreze, 1974). In contrast to a vast body of


game theory literature dealing with coalition structure and coalition formation issues (see e.g. Greenberg, 1994), the OR/MS literature is more sparse in this stream. Some of the exceptions
addressing issues related to coalition structures in OR are
Axelrod, Mitchell, Thomas, Bennett, and Bruderer (1995), Leng
and Parlar (2009), Nagarajan and Sosic (2009) and Sosic (2006,
2010). Other recent works have acknowledged the relevance of
the coalition structure issues, although without addressing them
(Chen, 2009; Kim & Jeon, 2009; Lozano et al., 2013).
In this article, we propose mixed integer linear programming
formulations to simultaneously approach the coalition structure
and cost allocation problems. Our formulations aim at deciding
which coalitions should be formed and the cost allocation to the
players in each of them, subject to stability conditions. These stability conditions are of two types. First, we formulate constraints
to model internal stability, i.e., the core constraints for the members within a coalition. Second, we formulate constraints to model
the strong stability, i.e., where no group of players, whether from
the same coalition or from different ones, can get together and
form new coalition(s) in such a way that they are all better off
(Aumann, 1959; Hart & Kurz, 1983). Concerned about the complications of implementing large coalitions in practical situations, we
also incorporate conditions on the maximum number of players
that can form a coalition. We test our models in two applications.
The rst is a collaborative forest transportation problem presented
by Frisk et al. (2010), which includes eight companies operating in
Sweden. As the companies operate in similar regions, the opportunities for collaborative transportation represent important savings,
in the range 515%. The second is a problem on collaboration in
inventory of spare parts for oil operations, where risk pooling represents an important source of savings of about 20% of the inventory costs (Guajardo, Rnnqvist, Halvorsen, & Kallevik, 2014).
Our work contributes in both practical and theoretical aspects.
In the practical aspects, we approach two cases motivated in forestry and oil operations, where collaboration in transport and
inventory operations improve the non-cooperative case. Since we
do not assume any particular form of the cost functions, the potential for applications may also be extended to other industries using
collaborative logistics. In the theoretic aspects, we introduce model
formulations for a relevant problem in the interface of Game Theory and Operations Research. The models consist of a set partitioning problem subject to stability constraints. Because of the
combinatorial structure of this problem and its relevance in theory
and practice, we believe our work opens a rich source of challenges
where OR solution methods can play an important role.
The remainder of this article is organized as follows. In Section 2,
we summarize concepts of stability in cooperative games. In
Section 3, we formulate models for coalition structure and cost
allocation. In Section 4, we apply the models in collaborative problems in the forestry and oil industries, and also discuss its applicability in larger scale problems. Our conclusions are presented in
Section 5.

2. Stability in cooperative games


A growing body of literature has incorporated concepts from
cooperative game theory in collaborative logistics. FiestrasJaneiro, Garca-Jurado, Meca, and Mosquera (2011) review a number of references on collaborative inventory management. In other
areas of operations management, however, the use of cooperative
game theory is much less prevalent (Hu, Caldentey, & Vulcano,
2013). In the following, we introduce some notation and summarize the concepts on stability in cooperative games that are relevant for our article.

We consider a game consisting of a set of players N. We refer to


the cardinality of this set by n jNj. Players can form coalitions, in
order to take advantage of collaboration opportunities. The set N is
the grand coalition. We refer by C Z to the cost incurred by coalition
Z, where Z # N. We assume C Z P 0 for all Z. The cost function C is
called the characteristic function of the game. We refer by uj to the
cost allocated to player j, for each player j 2 N. We assume uj P 0
for all j. A cost allocation vector u u1 ; u2 ; . . . ; un is said to be in
the core of the game (Gillies, 1959) if it satises the following
constraints:

X
uj 6 C Z

8Z  N

j2Z

X
uj C N

j2N

Constraints (1) correspond to the rationality conditions, which state


that there is no subset Z of players such that should they form a coalition separately from the rest they would perceive less total cost
than the total cost allocated to them in u. In the case of Z containing
only one player, the constraint corresponds to the individually
rational condition, which states that the cost allocated to each
player j must not be greater than its stand-alone cost. Constraint
(2) corresponds to the efciency condition, which states that the
sum of the costs allocated to all the players must be equal to the
optimal expected cost of the grand coalition. The core of the game
is the set of all vectors u satisfying constraints (1) and (2). An allocation that belongs to the core is said to be stable. In order to differentiate this type of stability from a following concept, we will refer
to it as core stability.
As mentioned in the introduction, several reasons might exist
for the grand coalition not to be formed. Therefore, it becomes relevant to study the stability of cooperative situations where different group of players may form different coalitions. The formed
coalitions dene what in game theory is called a coalition structure
(Aumann & Dreze, 1974). Following the notation of Sosic (2006), a
coalition structure is a partition on N, that is, fZ 1 ; . . . ; Z r g;
Sr
i1 Z i N; Z i \ Z j ;; i j.
We assume the cost of all coalitions are given parameters. This
in practice does not necessarily mean that the costs and the information needed to compute them are known by all the players, but
it is enough if there would be a decision maker who could gather
this information. Then, this decision maker needs to nd the coalitions that should be formed, aiming at minimizing the total allocated cost while satisfying the constraints. In the two cases that
primarily motivated our work, this decision maker emerges in
two different ways. In the forestry case, there is a consulting team
that gathers the information from the companies and conduces the
analysis, in order to come up with a suggestion on how to implement the collaboration among the companies. It is crucial that this
third party manages the information under condentiality, since
part of the information is sensitive to the companies. In fact, this
has been the only viable way for them to provide the information.
The third party also needs to be impartial and, therefore, the minimization of total cost is an appropriate overall objective when
making the suggestion on how to implement the collaboration.
This objective is also convenient from an environmental point of
view, as usually the savings in transportation cost are associated
to a reduction of emissions from the trucks. In the oil case, the suggestion is made by a main company which has the operational
responsibility for all the warehouses that hold inventory of spare
parts, despite some different ownership structures involving other
companies and stakeholders. The data from all these warehouses
are managed in the same enterprise resource planning system
SAP. The operational responsible acts with impartiality and a cost

M. Guajardo, M. Rnnqvist / European Journal of Operational Research 240 (2015) 147159

efcient criterion, therefore the solution is also driven by an overall


cost minimization objective.
Although in both cases all players are willing to conform to the
suggested coalitions, we would like to prevent deviations that can
affect the stability of the coalition structure. We quote Hart and
Kurz (1983) to emphasize that forming a coalition does not eliminate the individual players as decision makers. In all interactions
with the other players, coalition members act as one unit (it may
be useful to think of a representative agent taking their place);
however, this arrangement will continue only as long as each
player nds it desirable to act this way. For this purpose, they
refer to the notion of strong equilibrium introduced by Aumann
(1959). This refers not only to the stability within the formed coalitions, but also in the whole structure, so that for all players there
would be no incentives to form a coalition that is not contained
in the structure. For example, let us assume a structure contains
coalitions Z 1 and Z 2 . The coalition structure together with a cost
allocation vector u is not strongly stable if coalitions k1 # Z 1 and
k2 # Z 2 exist such that the sum of the costs allocated to the players
in k1 [ k2 is greater than the cost C k1 [k2 .
Given the set of players N, the cost parameter values C Z for all
coalitions and a positive integer m, we are interested in two version of the coalition structure problem: the core stability problem
and the strong stability problem. The rst one consists of nding a
coalition structure fZ 1 ; . . . ; Z r g and a cost allocation
u u1 ; . . . ; un such that:

3. MIP models
In the following, we formulate mixed integer programming
models for the core stability and the strong stability versions of
the coalition structure problem. First, let us introduce the notation
on sets, parameters and decision variables.
Sets
N: set of players.
K: set of coalitions.
Parameters
C k : Total cost of coalition k.

aj;k

bk;k


The strong stability problem is the same as the rst problem,


but it also requires that the coalition structure together with
the cost allocation vector u satises for all coalition Z (jZj 6 m)
P
not contained in the condition j2Z uj 6 C Z . The set of solutions
; u to this strong stability problem composes the core of a game
in the coalition structure sense dened by Aumann and Dreze
(1974).
To our knowledge, no general optimization model has been
proposed for these problems. The few works in OR/MS addressing
related issues have rather focused in more particular settings.
Also, they have adopted a predictive perspective on how the
agents will group rather than a decision making perspective on
how the agents should group. Such predictive perspective is presented by Axelrod et al. (1995), in a setting of alliances forming
to develop and sponsor technical standards. Nagarajan and Sosic
(2009) focus on a dynamic setting where different suppliers can
create alliances for selling components to an assembler. They
assume the coalition members divide the prot equally, unlike
our case where the allocations are decision variables of the problem and, therefore, may differ from one to another member. Sosic
(2006, 2010) incorporates the notion of farsighted stability, proposed by Chwe (1994), in an inventory sharing setting and in an
information sharing setting rstly approached by Leng and
Parlar (2009). In the dynamic of these games, the outcomes are
described by a set of possibilities within several time periods
and the relevant information is common knowledge for all players.
In contrast, we focus on single-period games with side payments,
i.e., where the outcome to a coalition can be captured by just one
real number, namely the characteristic function (Kaneko &
Wooders, 2004), and this function is not necessarily known by
all the players.

