Professional Documents
Culture Documents
1 s2.0 S030505481730028X Main
1 s2.0 S030505481730028X Main
a r t i c l e i n f o a b s t r a c t
Article history: In this paper, we propose a refinement of the polynomial method based on sampling theory proposed
Received 5 April 2016 by Castro et al. (2009) to estimate the Shapley value for cooperative games. In addition to analyzing
Revised 26 January 2017
the variance of the previously proposed estimation method, we employ stratified random sampling with
Accepted 29 January 2017
optimum allocation in order to reduce the variance. We examine some desirable statistical features of the
Available online 31 January 2017
stratified approach and provide some computational results by analyzing the gains due to stratification,
Keywords: which are around 30% on average and more than 80% in the best case.
Computer science
© 2017 Elsevier Ltd. All rights reserved.
Game theory
Operations research
Shapley value
Stratified sampling algorithm
1. Introduction exactly, some specific features of the problem and/or the game
are assumed. For instance, the Shapley value was obtained in a
One of the most important one-point solution concepts in the polynomial manner for some voting games by Bilbao et al. (20 0 0).
framework of coalitional games with side payments is the Shapley (Granot et al., 2002) developed a polynomial algorithm to obtain
value (Shapley, 1953), which proposes the allocation of resources the Shapley value of an extended tree game, where a tree network
in multiperson interactions by considering the power of the play- serving heterogeneous customers was constructed. (Castro et al.,
ers during their various opportunities for cooperation. Since the 2008) computed the Shapley value for an airport game in polyno-
pioneering work by Shapley, the Shapley value has been studied mial time by considering that this value was obtained using the se-
widely from a theoretical viewpoint. Many studies have also fo- rial cost sharing rule. In addition, Deng and Papadimitriou (1994),
cused on the potential applications of the Shapley value to specific Bolus (2011), Chalkiadakis et al. (2012), and Conitzer and Sand-
problems (e.g., see Lucchetti et al., 2010; Moretti, 2010). However, holm (2006) developed efficient strategies for computing the Shap-
since computing the Shapley value is an NP-complete problem (see ley value by finding alternative efficient representations for transfer-
Deng and Papadimitriou, 1994; Fernández et al., 2002, or Faigle able utility (TU) games. However, although these methods aim to
and Kern, 1992, for more details), the problem of its calculation serve as general tools, their applications are constrained to some
must be addressed before it can employed as a useful tool in real specific classes of games, particularly weighted majority games and
situations. combinatorial optimization games. Moreover, all of these methods
Two main approaches have been employed to address the cal- require a specific representation of the game.
culation of the Shapley value: developing efficient strategies to By contrast, few studies have focused on the approximation of
compute the Shapley value exactly. the Shapley value. Considering the widespread application of game
Different approaches have been considered for the exact calcu- theory to real-world problems where exact solutions are often not
lation of the Shapley value. In order to compute the Shapley value possible, then it is necessary to develop algorithms that may facil-
itate this approximation.
∗
Corresponding author. (i) To the best of our knowledge (also see Maleki et al.,
E-mail addresses: jcastroc@estad.ucm.es (J. Castro), dagomez@estad.ucm.es (D. 2014), the first attempt to estimate the Shapley value for
Gómez), elisenda.molina@uc3m.es (E. Molina), jtejada@mat.ucm.es (J. Tejada). a large class of games was the algorithm proposed by
http://dx.doi.org/10.1016/j.cor.2017.01.019
0305-0548/© 2017 Elsevier Ltd. All rights reserved.
J. Castro et al. / Computers and Operations Research 82 (2017) 180–188 181
Castro et al. (2009), who suggested the estimation of the recall the original sampling algorithm proposed by Castro et al.