1 if player j belongs to coalition k


0 

8
>
<1
>
:

 is a sub  coalition of coalition k i:e:; all


if coalition k
players in k belong to k


Variables

xk

(i) the maximum cardinality per coalition constraint is satised,


i.e., jZj 6 m 8Z 2 ;
(ii) the core stability constraints are satised within each Z, i.e.,
P
P
u 6 C Z 8Z  Z; Z 2 and j2Z uj C Z 8Z 2 ;
j2Z j
P
(iii) the cost of the structure, C ri1 C Zi , is the minimum
cost that can be achieved by a structure that satises conditions
(i) and (ii).

149

1 if coalition k is formed
0 

uj;k : Cost allocated to player j in coalition k.

3.1. The structure of core-stable coalitions at minimum cost


The following mixed integer linear programming model looks
for a structure such that all coalitions k 2 are stable and the
total allocated cost is minimized.
Objective function

min

XX
uj;k

j2N k2K

Constraints

 k 2 K : b 1
aj;k  uj;k 6 C k  xk 8k;
k;k

j2N

aj;k  uj;k C k  xk 8k 2 K

j2N

aj;k  xk 1 8j 2 N

k2K

xk 2 f0; 1g; uj;k P 0 8j 2 N; k 2 K

We refer to this model by M1. Objective function (3) minimizes the


total allocated cost among the players. Constraints (4) correspond to
the rationality conditions, which state that in the resulting coalition
  k such that should they form
structure there is no set of players k
a coalition separately from the rest they would perceive less total cost
than the total cost allocated to them in the allocation u. Constraints (5)
correspond to the efciency condition, which states that the sum of
the costs allocated to all the players in a formed coalition equals the
optimal cost of such coalition. Constraints (6) state that each player
is assigned to one and only one coalition, thus the solution to the
model, if such exists, denes a partition of set N. Constraints (7) state
the binary nature of variables x and the non-negativity of variables u.
Note that constraints (4) and (5) recover the stability and efciency constraints (1) and (2) for any coalition k such that xk 1,
so that they provide the allocations uj;k are core-stable within each
formed coalition k. Since the objective function minimizes the total

150

M. Guajardo, M. Rnnqvist / European Journal of Operational Research 240 (2015) 147159

allocated cost, the optimal value of variable uj;k is zero for all j; k
such that coalition k is not formed and also for all j; k such that
player j does not belong to coalition k.
Model M1 is a mixed integer programming model that solves a
set partitioning problem, subject to the stability constraints. A
solution to this model exists (in the worst case where collaboration
does not achieve any savings, the structure of all players standing
alone is stable). If we knew beforehand that the core of each coalition (or at least the core of each of those coalitions that minimize
cost) is not empty, an equivalent approach to solve the problem
could be given by the following two-stages procedures. In the rst
stage, a pure set partitioning problem is solved, for instance,
through an integer programming model in the binary variables
P
xk , dened by the objective function min k2K C k xk subject to constraints (6). In the second stage, a stable allocation must be determined for each coalition k such that in the optimal solution to the
set partitioning model xk 1. However, the information on the
non-emptiness of the core in our setting is not known beforehand,
since we do not make any related assumption on the cost parameter values C Z . That is why it is important to consider the coalition
structure and the cost allocation problems simultaneously in the
same model. Then, if a solution (; u) to this model is found, it
assures the cost allocation vector u satises the stability conditions
within each coalition in the structure . Then, the implementation
of the structure could re-allocate the cost within each formed coalition through any other desirable allocation rule.
3.2. The structure of strongly stable coalitions at minimum cost
The coalition structure resulting from model M1 assures that
players from the same coalition that is formed have no incentives
to deviate and create their own coalition. However, the structure
does not assure that players who belong to different coalitions
have no incentive to conform a new coalition. This needs incorporating the notion of strong equilibrium, in order to assure not only
stability within the formed coalitions, but also in the whole structure, so that for players who were assigned to different coalitions
there would be no incentives to form a different coalition. We formulate a mixed integer linear programming model for this problem in a similar way to model M1, maintaining objective
function (3) and constraints (5)(7), but replacing constraints (4)
by the following constraints:

X
j2N

aj;k 

X
uj;k

!
6 C k 8k 2 K:


k2K

We refer to this model by M2. Constraints (8) correspond to a strong


stability condition, assuring that no players, either from the same or
different coalitions in the structure, have incentive to form their
P
own coalition. Notice k2K
 is the actual allocation to player j in
 uj;k
the coalition it will belong to. In fact, for all the other coalitions,
its cost uj;k will be zero because of the objective function (using
another objective function may require addition of the following
constraint to assure that: uj;k 6 L  xk 8j 2 N; k 2 K, where L is a sufP
P
ciently large number). Then, the sum
 repre uj;k
j2N aj;k 
k2K
sents the sum of all costs allocated to players of coalition k. For
the structure to be strongly stable, we need this sum to be less than
or equal to C k , for all k 2 K, even if coalition k is not formed.
3.3. Maximum cardinality constraint
Concerned about the complications of managing large coalitions
in real-world situations, an interesting question is how to structure
the players in sub-coalitions such that a maximum cardinality constraint is satised. That is, each structured coalition must contain
at most m players, where m is a positive integer parameter that

represents the maximum number of players allowed in each coalition. The size of a coalition has been identied by DAmours and
Rnnqvist (2010) as an important aspect in collaborative logistics.
In practice, the upper bound on the coalitions size might be motivated for several reasons. For example, Lozano et al. (2013) point
out that although transportation costs are always reduced through
collaboration, as the number of partners grows coordinating the
cooperation becomes more problematic and/or costly, which is
one of the reasons why in practice cooperation in transport usually
involves just a few partners. Other reasons are argued by Stein
(2010), who remarks that the formation of coalitions might be
restricted due to structural issues, laws and culture ideology. He
refers to some other references in restrictive cooperative game theory (so called network games), and remarks that although some
of them date from early times, the literature in this eld is still
sparse and fragmentary. As another reason, in one of the cases that
primarily motivated our work, we have observed that managers
are subject to some political issues which do not allow the implementation of a grand coalition, but only of smaller sub-coalitions.
For example, in our inventory management case, it might be clear
that because of the risk pooling effect, a completely centralized
solution performs better in terms of costs than the decentralized
solution where there is no collaboration. However, in order to
change a previously implemented decentralized solution, the managers decision needs to consider that a centralized solution would
perhaps need signicantly less staff. In some organizations this
might involve some difcult issues.
In general, all these reasons affecting the maximum cardinality
of a coalition may not be easy to reect in the cost functions.
Therefore, it becomes relevant to analyze the coalition structure
problem subject to a maximum cardinality constraint. An obvious
alternative to incorporating this condition in models M1 an M2 is
to simply re-dene the set K including only coalitions of cardinality
less than or equal to m (or just to set xk 0 8k : Sk > m, where Sk is
a parameter indicating the cardinality of coalition k). For an upper
bound m, we will refer by fm and f m to the optimal objective value
of models M1 and M2, respectively.
Note that the coalition structure problem subject to the maximum cardinality constraint presents similarities with k-partitioning problems, which arise in multiprocessor scheduling of jobs
on machines at minimum makespan (Babel, Kellerer, & Kotov,
1998, Kellerer and Kotov, 2011). These problems also look for a
partition into subsets subject to a cardinality limit. However, while
in k-partitioning problems the weight (processing time or cost)
depends only on the item and not in the subset (machine) it is
assigned to, in the coalition structure problem the cost is associated to the subset rather than the sum of the costs of its members.
Also, the coalition structure problem includes the stability constraints and the allocation variables which do not appear in k-partitioning problems.
3.4. Relation between the solutions of models M1 and M2
In the following, we present a proposition that relates the solutions of models M1 and M2.
Proposition 1. If the strongly stable coalition structure model (M2) is
feasible, then its optimal objective value is equal to the optimal
objective value of the core-stable coalition structure model (M1).
Likewise, if model M2 is feasible, a coalition structure that solves any
of the two models, also solves the other model.
This relationship between the solutions to the two models is
fairly intuitive because the strong stability condition is more
restrictive than the core stability condition. In fact, since constraint
 for
(8) must hold for all k 2 K, in particular it must hold for any k

M. Guajardo, M. Rnnqvist / European Journal of Operational Research 240 (2015) 147159

which constraint (4) is formulated. This gives the inequality


P
P
 
 . Since uj;k 0 for all j and k such that
^ uj;k
^ 6 Ck
j2N aj;k
k2K
xk 0, then the left-hand side of this inequality is equal to
P
  uj;k , where k is such that xk 1. Then, it follows that if a
j2N aj;k
coalition structure together with a cost allocation vector satises
constraints (8), then it also satises constraints (4). Since all the
other constraints and the objective function are the same in both
models, the optimal objective value f  in M2, if such exists, cannot
be lower than the optimal objective value f  in M1. This proves that
f  6 f  . Now, to prove the equality, let us argue by contradiction.
 u
 an optimal
Let ; u be an optimal solution to model M1 and ;
P P
P

solution to model M2. Then, f j2N k2K uj;k Z2 C Z and
P
P
P
P
P
P

f 


 j;k .
u
 C Z . If f < f , then
j2N
k2K uj;k
Z2
Z2 C Z <
Pj2N k2K
 j;Z . That
Then, at least one coalition Z 2 exists such that C Z < j2Z u
is, constraint (8) is violated for coalition Z, but then model M2
would be infeasible. We conclude that if model M2 is feasible, then
the optimal objective value of model M1 is equal to the optimal
objective value of model M2.
A straightforward implication is that if the structure
 fZ k : xk 1g solves model M1 for some cost allocation vector
u, then this same structure must solve model M2 for some cost
 (if model M2 is feasible). This does not need
allocation vector u
 . Note that, given the coalition
the allocation u to be the same as u
structure is xed, a cost allocation vector u which solves model M1
does not necessarily solves model M2, but an allocation that solves
model M2 must also solve model M1.
Let us consider an example consisting of three players, with a
characteristic function dened as follows: C f1g 50; C f2g 65;
C f3g 70; C f1;2g 100; C f1;3g 115; C f2;3g 130; C f1;2;3g 180. The
optimal structure obtained by model M1 in this problem is
 ff1; 2g; f3gg, with objective function value f  170. The allocation vector u 49; 51; 70 for this structure solves model M1,
since the coalitions f1; 2g and f3g with this allocation satisfy the
core stability constraints within each of these two coalitions. However, this allocation does not satisfy the strong stability condition,
because u1 u3 119 > C f1;3g 115. Therefore, there would be
incentives for players 1 and 3 to form their own coalition. For
example, they could agree on allocations u1 48 and u3 67 to
split the cost of coalition f1; 3g and leave player 2 alone. The resulting structure ff1; 3g; f2gg is sub-optimal, with objective value
f 180, 10 units more expensive than the optimal. There exist innite other allocations that can overcome this problem. For exam 44; 56; 70 together with the same
ple, the allocation vector u
structure  solve both models M1 and M2. In fact, this allocation
satises the core stability constraints within each formed coalition
and the strong stability condition, while minimizing the total cost.
4. Applications
In this section we present two examples of applications of the
MIP models presented in the previous section. The rst one corresponds to a problem in forest transportation, where different players can collaborate by means of backhauling and wood bartering.
The second one corresponds to a problem of inventory of spare
parts for oil operations, where different players can collaborate
by means of inventory pooling. In order to nd optimal coalition
structures and cost allocations, we proceed in two stages. First,
we solve the corresponding cost minimization problem (transportation or inventory) for each possible coalition k. This cost minimization problem provides an optimal plan for each coalition and the
characteristic function C k of the cooperative game. In the second
stage we use such cost as parameters in the MIP models and solve
them to nd the coalition structure and cost allocation. Then, the
coalitions contained in the structure implement the collaboration

151

according to the corresponding plan, which was obtained when


solving the cost minimization problem.
4.1. Forest transportation
Frisk et al. (2010) studied a problem of cost allocation in collaborative forest transportation, motivated by a real-world case. The
case study was done by the Forest Research Institute of Sweden
and included eight companies with operations in southern Sweden. In this context, collaboration among companies creates opportunities to better utilize their overlaps and distances to mills. The
authors studied a number of instances of this problem and found
that savings of up to 14.2% can be achieved by implementing collaboration. This can be done by wood bartering and/or backhauling. Wood bartering can be used in such a way that destinations
between supply and demand nodes are changed. This means that
one company may send some of its own supply to another companys mill, and in return get the same volume back to its own mill
but from another company. This can be achieved when the supply
areas of the companies cover each other as this lowers the overall
average transportation distances. The ows can be direct ows
between a supply and a demand point or more complex routes
where backhauling possibilities are included. Backhauling can be
used to nd better routes by combining two or several destinations
(combination of supply and demand points) to reduce the proportion of empty driving. All computations were done in the decision
support system Flow-Opt (Forsberg, Frisk, & Rnnqvist, 2005). This
system was used to solve a transportation model for each possible
coalition of companies. The transportation model is a column
based model, where the decision variables represent the ows in
routes or backhauls. The model to nd the coalition costs C k can
be written as follows:

min C k
s:t:

cp wp

p2P k

aijp wp 6 sij ; 8i 2 Ij ; j 2 J k

10

dljp wp dlj ; 8l 2 Lj ; j 2 J k

11

p2Pk

p2Pk

wp P 0; 8p 2 Pk

12

The variable wp denotes the ow in backhaul route p and cp the corresponding unit cost. The set of routes for each coalition is given by
the set Pk . The coefcients aijp have value 1 if route p picks up at
supply point i at company j and 0 otherwise. In the same way dljp
has value 1 if route p delivers at demand point l at company j and
0 otherwise. The set J k denotes the set of companies in coalition k,
and Ij and Lj the set of supply and demand points, respectively, at
company j. Constraint sets (10) and (11) represent the supply and
demand respectively at each company j in coalition k, and (12)
the non-negativity conditions. We note that direct ow variables
are represented using columns including only 2 nonzero elements
(i.e. one pick up and one delivery). There is often a huge number
of potential backhaul routes (often several hundred millions) and
they can typically not be used explicitly in a solver. Instead, they
are generated in a column generation approach (Carlsson &
Rnnqvist, 2007). The restrictions imposed on the backhaul routes
are e.g. driving time or the number of pick ups and deliveries. In tactical planning, there is often a practical limit for using backhaul
routes consisting of two direct ows.
After having computed the costs C k for all coalition k 2 K, we
use them as parameters in the MIP models of Section 3. We implement these models in AMPL and solve them using the software
CPLEX 12.5 on an Intel Core 2 Duo 2.26 gigahertz processor with
8 gigabyte of RAM. The solutions to models M1 and M2 in these
instances are quickly obtained, in a matter of a second. Note, since