Shapley value (or any semivalue) for cooperative games (2009) upon which our refinement is based.
based on sampling. The only constraint on this algorithm be- A cooperative game in coalitional form with side payments, or
ing polynomial is the fact that the worth of any coalition with TU, is an ordered pair (N, v), where N is a finite set of players
must be computed polynomially. This approach generalizes and v : 2N → R, with 2N = {S | S ⊆ N}, is a characteristic function on
the earliest approximation method proposed by Mann and N that satisfies v(∅ ) = 0. For any coalition S ⊆ N, v(S ) ∈ R is the
Shapley (1960) for estimating the power index of a weighted worth of coalition S and it represents the reward that coalition S
voting game. In fact, Mann and Shapley (1960) stated that: can obtain by itself if all its members act together. We restrict to
“These methods have potential applications beyond the par- the case of TU games in the sequel, so we refer to them simply as
ticular problem solved here.” Moreover, as proposed in the games. For brevity, throughout this paper, the cardinalities of the
present paper, Mann and Shapley also considered alterna- sets (coalitions) N and S are denoted by appropriate small letters n
tive sampling methods for reducing the variance of the basic and s, respectively. In addition, for convenience, we write the sin-
Monte Carlo sampling approach. gleton {i} as i if there is no ambiguity. When the worth v(S) repre-
(ii) If an efficient general purpose algorithm is available, then it sents the cost that players must be charged, the game is known as
will be possible to estimate the Shapley value by refining a cost-game and we denote these types of games as (N, c).
the original procedure by Castro et al. in order to exploit the Let GN be the vector space of all the games with a fixed player
specific properties of the problem considered. Thus, Maleki set N, and identify (N, v ) ∈ GN by its characteristic function v if
et al. (2014) proposed the use of stratified random sam- there is no ambiguity. A value ϕ is an allocation that associates
pling to reduce the estimated error in the original method with each game (N, v ) ∈ GN a vector ϕ (N, v ) ∈ RN , where ϕi (N, v ) ∈
described by Castro et al. They also considered the same R represents the value of player i, i ∈ N. Shapley (1953) defines his
strata treated in the present paper. However, their allocation value as follows:
method ignores the specific features of the game, thereby ( n − s − 1 )! s !
resulting in an estimation method that improves the origi- Shi (v ) = (v(S ∪ {i} ) − v(S )), i ∈ N. (1)
n!
S⊆N
nal method based on simple random sampling, but it gen- i∈S
eral performs poorly compared with the refinement that
The value Shi (v) of each player, which is a weighted average of his
we propose in the present study. The reader may refer to
marginal contributions, allows different interpretations, such as the
Section 6 for a comparative analysis using the cooperative
payoff that player i receives when the Shapley value is used to pre-
games analyzed by Castro et al. (2009).
dict the allocation of resources in multiperson interactions, or his
Few studies have developed approximations to estimate the power when averages are used to aggregate the power of players
Shapley value and most of them are problem specific. Fatima et al. in their various opportunities for cooperation.
(2006) presented a randomized polynomial method for determin- Weber (1988) gave an alternative characterization of the Shap-
ing the approximate Shapley value for weighted voting games and ley value in terms of all the possible orders of the players, which
k-majority voting games. Crama and Leruth (2007) also proposed was employed for estimation by Castro et al. (2009) . Let O: N
the estimation of the Shapley value for a complex voting game → N be a permutation that assigns the player O(k) to each po-
that reflects the structure of shareholdings in complex interlocked sition k and let π (N) denote the set of all possible permutations
shareholding structures based on a Monte Carlo simulation. of the player set N. Given a permutation O ∈ π (N), let us de-
In this paper, we propose a refinement based on stratified ran- note Prei (O) as the set of predecessors of the player i in the or-
dom sampling, which we use instead of simple random sampling der O (i.e., P rei (O ) = {O(1 ), . . . , O(k − 1 )}, if i = O(k )). In this set-
to estimate the Shapley value. As found with the original algo- ting, the vector of marginal contributions for a given order O ∈ π (N),
rithm, these refined estimations are polynomial if each marginal x(O ) = (x(O )i )i∈N is defined as:
contribution can be calculated in polynomial time. The remainder x(O )i = v(P rei (O ) ∪ i ) − v(P rei (O )), i ∈ N.
of this paper is organized as follows. In Section 2, we provide some
basic concepts and notation. We also describe the ApproShapley al- Weber (1988) showed that the Shapley value can be expressed
gorithm proposed by Castro et al. (2009), which is the basis of our as the following expectation, where it is assumed that all different
propose method. In Section 3, we analyze the sources of the vari- orders have equal probability:
ance in the estimations derived from ApproShapley and we propose 1
Shi (v ) = x(O )i , i ∈ N. (2)
three methods for reducing this variance. First, we propose a gen- n!