152

M. Guajardo, M. Rnnqvist / European Journal of Operational Research 240 (2015) 147159

there are eight companies, there are 255 coalitions, and consequently 255 binary variables xk . In addition, there are 2040 continuous variables uj;k and 6050 rationality constraints.
The results of solving the models for different values of the
maximum cardinality bound m are shown in Table 1.
From these results we can observe that the optimal value fm of
model M1 is strictly decreasing in m. This means that the structure
of stable coalitions that minimizes total cost always include at
least one coalition of the maximum cardinality allowed by the
upper bound m. This suggests that the more companies are
allowed to collaborate, the less total cost, as expected in this case.
However, the relative cost savings of including an additional company decreases in most cases when increasing the upper bound
from m to m 1 (except when m 6). These relative cost saving
 

values, calculated as jfm1
 fm j=fm  100%, are shown in the
fourth column of Table 1 and plotted in Fig. 1. Note from m 7
to m 8 the relative savings amount to only 0.08%.
The effect of increasing the maximum cardinality allowed for a
coalition is very important in the lowest interval and decreases
with higher m. Fig. 2 shows the cumulative savings with respect
to the base case m 1, i.e., jfm  f1 j=jf8  f1 j  100%. It can be
observed that the cumulative savings of allowing coalitions of cardinality up to m 3 already account for more than a half of the
total savings achieved by the grand coalition. And when setting
m 6, more than 90% of these savings are achieved.
This analysis provides insights about the marginal savings of
allowing one more player into the coalitions. These marginal savings becomes considerably less signicant when m approaches
the total number of players. In consequence, the decision maker
in practice should analyze whether these savings overwhelm or
not the managerial complexities of forming larger coalitions.
In order to illustrate how the structures change from one to
another value of m, we have included Fig. 3. This gure shows
the coalition structures resulting for m 2 f2; . . . ; 7g (for m 1 the

Table 1
Optimal objective values, relative savings and cumulative savings of the application in
collaborative forest transportation.
m
1
2
3
4
5
6
7
8

fm

f 
m

42.96
41.24
40.58
40.12
39.70
39.52
39.28
39.25

42.96
41.24
40.58
40.12
39.70
Infeasible
39.28
39.25


fm j
jfm1

fm

4.00
1.60
1.13
1.05
0.45
0.61
0.08

jfm f1 j
%
jf8 f1 j
46
64
77
88
93
99
100

optimal coalition structure obviously consists of all players standing alone, while for m 8 it consists of the grand coalition). The
driver to create a coalition depends on the participating companies, geographical distribution and their assortment distribution.
If the companies have more geographical overlap, it creates better
potential savings. In the case study, there are close to 40 assortments based on specie, diameter, length and quality. Any delivery
between companies must be balanced for each assortment.
Changing from m 2 to m 3 makes player P3 to join P2 and
P5, and player P4 to join P1 and P6. This leaves players P7 and
P8 alone. In this instance, these two player have no savings when
acting in coalition (moreover, they would incur more costs by
implementing the coalition because in this instance
C fP7;P8g > C fP7g C fP8g ). Changing from m 3 to m 4 involves
three deviations. For m P 4, the collaboration relations vary less,
with only one deviation when increasing the upper bound by
one. It can also be observed that only players P2 and P5 remain
together in the same coalition for all m 2 f2; . . . ; 7g. These observations illustrate that the upper bound on the cardinality may have
important implications.
Another observation from the results in Table 1 is that model
M1 always had a solution, which is not surprising because a coalition structure satisfying the core stability within each coalition
formed always exists. Also, for all m 6, model M2 was feasible
and, consistently with Proposition 1, the optimal objective value
was the same as the optimal objective value of model M1. Interestingly, when m 6 there exists no strongly stable coalition structure, thus if the upper bound would be xed to a maximum of
six companies per coalition, there will always be incentives for
some of them to deviate from their coalition to join with other
players. For example, the solution we obtained by running model
M1 to this instance results in the structure  fZ 1 ; Z 2 g, where
Z 1 fP1; P2; P3; P5; P6; P7g and Z 2 fP4; P8g. This structure
reaches the optimal objective value f  39:52. In the obtained
allocation,
we verify that u3 u4 4:69 2:07 6:76 >
C fP3;P4g 6:73. That is, players P3 and P4 would have incentives
to form their own coalition. However, should we add a constraint
for this example requiring that players P3 and P4 must belong to
the same coalition, the resulting structure reaches a sub-optimal
objective value equal to 39.53 cost units and players P1 and P4
would have incentives to form their own coalitions. These types
of features are not easy to identify beforehand, especially since
the characteristic function is a sort of black box with no particular
properties. Therefore, the modeling approach we present can serve
as a useful support tool for the decision maker.
4.1.1. Individual cost minimization
When the strong equilibrium does not exist, it is worthy to
explore the possibilities for some of the players to conform

Fig. 1. Relative cost savings from allowing one more player in the maximum allowed cardinality of a coalition.

153

M. Guajardo, M. Rnnqvist / European Journal of Operational Research 240 (2015) 147159

Fig. 2. Cumulative cost savings from allowing one more player in the maximum allowed cardinality of a coalition.

Fig. 3. Optimal coalition structures for different upper bounds m in the forestry
case.

coalitions. It can be argued that the goal of any individual player is


to minimize his own cost while keeping stability for all the players
who conform his coalition. We tackle this problem by introducing
some modications to model M1. Without lost of generality, let
player j 1 be the one for which we minimize the cost. Let wj be
a binary decision variable, which takes the value 1 if player j joins
the coalition of player P1, and zero otherwise. In order to minimize
the total cost of P1, we replace the objective function (3) by the
following one:

min

X
u1;k

13

obtained when using the individual cost minimization for each


player in the instance with m 6.
It can be seen that the solution to the problem of a given player j
may result in a cost uj greater than the cost for this same player j in
the solution to the problem of another player i. Moreover, this is even
possible when the coalition of player j is the same in both problems.
For example, when solving the model for P1, he gets allocated
u1 3:76, which is more costly for him than the solution obtained
when solving the model for player P4, despite for both problems
the optimal structure is conformed by Z 1 fP1; P2; P3; P5; P6; P7g
and Z 2 fP4; P8g. As in farsighted coalitional stability (Chwe,
1994; Sosic, 2006), this is rational in order for player P1 to keep stability of the coalition it conforms, despite being opportunities for
him to collaborate with players who are left out of his coalition.
While such opportunities may be better for P1 from a myopic perspective, if he breaks out then some other moves would trigger
and the overall situation becomes instable, which is worse for him.
In fact, if constraint (15) is formulated for j 1, the model turns
infeasible. When the model is solved for j 4, the best alternative
for player P4 is to form a coalition with player P8. Assuming this coalition fP4; P8g forms rst (and does not break), the remaining universe of players is such that P1 can get a better (lower) allocation
because the opportunities for the players to collaborate with P4
and P8 are not longer possible. This suggests that in a game where
a single player motivates the formation of a coalition, the nal structure may strongly depend on who is this player. Note, however, that
the individual cost minimization problem may require this player to
have the information from all players, which in practice can turn to
be inviable as we pointed out in Section 2.

k2K

As for the constraints, we keep constraints (4)(7) from model


M1 and, in addition, we incorporate the following constraints:

uj;k 6 M  xk 8j 2 N; k 2 K; M  0
X
X
ai;k 
ui;k 6 C k  wj M  1  wj

k2K

i2N

8k 2 K; j 2 N; j 1; aj;k 1; M  0
aj;k  xk a1;k  xk 6 1 wj 8k 2 K; j 2 N; j 1
wj 6

14

aj;k  a1;k  xk 8j 2 N; j 1

15

4.1.2. Bounding the maximum difference between relative savings


In their case study, Frisk et al. (2010) discussed the suitability of
traditional cost allocation methods for sharing the costs in collaborative forest transportation, such as the Shapley values and the
nucleolus. When it came to the acceptance of the cost allocation
among the companies participating in the case study, the authors
found some disadvantages of using these traditional methods,

16
17

k2K

Constraints (14) assure that uj;k is greater or equal than zero only if
coalition k forms (M is a sufciently large number, e.g. M C k ).
Constraints (15) provide strong stability for the players who join
the coalition of player 1. Constraints (16) and (17) are logical relationships assuring that wj takes value 1 if and only if player j joins
the coalition conformed by player 1. Note that the strong stability
condition in this model is not formulated for player 1, but only
for those players he collaborates with. Table 2 shows the allocations

Table 2
Optimal allocations obtained for the individual cost minimization problems.
O.F.
Min
Min
Min
Min
Min
Min
Min
Min

u1
u2
u3
u4
u5
u6
u7
u8

u1

u2

u3

u4

u5

u6

u7

u8

TOTAL

3.76
3.78
3.78
3.67
3.78
3.78
3.52
3.52

12.86
14.86
14.86
14.07
14.86
14.86
14.7
14.7

4.13
4.74
4.74
3.9
4.74
4.74
3.76
3.76

2.05
2.07
2.07
2.05
2.07
2.07
1.81
1.81

10.11
10.34
10.34
9.54
10.34
10.34
9.44
9.44

4.54
4.96
4.96
4.39
4.96
4.96
4.27
4.27

1.74
1.88
1.88
1.57
1.88
1.88
1.88
1.88

0.33
0.33
0.33
0.33
0.33
0.33
0.33
0.33

39.52
42.96
42.96
39.52
42.96
42.96
39.71
39.71

154

M. Guajardo, M. Rnnqvist / European Journal of Operational Research 240 (2015) 147159

because of the large difference in size among the companies. In


order to overcome this problem, they proposed a new method
called Equal Prot Method (EPM). This method looks for a stable allocation such that the relative savings are as similar as possible for all
participants. Such allocation solves a linear programming model,
whose objective function is to minimize the maximum difference
in pairwise relative savings between the players. Being g a continuous variable, such a linear programming model minimizes g subject
u
to the core stability constraints and to g P Cufigi  C fjgj 8i; j 2 N. A similar principle of fairness can be extended to our problem, by considering the difference in pairwise relative savings between coalitions
instead of players. However, the minimization of the maximum of
these differences would lead to a coalition structure where all
player stand alone, because in this fully non-cooperative structure
all coalitions save zero and, therefore, the difference among their
relative savings is also zero (which is the minimum that can be
achieved). Therefore, we rather keep the objective function (3) that
minimizes the total allocated cost, and introduce a parameter u as
an upper bound on this difference. We also introduce a binary var form and 0
iable yk;k , which takes the value 1 if coalitions k and k
otherwise. Together with these variables, we incorporate the following constraints to the models:

yk;k P

Ck
j2N aj;k C j

P

yk;k 6 xk

8k; k 2 K

yk;k 6 xk

8k; k 2 K

xk xk 6 1 yk;k

C k
j2N

aj;k C j

2K
6 u 8k; k

18
19
20

8k; k 2 K

21

Constraints (18) restrict the difference between the relative savings


among the formed coalitions to be no greater than the upper bound
u. Constraints (19)(21) are logical relationships linking variables x
and y, that provide variable yk;k being 1 if and only if xk and xk are 1.
Table 3 shows the results obtained in this forestry case using
ve different values of u ranging from 0.1% to 10%.
It can be observed that relatively low values of u often leaves
the strong stability problem without any feasible solution. This
can be explained because the coalition structure of minimum cost
does not necessarily fulll the constraint on the bound u. In fact
only for u 10% do the results match with the minimum cost
solutions (which were reported in Table 1). Comparing the optimal
objective values with those obtained by the original core stability
model gives a sort of price of fairness, understanding by fairness
the EPMs notion of equality in relative savings among coalitions.
This price of fairness obviously turns to be higher for lower values
of u, where the space of feasible solutions becomes smaller, while
lower for higher values of u. A comparison of the values of fm in
Table 3 with those in Table 1, reveals the solutions for
m 2 f2; . . . ; 7g when u 0:1% are on average 3:7% more expensive
than the minimum cost solution. When u 7:5% the results are
on average only 0.3% more expensive than the minimum cost

solutions. When u 10% the solutions coincide, because the minimum cost solution satises the upper bound constraint on the
relative savings difference.
4.1.3. The largest cardinality of a stable coalition
Looking for the maximum number of players in a game for
which the core exists is mentioned by Frisk et al. (2010) as an alternative to the EPMs objective function. How this subgroup of players should be selected was, however, left for further research in
their article. We can easily answer this question by using model
M1 with the following two modications. First, the following
objective function is used:

max ^f

S k  xk ;

22

k2K

where Sk is a parameter dened as the cardinality or number of


players of coalition k, for all k 2 K.
Second, the following constraint is added:

xk 1

23

k2K

Constraint (23) states that only one coalition will be formed, while
objective function (22) maximizes the number of players in it.
When running the resulting model for this instance, we obtain
^f jNj 8. This is not surprising, since it was known from Frisk
et al. (2010) that the core of the game considering the grand coalition in this instance was not empty.
4.1.4. Cost per cardinality
Adding a cost on the coalition cardinality might be useful for
reecting the issues of managing large coalitions. How this cost
should be calculated is, however, a tough task. As we discussed
earlier, these issues are not easy to translate into a cost function.
In order to illustrate the effects that this type of cost could have,
we introduce an additional cost f on each coalition. We assume
that f can be modeled as a convex increasing function of the cardinality of the coalition. This cost does not affect the structure of the
models, but just modies the cost parameter values C in constraints (4), (5) and (8). Namely, the new cost parameter value
b Z C Z fZ.
for coalition Z is C
In our numerical runs, we utilize the function fZ a jZjjZj1
,
jZj1
where jZj is the cardinality of coalition Z and a is a positive parameter. Table 4 shows the results obtained using ve different values
of a.
Since the cost on the cardinality of the coalitions f increase with
a, the relative savings of allowing m 1 players with respect to m
players decrease with a. When f is relatively small, savings are still
realized for the largest values of m; for example, when a 0:1, setting m 7 leads to savings of 0.40% with respect to the case where
m 6. When f is relatively high, the large coalitions do not form,
thus increasing the upper bound on the number of payers per coalition has no impact. For example, when a 0:5 any bound m P 4

Table 3
Optimal objective values of the instances considering a bound u on the maximum difference between relative savings among coalitions.
m

1
2
3
4
5
6
7
8

u 0:1%

u 2:5%

u 5%

u 7:5%

u 10%

fm

f 
m

fm

f 
m

fm

f 
m

fm

f 
m

fm

f 
m

42.96
42.95
42.95
40.87
40.84
40.84
40.84
39.25

42.96
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
39.25

42.96
42.25
41.62
40.34
40.20
40.19
40.19
39.25

42.96
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
39.25

42.96
41.59
40.70
40.16
40.13
40.13
40.13
39.25

42.96
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
Infeasible
39.25

42.96
41.24
40.58
40.12
39.70
39.70
39.70
39.25

42.96
41.24
40.58
40.12
39.70
Infeasible
Infeasible
39.25

42.96
41.24
40.58
40.12
39.70
39.52
39.28
39.25

42.96
41.24
40.58
40.12
39.70
Infeasible
39.28
39.25

155

M. Guajardo, M. Rnnqvist / European Journal of Operational Research 240 (2015) 147159


Table 4
Optimal objective values and relative savings of the instances considering an additional cost per cardinality of the coalitions.
m

a 0:1
fm

1
2
3
4
5
6
7
8

42.96
41.47
40.88
40.46
40.12
39.97
39.81
39.81

a 0:2

jfm1
fm j
%
fm

3.47
1.42
1.03
0.84
0.37
0.40
0.00

fm
42.96
41.65
41.13
40.71
40.47
40.40
40.33
40.33

a 0:3

jfm1
fm j
%
fm

3.05
1.25
1.02
0.59
0.17
0.17
0.00


jfm1
fm j
%
fm

42.96
41.86
41.35
40.95
40.80
40.80
40.80
40.80

leads to the same solutions, so coalitions of cardinality greater than


4 are not part of the optimal coalition structure.
When considering the cost f, the strong stability model has
solution for all values of m in the four instances where a P 0:2.
This happens still in the case where m 6, which without this cost
had turned to be infeasible.
As for the optimal structures, for relatively small values of m,
we observe the most convenient coalitions tend to form despite
the increase in cost f. For example, when m 2 we nd four coalitions take part in the optimal structure for all values of a. However,
when m increases, the impact of changing a in the optimal structure is more notorious. For example, when xing m 8, for
a 0:1 the optimal structure contains only two coalitions, while
for a 0:5 it contains ve coalitions.
These results illustrate that a cost on the cardinality of the coalitions may have an important effect in the resulting coalition
structure.
4.2. Inventory pools in oil operations
Guajardo et al. (2014) studied a case of inventory of spare parts
in an oil and gas company headquartered in Norway. The operations at the platforms offshore involve sophisticated equipment,
usually expensive and subject to long lead times. Therefore, the
inventory of spare parts plays an important role to facilitate operating conditions. Since some identical pieces of equipment are
used at different platforms, pooling inventory presents important
opportunities for savings. The authors estimated cost savings of
about 20% from risk pooling in comparison to the case where
inventory is carried separately for each platform. The spare parts
are held at inventory plants or licenses located along the Norwegian coast. Some of these plants lie right next to each other, within
a same warehouse. Although the company is the majority owner of
the licenses, there is also a set of other companies owning a share
of them. Given the important consequences of stock out in oil operations, in terms of safety and downtime costs, an approach divided
in stages as in Anupindi and Bassok (1999) and Granot and Sosic
(2003) where the stage of collaboration is agreed after demand is
realized, is not appropriate in this case. The coalitions rather need
to be formed beforehand. On the other hand, due to several reasons, including political, operational and managerial difculties,
it is not straightforward to implement an inventory pool of spare
parts to cope centrally with the demand from all the platforms.
For each coalition, we consider a single-item problem under
continuous review (S  1; S) policy with xed lead time, Poisson
distributed demand and backorders. The optimal base-stock S for
each coalition k is found through a procedure aimed at minimizing
ordering plus holding cost subject to a ll-rate constraint, that is, S
is the positive integer number that solves the following inventory
control problem:

min C k Alk h  S  lk L Bk S

a 0:4

fm

24

2.56
1.22
0.97
0.37
0.00
0.00
0.00

s:t:

fm
42.96
42.03
41.53
41.19
41.13
41.13
41.13
41.13

X l Lj elk L
k
Pk
j!
j6S1

S 2 Z

a 0:5

jfm1
fm j
%
fm

2.16
1.19
0.82
0.15
0.00
0.00
0.00

fm
42.96
42.15
41.68
41.43
41.43
41.43
41.43
41.43


jfm1
fm j
%
fm

1.89
1.12
0.60
0.00
0.00
0.00
0.00

25

26

In the cost function (24), A is the xed cost per order, lk is the average demand of coalition k per unit of time, h is the carrying cost, L is
the lead time and Bk S is the expected number of backorders given
P1
j lk L
by
=j!. Constraint (25) states a service level
jS j  Slk L e
constraint, where the left-hand side IS is the ll-rate as a function
of S and k 2 0; 1 is a parameter expressing the target service level.
All these expressions come from standard formulae in inventory
control literature (see e.g. Hadley & Whitin, 1963). The solution to
this inventory control problem is simply the minimum positive
integer S that satises the service level constraint (25). This optimal base-stock level S can be computed by evaluating the ll-rate
starting from S 1 and consecutively increasing S by one, until the
target k is fullled. This procedure tends to nish quickly in inventory control of spare parts, where the optimal base-stock levels are
relatively low quantities. It not rarely happens that the optimal
base-stock level for a coalition is lower than the sum of the optimal
base-stock levels of its players standing alone and, therefore, the
potential for collaboration appears. We implement and solve this
inventory control problem by using Matlab. The solution to each
instance is obtained in a matter of a second. More details about this
case can be found in Guajardo et al. (2014) and Guajardo and
Rnnqvist (2014).
We test our models in an instance including 16 players (each
representing an inventory plant), 19 spare parts and 53 possible
coalitions. These coalitions result from limiting the possibilities
to only those where at least one item is used by all the players in
the same coalition. The resulting coalitions include 2 sets of 4 players, 13 sets of 3 players, 22 sets of 2 players and 16 single-player
coalitions. All players in this instance are somehow related with
each other, either because they can directly collaborate in a same
coalition or because another coalition includes one of its collaborator or another descendent in the collaborative relations. Therefore,
the problem of nding the best structure cannot be separated in
sub-problems for some players beforehand.
After having computed the costs C k for all coalition k 2 K, we
use them as parameters in the MIP models of Section 3. The coalition structure resulting from the optimal solution to model M1
without cardinality constraints (equivalently, when m 4),
includes eight coalitions as illustrated in Fig. 4. This structure also
solves model M2. What we believe is interesting from this instance
is that only one of the two coalitions of four players (the largest
cardinality of the possible coalitions) are part of the optimal solution. In fact, coalition fP1; P3; P4; P5g is not formed. Although P3
contributes to coalition fP1; P4; P5g (because the cost parameters
are such that C fP1;P3;P4;P5g 83; 424 < C fP3g C fP1;P4;P5g 90; 984),
more savings can be achieved if P3 leaves this coalition to join with

156

M. Guajardo, M. Rnnqvist / European Journal of Operational Research 240 (2015) 147159

sense that the target service level for a coalition between players
i and j can be satised by a base-stock level lower than the sum
of the base-stock levels for the separate players. Because of the discrete nature of the base-stock level, the solution S to the inventory
control problem often provides some slack in constraint (25), that
is, lS > k. This encourages some players to nd collaborators to
share the inventory costs with, while conforming to the target service levels. These drivers explain that some of the players are especially good collaborators between each other. For example, player
P1 with P4; P2 with P3; P7 with P8 and P9 with P10, who as can
be seen in Fig. 4 remain together in the solution to all the instances.
4.3. Discussion

Fig. 4. Optimal coalition structures for different upper bounds m in the oil case.

P2 only. In fact, the cost parameters give C fP2;P3g C fP1;P4;P5g


105; 089 < C fP2g C fP1;P3;P4;P5g 133; 818. In contrast, the other
coalition of largest cardinality, fP12; P13; P15; P16g, is formed.
Other players are left alone in this solution, namely P6; P11 and
P14. These do not form any coalition among themselves, because
they do not use a same item, from the 19 items in this instance.
On the other hand, these players could have formed coalitions with
other players, but these other players are better forming coalitions
with some others.
Table 5 shows the results of imposing an upper bound on the
maximum number of players that can form a coalition.
Allowing coalitions of up to two players leads to savings which
represent 91% of the savings achieved when allowing coalitions of
up to four players. Note here that the optimal coalition structure
obtained through model M1 is not strongly stable. In fact, model
M2 is infeasible. As can be seen in Fig. 4, the optimal coalition
structure of model M1 when m 2 consists of seven two-player
coalitions, and players P5 and P6 stay alone (note these two players
cannot collaborate between themselves, because they have no
common item). We have veried that such structure with the allocation vector found for model M1, violates the strong stability in
ve cases. The individual cost minimization model of Section (4.1.1) can be run for each player j, requiring the strong stability constraints only for the player who joins j. Similarly as in the
forestry case, the structure and allocations so obtained may
strongly depend on j.
Players P6 and P11 are also left alone in the solution for the
instance where m 3 and m 4, while player P14 is left alone
when m 4. These three players do not have any item in common,
and that is why they do not form any coalition among themselves.
Players who could collaborate with them are better joining other
coalitions.
The number of common items and their value is a main driver
for the players to conform a coalition. Another driver is given by
average demand values which complement each other, in the

Table 5
Optimal objective values, relative savings and cumulative savings in the problem on
inventory of spare parts for oil operations.
m

fm

f 
m

1
2
3
4

561,083
395,250
384,004
378,810

561,083
Infeasible
384,004
378,810


jfm1
fm j
%
fm

jfm f1 j
%
jf8 f1 j

29.56
2.85
1.35

91
97
100

Although our instances have been solved by the straight use of a


solver, we would like to conclude this section with a brief discussion on the applicability of the models in larger scale instances.
Note the dimension of the models increases fast with the number
of players n. There exist 2n  1 coalitions and n  2n  1 tuples
(j; k). The number of coalitions equals the number of binary variables x and the number of tuples (j; k) equals the number of nonnegative variables u. The number of constraints (4) or, equiva such that b 1, is given by
lently, the number of tuples (k; k)
k;k
 


Pn1 n Pnk n  k
n
n1
1. In consequence, the num3 2
k1
i1
k
i
ber of constraints and the number of variables of the models
reaches the order 106 when n 13 and n 20, respectively. In
order to explore the solvability, we have built a couple of random
instances (available from the authors upon request) following the
procedure described below.
First, we dene a basis cost C 100; 000. The stand-alone cost
of player i is calculated as C  1  hi , where hi are i.i.d continuous
random variable uniformly distributed in 0:01; 0:02 .
In the rst set of instances, we dene the cost of coalition Z as