O∈π ( N )
eral method based on stratified random sampling St-ApproShapley,
Shst , and we then determine the optimum allocation of the sam- Thus, since the Shapley value is an expectation, Castro et al.
ple sizes in the corresponding strata St-ApproShapley-opt, Shst, opt . (2009) proposed its statistical estimation. Next, we describe the al-
We also propose a two-stage estimation method that preserves the gorithm that they proposed (emphApproShapley), which is based
efficiency of the Shapley value, St-ApproShapley-eff, Shst, eff , in the on a unique simple random sampling process for estimating the
case where efficiency is a demanding property. In Section 4, we Shapley value Shi of all players i ∈ N.
analyze the estimated error for stratified random sampling with 1. The population of the sampling process P is the set of all possi-
the optimal allocation approach, and we compare it with the es- ble orders of players, i.e., P = π (N ). Each sampling unit repre-
timation method proposed by Maleki et al., which is described in sents an order O ∈ π (N).
Section 3. In Section 5, we present the proposed estimation algo- 2. The vector parameter under paper is Sh = (Shi )i∈N .
rithm. Finally, in Section 6, we present some computational results 3. The characteristic observed in each sampling unit, O ∈ π (N), is
to show the improvements obtained using the proposed method the vector of the marginal contributions for that order O, i.e.,
for the cooperative games analyzed by Castro et al. (2009). We give x ( O ) = ( x ( O )i )i∈N .
our conclusions in Section 7. 4. The estimate of the parameter Sh, Sh ˆ , is the mean of the
marginal contributions over the sample M, i.e., Sh ˆ = (Sh
ˆ i )i∈N ,
2. Preliminaries where
In this section, we introduce some sampling notation and theo- ˆi= 1
Sh x(O )i , i ∈ N.
retical games in order to understand the rest of this paper. We also m
O∈M
182 J. Castro et al. / Computers and Operations Research 82 (2017) 180–188
5. Finally, the selection process used to determine the sample M St-ApproShapley-eff, Shst, eff , for the case where efficiency is a de-
will take any order O ∈ π (N) with a probability of n1! . They con- manding property.
sider a sample M as a subset of P × P . . . × P, i.e., a sample that The first stratified random sampling process, which we refer to
m times as St-ApproShapley, is defined as follows.
has been obtained with replacement.
1. The population in the sampling process is P and each sampling
unit represents a marginal contribution X(O)i ∈ P.
ˆ can be calculated in polyno-
Note that in order to assure that Sh
2. The parameter under paper is Sh.
mial time, it is assumed that the worth of any coalition S, v(S), can
3. The characteristic observed in each sampling unit is x(O)i .
be calculated in polynomial time. In fact, it is sufficient to assume
4. The strata considered are Pi for all i = 1, . . . , n and = 1, . . . , n,
that the vector of the marginal contributions of all orders can be
where Pi is the set of all possible marginal contributions in
obtained in polynomial time.
which player i is in position .
We select a random sample Mi of size mi , which is obtained
with replacement from each stratum Pi . Thus, the selection
3. Improving the estimation of the Shapley value: Stratified
process used to determine the sample Mi will take any ele-
random sampling approach
ment of Pi with a probability of (n−11
)! .
In some situations, the ApproShapley algorithm proposed by 5. Let Shi be the mean of the marginal contributions over the
Castro et al. (2009) might not be sufficiently precise for our pur- sample Mi , i.e.
poses. In these cases, a method must be employed to reduce the 1
variance of the estimations obtained by simple random sampling. A Shi = x ( O )i , ∀i = 1, . . . , n ∀ = 1, . . . , n.
mi
common technique for variance reduction when using Monte Carlo X (O )i ∈Mi
methods is to estimate the expectations from a known population Then, the estimate of the parameter Shi obtained by stratified
by stratification (e.g., see Cochran, 1977 or Lohr, 1999), so we pro- sampling is Shst
i
, where
pose the use of stratified random sampling instead of simple ran-
n
1
dom sampling to improve the performance of the original method. Shst
i = Sh , ∀i = 1, . . . , n.
Stratification may enhance the estimated precision for the expecta- n i
=1
tions in the whole population if it is possible to divide a heteroge- Note that a procedure for determining the sample size mi
neous population into subpopulations, each of which is internally taken from each stratum Pi must be defined in order to apply
homogeneous. In many cases, one of the sources of the variance the St-ApproShapley algorithm described above. We propose the
in each marginal contribution is the player and the order in which use of stratification with the optimum allocation of mi (Neyman,
this player arrives, so we consider these two sources of variation 1934), which means that the sampling sizes mi in the correspond-
when building our subpopulations. ing strata are selected to minimize the estimation error. The next
For example, let us consider the airport game described in proposition shows that Shst is unbiased, and thus we select the
i
Section 6, where the marginal contributions of larger planes vary sample sizes in order to minimize the variance of the estimator,
substantially depending on whether the plane arrives at the begin- V [Shst ].
i
ning or the end of the corresponding order, where they are inter-
nally homogeneous in each of these positions. This is generally the Proposition 3.1. The estimator Shst
i
is unbiased, i.e., E[Shst
i
] = Shi ,
case for games determined by a graph structure where hub players and its variance is given by:
exist that connect large connected components. In addition, in the
1 σi2 (x(O )i − Shi )2
X (O )i ∈Pi
case of voting games, the marginal contributions can vary from 0 V [Shst
i ] = 2 , where σ =
2
, (3)
to 1 depending on the position in which the voter arrives, and in
n ∈L mi i
( n − 1 )!
i
the case of extreme weights, they tend to be stable for each given and Li ⊆ {2, . . . , n − 1} is the set of the indices of the strata for player
position. i with strictly positive variability and Shi is the expected marginal
Therefore, criteria for determining the strata may comprise the contribution of player i when he occupies the th position:
players and the order in which each player arrives. In this case,
X (O )i ∈Pi x ( O )i
we define n × n strata, Pi , i ∈ N, = 1, . . . , n, where the stratum Shi = .
Pi is the set of all the possible marginal contributions in which ( n − 1 )!
player i occupies position . Let us note that the population (from Proof 3.1. The proof is straightforward given the properties of the
which we take random samples) must include all possible marginal sample mean as an estimator of the population mean. Note that
contributions, so we have n · n! possible units. This number (n · in St-ApproShapley, the samples are drawn independently from dif-
n!) is obtained because for each n marginal contributions, we have ferent strata. Moreover, since all the values x(O)i in strata Pi1 are
n! possible orders. The population with size n · n! must be di- v(i) and all the values x(O)i in strata Pin are v(N ) − v(N \ i ), then
vided into non-overlapping subpopulations, Pi for all i = 1, . . . , n σi21 = σin2 = 0 for every player i ∈ N.
and = 1, . . . , n, which comprise the whole of the population. Af-
ter the strata have been determined, a random sample is taken According to expression (3), the sum of variance of all play-
n st
from each stratum. ers, i=1 V [Shi ], is minimized when mi is proportional to the
In this paper, we propose three methods for reducing the vari- true variance σi2 in each stratum i = 1, . . . , n and = 1, . . . , n. Let
ance in the original estimation method (Castro et al., 2009). First, Shst, opt be the estimator of the Shapley value obtained by stratified
we propose a general method based on stratified random sampling random sampling with the optimum allocation, i.e., with mi ∝ σi2 .