C Z jZj  C 

1  hi :

i2Z

In the second set of instances, we dene the cost of coalition Z as



Y
jZjx
 C  1  hi ;
C Z jZj  1 a
jZj 1
i2Z
where a is a deector parameter (we use a 0:1) and x is a binary
random variable equally likely in f0; 1g.
Note the difference between these two sets of instances is given
by the penalty term involving the deector parameter a and a term
that, subject to x 1, increases with the number of players. This
attempts to consider that the implementation of larger coalitions
is likely to involve larger costs.
We set a time limit of 1 hour in each run of the models. For both
data set instances, we have veried in our computational experiments that the models turn to be directly solved at least for a number of players up to n 17. An example of how the solution time
varies with the cardinality bound is shown in Fig. 5. This gure
shows the solution times of both models for the second set of
instances with n 12. It can be observed that model M1 runs
quickly for all values of m. The instances that take longer times
to be solved are those corresponding to model M2 where the upper
bound on the maximum number of players per coalition lies
between m 6 and m 8. For all other values of m, model M2 runs
relatively fast, in no more than three minutes. We explain this relation between m and the time solutions by the combinatorial structure of the problem and the tendency of the optimal structure to
include coalitions of cardinality m or very close to it. Note when
m n, there is only one structure of cardinality m, obviously the
grand coalition. If the grand coalition is the optimal structure, we

M. Guajardo, M. Rnnqvist / European Journal of Operational Research 240 (2015) 147159

157

Fig. 5. Solving times of models M1 (triangular markers) and M2 (circular markers) of the second set of instances for n 12 players.

interpret it is evident for the solver to nd this out quickly. If the


grand coalition is not the optimal structure, the optimal is likely
to be found in a structure containing one coalition of cardinality
close to it, because the savings usually tend to increase with the
number of players. From combinatorics, it is known that the number of coalition structures that can be formed from a set of n players is given by the Bell number, recursively dened as B0 1 and
 
P
n
Bn1 nk0
Bk (Rota, 1964). When n 12 players, the number
k
of coalition structures is Bn 4; 213; 597. Note only one of these
structures contains a set with 12 players (namely, the grand coalition) and only 12 of them contain a set with 11 players. On the
other extreme, when m 1, there also exists only one coalition
structure, obviously the one conformed by all players standing
alone. When m is about the half of the total number of players,
then there are more coalition structures of maximum allowed cardinality and it might become trickier for the solver to nd the optimal solution, and we interpret that is why the solution time
presents the shape in Fig. 5.
When n 13, model M1 was also solved directly, within only
90 seconds. However, the run of model M2 formulated in Section 3.2 exceeded the time limit of one hour without providing
any feasible solution. We noted the problem was caused by conP
straints (8) involving the sum k2K
 , which appears to be a quite
 uj;k
time consuming computation when formulated explicitly in that
form. We introduced a new continuous variable v j dened for all
P
j 2 N, and a constraint v j k2K
 8j 2 N. Then we expressed
 uj;k
constraints (8) in terms of v j instead of the sum. This simple re-formulation provided us with a quick time solution of only 25 seconds. And it also provides us with a solution for the instances
with a number of players up to 17. For n P 18, the time limit
was exceeded without any feasible solution. Moreover, we observe
the generation of the model itself is very time consuming, because
of the large number of parameters. Since many of the aj;k and bk;k

parameters have value zero, it is possible to incorporate sets with
 k instead, but we veried this modthe relevant pairs j; k and k;
ication did not conduce to a solution within the limit time for
instances with n P 18.
For a given number of players n, in order to solve the instances
of model M1 for all values of m, it might be useful to start by solving the instance with the lowest relevant value of m (i.e., m 2). It
is easy to note that the solution to the instance where the bound is
set at m must be feasible in the instance where the bound is set at
m 1. Then the run of an instance with the bound set at m 1 may
use the optimal solution of the instance where the bound was set
at m as initial solution. Similar procedures have proved to be effective in other problems in need of solving several similar instances,
such as the multiple scenarios solved in a combinatorial auction by
Cataln, Epstein, Guajardo, Yung, and Martnez (2009). Also,

heuristics based on initial solutions have proved to be effective in set


partitioning problems, for example, in the ship scheduling problem
studied by Brnmo, Christiansen, Fagerholt, and Nygreen (2007).
As for solving model M2, because of Proposition 1, it is enough
to verify whether a cost allocation vector exists such that the optimal structure obtained when running model M1 satises the
strong stability condition. This verication can be done without
the use of binary variables, since the coalition structure is already
xed by the solution obtained when running model M1. Instead,
running a purely linear programming formulation can complete
the verication. This linear programming formulation consists only
of variables v j P 0, which refers to the cost allocated to player j.
These variables must satisfy efciency and strong stability constraints. The efciency constraints can be written as
P
j2N aj;k  v j C k for all coalition k contained in the optimal structure to model M1. The strong stability constraints can be written
P
as
j2N aj;k  v j 6 C k for all coalition k. Alternatively, one can
attempt to rst solve model M2 and, if an optimal solution is
found, Proposition 1 assures this solution also solves model M1.
Further testing of solution methods falls beyond the scope of
this article. Some literature has dealt with the combinatorial issues
in cooperative games where the characteristic function is solution
to a problem with a particular structure. These include assignment
games and travelling salesman games, among others. We refer the
reader to Curiel (2008) for a detailed overview on these types of
games. Another stream of literature has introduced some concepts
from network problems in restricted cooperative games. In a
recent article, Stein (2010) gives account on the immaturity of
the literature on this stream. A couple of decades ago, Axelrod
et al. (1995) point out the identication and quantication of the
payoffs of each player in every conceivable set of alliances as a
shortcoming of the game theoretic approach. Although we would
certainly agree it is not straightforward to compute all relevant
cost parameter values, current information technologies have
made it possible to advance in this direction. And as for the solvability of the model, in this article we have provided empirical evidence this can be carried out for a relevant number of players in
two applications and some experimental instances. Recent
advances in core stability problems based on mathematical programming have proved to be effective in large-scale problems, as
reported by Drechsel and Kimms (2010), who propose a procedure
that allows them to compute core elements in games with up to
150 players. Other recent advances in set partitioning problems
with side-constraints by Delorme and Soumis (2013) also support
the possibilities of solving large-scale problems of similar characteristics. Overall, because of the combinatorial features of the problem we have studied in this article and its relevance in theory and
practice, we believe our work opens a rich source of challenges
where OR solution methods can play an important role.

158

M. Guajardo, M. Rnnqvist / European Journal of Operational Research 240 (2015) 147159

5. Conclusions
Motivated by the context of collaborative logistics, we have formulated mixed integer linear programming models for coalition
structure and cost allocation. The models consist of a set partitioning problem together with stability constraints. These stability
constraints come from game theoretic principles, and consider
both core stability and strong stability. We have tested the models
in two applications; one in collaborative forest transportation and
one in inventory pooling of spare parts for oil and gas operations.
By running the models, we were able to determine coalition structures that satisfy the stability constraints at minimum cost, or, in
the strong stability problem, to determine whether the problem
has or not a feasible solution. If the latter is feasible, we have
proved that a same structure solves both the core stability and
the strong stability problems. The consideration or not of the
strong stability conditions can have important effects. In some
instances of the cases we have studied, it was veried that while
the core stability problem is solved, the strong stability problem
turned to be infeasible. This certainly can have practical implications, thus highlighting the notion of strong equilibrium in addition to the traditional core stability concept.
Our approach also allowed us to study the effects of restricting
the maximum cardinality of the coalitions contained in a structure.
We reported results where most of the savings from collaboration
are achieved by allowing a certain number of players in the coalitions. Still if the grand coalition achieved more savings than the
coalition structure subject to the maximum cardinality constraint,
in practice a decision maker should analyze whether these savings
overwhelm or not the managerial complexities of forming large
coalitions. In the forestry case, for example, we found that increasing the upper bound on the cardinality of the coalitions from 7 to 8
companies provided only 0.08% of relative savings.
Although previous literature in OR has incorporated game theoretic principles into collaborative problems, most of it has paid
attention to problems assuming the grand coalition is formed.
Instead of making this assumption, we approached the cost allocation problem simultaneously with the problem of nding an optimal coalition structure. A vast source of opportunities derives from
this research. First, although the coalition structure problem has
been studied in game theory for several decades, there appear to
be no particular solution methods to deal with the core stability
and strong stability problems with general characteristic functions.
Since this is basically a set partitioning problem subject to additional constraints, a direction for further research is to formulate
solution methods that could efciently nd the optimal structures
and cost allocation vectors. This is specially challenging in largescale problems, because the number of coalitions increases exponentially with the number of players. We have reported examples
where the models can be directly solved in short time including up
to 17 players and about 131,000 coalitions. Any further dimension
remains worthy of exploration. Second, it might be interesting to
set conditions for the uniqueness of the solution to these problems,
since beforehand there is no guarantee that the optimal structure
and the cost allocations are unique. Also, since the costs have been
assumed as parameters of the models, a natural extension is to
study the case where these parameters are exposed to some type
of uncertainty or where some data are missing. We have also
assumed a fully cooperative behavior of the players. Another
extension could consider the case where they can adopt non-cooperative or partly cooperative strategies. Finally, the implementation in practice of these types of models appears to be one of the
challenges for achieving the benets of collaboration. As referred
by Chen (2009), getting all players to agree on how to share costs
and benets was identied as one of the major barriers to
collaboration in practices on logistics and commerce. We believe

our article has pointed in the correct direction to break these


barriers.