(which we refer to as St-ApproShapley and its solution is Shst ). Next, In this case, m → ∞ implies that mi → ∞ for every stratum Pi
we determine the optimum allocation of the sample sizes in the with σi2 > 0. Therefore, the consistency of Shst, opt can be estab-
corresponding strata (we refer to this method as St-ApproShapley- lished as a corollary of Proposition 3.1.
opt and we denote its solution as Shst, opt ). We also describe the
estimation method proposed by Maleki et al. (2014) as a particu- Corollary 3.1. The estimator Shst, opt (m) is consistent in probability,
i.e.,
lar case of St-ApproShapley. Finally, we present a two-stage estima-
tion method, which preserves the efficiency of the Shapley value, lim P |Shst,opt
i
(m ) − Shi | > = 0, ∀ > 0 ∀ i = 1, . . . n, (4)
m→∞
J. Castro et al. / Computers and Operations Research 82 (2017) 180–188 183
where Shst, opt (m) denotes the estimator of Sh obtained by stratified n strata Si0 , Si1 , . . . , Sin−1 defined as Sik = {C ⊆ N \ {i}, |C | = k}. In par-
random sampling with the optimum allocation based on a total sam- ticular, they proposed estimating the average marginal contribu-
ple M = ∪ni=1 ∪n=1 Mi of size m = ni=1 n=1 mi . tion of player i in each stratum:
1
In the case where efficiency is a desirable property, we propose φik = (v(C ∪ {i} ) − v(C )),
an alternative estimation method referred to as St-ApproShapley-eff, |Sik | C∈Sk
i
which preserves the efficiency, reduces the variance in the original
simple random sample estimation approach ApproShapley, and re- k = 0, 1, . . . , n − 1, i = 1, . . . , n,
mains an unbiased estimator of Sh. We propose a two-stage esti- by simple random sampling, and then obtained the final estimation
mation approach by combining the original unbiased and efficient of Shi as the average
ApproShapley approach with the St-ApproShapley with optimum al-
n−1
location approach. 1 ˆk
Let m be the total sample size, and let msim and mst be such
φˆ i = φi
n
k=0
that m = msim + mst . Then, the propose two-stage efficient strati-
fied method is defined as follows. They proposed the calculation of the sample size in each stratum,
as follows. Let
• Stage 1. Efficient estimation based on a simple random sam-
mi ( k + 1 ) 3
2
ˆ.
original ApproShapley method to obtain the estimator Sh Then, allocate mki = m∗k i
samples to stratum Sik , k = 0, . . . , n − 1,
• Stage 2. Perform a St-ApproShapley with optimum allocation us-
st,opt and sequentially increase the value of mki by 1 from k = 0 to (n −
ing a total sample size of mst . Let Shst,opt = (Shi )i∈N be the
1 ) until it exceeds mi , where mi is the total sample size allocated
estimator obtained.
to the estimation of Shi .1 . The reader is referred to Remark 4.2 for
• Estimator. The final estimate of the parameter Sh, Shst, eff , that
the motivation of m∗ki
.
we propose is given by:
Now, each stratum Sik is equivalent to Pik+1 , so the method pro-
n Sˆh posed by Maleki et al. is a particular case of St-ApproShapley with
( ) , ∀ i ∈ N.
i
Shst,e ff
= Shst,opt + v N − Sh st,opt
(5) allocation
i i
i=1
i v ( N)
m · 3
2
for every marginal contribution in P31 ∪ P32 and x(O )3 = 1 for every Proposition 4.3. Let ei > 0 be the maximum estimation error for
marginal contribution in P33 ∪ P34 . player i’s Shapley value Shi . Then,
In the sequel, we will consider that the conditions established
in Proposition 4.1 do not hold and that we use an optimal alloca-
2e2i
P | Shst,opt − Shi | ≤ ei ≥ 1 − 2exp − . (11)
i n−1 (xmax −xmin )2
i i
tion, i.e.,
=2 n2 m i
σ2
mik = n ik m, ∀ k = 1, . . . , n, ∀ i = 1, . . . , n. (8) Proof 4.3. Note that
∈L j σ j
2
j=1
n x ( O )i
st,opt Shst,opt = .
Thus, assuming that for all i = 1, . . . , n is normally dis-
Shi i n · mi
=1 x(O )i ∈Mi
tributed, which holds for large samples according to the central
limit theorem, then we conclude with the following result regard- Thus, since Mi1 , . . . , Min are random samples, then each of its el-
ing the precision of the estimate. ement is bounded by the two values xi min
and xi
max , so inequality
(11) holds (see Hoeffding, 1963, Theorem 2).
Proposition 4.2. Let 1 − α be a given confidence level and e > 0 is
the maximum estimation error for player i’s Shapley value Shi for all Remark 4.2. Maleki et al. (2014) used the estimation error (11) to
i = 1, . . . , n. If the sampling size m verifies: determine the sample sizes mi , ı = 1, . . . , n, = 1, . . . , n, allocated
n to each stratum in order to minimize the estimation error within
zα2 /2 max j=1,...,n |L j | j=1 ∈L j σ j2
m≥ , (9) each of them. In particular, they first derived an upper bound for
n2 · e2 the marginal contributions range xi i
max − xmin , for each = 1, . . . , n,
where zα /2 is the value such that P {Z ≥ zα /2 } = α2 , and Z is the stan- and they then determined a global upper bound for the estimation
error of Shi . They set the values m∗k
dard normal distribution, then P |Shi − Shi | ≤ e ≥ 1 − α for all in (6) to minimize the global
st,opt
i
i = 1, . . . , n. upper bound.
The main problem with their allocation m∗k i
is that it does not
Proof 4.2. The proof is a straightforward application of the central depend on either the specific player i ∈ N or the special features of
limit theorem, Proposition 3.1, and expression (8) above. the game. Moreover, it increases with position k by ignoring that
In general, the true variances σ j 2 , = 2, . . . , n − 1 are not σin2 = 0 for every player and for every game. The root cause of this
problem is the upper bound with which they work because it is
known. Thus, the variances must be estimated in order to deter-
excessively coarse. For example, if we consider the airport game
mine the minimum sample size, m, required to guarantee a given
(N, c) described in Section 6 (page 6), the range of the marginal
error with a certain probability.2 However, following the same ar-
contributions within each stratum are obtained as follows. Let F0 =
gument stated by Castro et al. (2009), we can give the upper
j j 0 and Fk = ki=1 fi , where fi is the number of planes for which the
bounds on the estimation error e. Let xmax and xmin denote the
cost is ci = i (i = 1, . . . , 10 in this example). Then, since fi ≥ 2 for all
greatest and least marginal contribution of player j in Pj , respec-
i, then it holds that xi min
= 0 for all ≥ 2 and for all i ∈ {1, . . . , 10}.
tively. Then, the following inequality holds:
With respect to the maximal marginal contributions, the following
(xmax
j j
− xmin )2 relations hold for every 2 ≤ i ≤ 10:
σ j2 ≤ , ∀ = 2, . . . , n − 1, ∀ j = 1, . . . , n.
4 xi
max = i − k, for all Fk−1 + 2 ≤ ≤ Fk + 1,
Therefore, when the confidence level 1 − α is given and assuming for all k = 1, . . . , i − 1. (12)
the optimal allocation, the estimation error can be bounded by
Thus, for every fixed i ∈ {2, . . . , 10}, the range of the marginal con-
n tributions within each stratum decreases with , so a good alloca-
zα /2 max j=1,...,n |L j | j=1 ∈L j (xmax − xmin )
j j 2
e≤ √ . (10) tion of the mi samples should prescribe smaller sample sizes for
2n m the latest positions. In Section 6, we use this game, among others,
Remark 4.1. If we need to reduce the estimation error, we can in order to measure the improvement obtained due to the optimal
−1 allocation of the m samples compared with that obtained using the
also compare the amounts mi and n−1 , which is an upper bound
of the cardinality of the different marginal contributions in Pi . allocation proposed by Maleki et al.