Acknowledgements
We thank Vivienne Bowery Knowles for her useful suggestions
that improved the writing of this article. We also thank the anonymous reviewers, whose comments helped us to improve the
article.

References
Anupindi, R., & Bassok, Y. (1999). Centralization of stocks: Retailers vs.
manufacturer. Management Science, 45, 178191.
Aumann, R. J. (1959). Acceptable points in general cooperative n-person games. In
A. W. Tucker & R. D. Luce (Eds.). Contributions to the theory of games, IV, annals of
mathematics studies (Vol. 40, pp. 287324). Princeton, NJ: Princeton University
Press.
Aumann, R. J., & Dreze, J. H. (1974). Cooperative games with coalition structures.
International Journal of Game Theory, 3, 217237.
Axelrod, R., Mitchell, W., Thomas, R. E., Bennett, D. S., & Bruderer, E. (1995).
Coalition formation in standard-setting alliances. Management Science, 41,
14931508.
Babel, L., Kellerer, H., & Kotov, V. (1998). The k-partitioning problem. Mathematical
Methods of Operations Research, 47, 5982.
Brnmo, G., Christiansen, M., Fagerholt, K., & Nygreen, B. (2007). A multi-start local
search heuristic for ship schedulingA computational study. Computers &
Operations Research, 34, 900917.
Carlsson, D., & Rnnqvist, M. (2007). Backhauling in forest transportation: Models,
methods, and practical usage. Canadian Journal of Forest Research, 37,
26122623.
Cataln, J., Epstein, R., Guajardo, M., Yung, D., & Martnez, C. (2009). Solving multiple
scenarios in a combinatorial auction. Computers & Operations Research, 36,
27522758.
Chen, X. (2009). Inventory centralization games with price-dependent demand and
quantity discount. Operations Research, 57, 13941406.
Chwe, M. S. (1994). Farsighted coalitional stability. Journal of Economic Theory, 63,
299325.
Curiel, I. (2008). Cooperative combinatorial games. In A. Chinchuluun, P. M.
Pardalos, A. Migdalas, & L. Pitsoulis (Eds.). Pareto optimality, game theory and
equilibria (Vol. 17, pp. 131157). Dordrecht: Springer.
DAmours, S., & Rnnqvist, M. (2010). Issues in collaborative logistics. In E. Bjrndal,
M. Bjrndal, P. M. Pardalos, & M. Rnnqvist (Eds.), Energy, natural resources and
environmental economics (pp. 115130). Berlin Heidelberg: Springer-Verlag.
Delorme, M., & Soumis, F. 2013. Integral simplex and additional constraints. In
Optimization days conference. Montreal.
Drechsel, J., & Kimms, A. (2010). Computing core allocations in cooperative games
with an application to cooperative procurement. International Journal of
Production Economics, 128, 310321.
Fiestras-Janeiro, M. G., Garca-Jurado, I., Meca, A., & Mosquera, M. A. (2011).
Cooperative game theory and inventory management. European Journal of
Operational Research, 210, 459466.
Forsberg, M., Frisk, M., & Rnnqvist, M. (2005). FlowoptA decision support tool for
strategic and tactical transportation planning in forestry. International Journal of
Forest Engineering, 16, 101114.
Frisk, M., Gthe-Lundgren, M., Jrnsten, K., & Rnnqvist, M. (2010). Cost allocation
in collaborative forest transportation. European Journal of Operational Research,
205, 448458.
Gillies, D. B. (1959). Solutions to general non-zero-sum games. In A. W. Tucker & R.
D. Luce (Eds.). Contributions to the theory of games, IV. Annals of mathematics
studies (Vol. 40, pp. 4785). Princeton, NJ: Princeton University Press.
Granot, D., & Sosic, G. (2003). A three-stage model for a decentralized distribution
system of retailers. Operations Research, 51, 771784.
Greenberg, J. (1994). Coalition structures. In R. J. Aumann & S. Hart (Eds.). Handbook
of game theory with economic applications (Vol. 2, pp. 13051337). Elsevier
Science B.V.
Guajardo, M., & Rnnqvist, M. (2014). Cost allocation in inventory pools of spare
parts with service-differentiated demand classes. Article 4 in Joint decision
making and cost sharing: Theory and applications in natural resource supply
chains, PhD thesis, Department of Business and Management Science, NHH
Norwegian School of Economics.
Guajardo, M., Rnnqvist, M., Halvorsen, A. M., & Kallevik, S. I. (2014). Inventory
management of spare parts in an energy company. Accepted for publication in
Journal of the Operational Research Society. http://dx.doi.org/10.1057/
jors.2014.8.
Hadley, G., & Whitin, T. M. (1963). Analysis of inventory systems. Englewood Cliffs,
NJ: Prentice-Hall.
Hart, S., & Kurz, M. (1983). Endogenous formation of coalitions. Econometrica, 51,
10471064.
Hu, X., Caldentey, R., & Vulcano, G. (2013). Revenue sharing in airline alliances.
Management Science, 59, 11771195.

M. Guajardo, M. Rnnqvist / European Journal of Operational Research 240 (2015) 147159


Kaneko, M., & Wooders, M. H. (2004). Utility theories in cooperative games. In S.
Barbera, P. J. Hammond, & C. Seidl (Eds.), Handbook of utility theory. Extensions,
Vol. 2 (pp. 10651098). Dordrecht: Kluwer.
Kellerer, H., & Kotov, V. (2011). A 3/2-approximation algorithm for ki -partitioning.
Operations Research Letters, 39, 359362.
Kim, T., & Jeon, Y. (2009). Stationary perfect equilibria of an n-person
noncooperative bargaining game and cooperative solution concepts. European
Journal of Operational Research, 194, 922932.
Leng, M., & Parlar, M. (2009). Allocation of cost savings in a three-level supply chain
with demand information sharing: A cooperativegame approach. Operations
Research, 57, 200213.
Lozano, S., Moreno, P., Adenso-Daz, B., & Algaba, E. (2013). Cooperative game
theory approach to allocating benets of horizontal cooperation. European
Journal of Operational Research, 229, 444452.
Meca-Martnez, A., Snchez-Soriano, J., Garca-Jurando, I., & Tijs, S. (1998). Strong
equilibria in claim games corresponding to convex games. International Journal
of Game Theory, 27, 211217.

159

Nagarajan, M., & Bassok, Y. (2008). A bargaining framework in supply chains: The
assembly problem. Management Science, 54, 14821496.
Nagarajan, M., & Sosic, G. (2009). Coalition stability in assembly models. Operations
Research, 57, 131145.
zen, U., Fransoo, J., Norde, H., & Slikker, M. (2008). Cooperation between multiple
newsvendors with warehouses. Manufacturing & Service Operations
Management, 10, 311324.
Rota, G. (1964). The number of partitions of a set. The American Mathematical
Monthly, 71, 498504.
Stein, H. D. (2010). Allocation rules with outside option in cooperation games with
time-inconsistency. Journal of Business and Economics, 11, 5696.
Sosic, G. (2006). Transshipment of inventories among retailers: Myopic vs.
farsighted stability. Management Science, 52, 14931508.
Sosic, G. (2010). Stability of information-sharing alliances in a three-level supply
chain. Naval Research Logistics, 57, 279295.

Potrebbero piacerti anche