−1
Let Ki := { ∈ {1, . . . , n} | mi < n−1 }. Thus, for all i = 1, . . . , n and 5. Algorithm description
∈ Ki , we can exactly calculate
( − 1 )! ( n − )! As mentioned earlier, the true variances σi2 must be estimated
Shi = (v(S ∪ i ) − v(S )). to determine the optimum sample sizes for each stratum Pi . In
S⊆N\i
( n − 1 )!
|S|=−1 this section, we describe our proposed method for use in practice,
which is a two-stage approach that divides the (total) sample size
In this case, the bound above (10) can be restated as follows: m into two parts, m = mexp + mst . Then, mexp samples are used to
obtain unbiased estimates of the true variances σi2 , and we also
zα /2 max j=1,...,n (n − k j ) nj=1 ∈K j (xmax
j j
− xmin )2
employ this sample to estimate the marginal contributions of all
e≤ n ,
2n m − j=1 m(K j ) players in all positions (the strata in the sampling model) using the
n−1 same sample size. Next, the remaining mst are used in the strata
where k j = |K j | and m(K j ) = ∈K j −1 . where the optimal sample gives a greater value than that assigned
in the previous step.
All the previous bounds are based on the central limit theo- Obviously, the selection of mexp and mst should also be decided
rem. Following Maleki et al. (2014), we can derive an alternative according to the characteristics of the current problem. However,
non-asymptotic bound, which is based on Hoeffding’s inequality we consider that the effort required to select these sizes is gen-
(Hoeffding, 1963). erally not worthwhile. Thus, we employ a balanced percentage of
50% in each part of the analysis to obtain the computational results
2
These estimations are also necessary for determining the optimal allocations. In in Section 6. Moreover, other features of the two-stage method
Section 5, we present the estimation process that we apply in practice. must be set, as follows.
J. Castro et al. / Computers and Operations Research 82 (2017) 180–188 185
• The sampling size in Stage 1 is the same in all strata Pi , i.e.,
exp
mi = 2m n2
.
• The total sampling size in each stratum is proportional to the
estimated variance s2i . The sampling size mst i
for each stratum
Pi , = 1, . . . , n, in the stratified paper is calculated as follows:
s2i
mi = m n n 2
,
i=1 k=1 sik
exp
mst
i = mi − mi .
Note that the sample sizes for each stratum are calculated over
the total sample size m because we use the information about
the marginal contributions x(O)i obtained from the pilot study
to give the estimation of Shi . Thus, in the second stage, only
exp
the remaining mi − mi samples are selected randomly. In the
st
case where mi ≤ 0, no more samples are generated for the
0
stratum Pi 0 and the sample size that we have not yet used
(−msti
) is allocated among the remaining strata in a propor-
0
exp
tional manner, mi − mi . We iterate this process until all mst
i
≥
0.
∀ = 1, . . . , n, ∀i = 1, . . . , n. 6. Computational results
• Stage 2. For each position ∈ {1, . . . , n} and for each player i ∈ We used the same examples employed by Castro et al.
{1, . . . , n}: (2009) as a benchmark in order to measure the improvement ob-
1. The population in the sampling process is Pi and each sam- tained due to stratification and to compare our allocation method
pling unit represents a marginal contribution x(O )i ∈ Pi . with that of Maleki et al. We selected five well-known games from
2. The parameter under paper is Sh. the literature where the worth can be calculated in polynomial
3. The characteristic observed in each sampling unit is x(O )i ∈ time for any coalition with known Shapley values. We found that
Pi . method proposed by Maleki et al., which is also based on stratifi-
4. We select a random sample Mi st of size mst , which is ob-
i cation, improved the performance compared with the original es-
tained with replacement from each stratum Pi . timation method described by Castro et al. (2009) based on sim-
st
5. Let Shi be the mean of the marginal contributions over the ple random sampling, but it performed poorly compared with the
st (if mst = 0 then Sh = 0), i.e., st two-stage optimal allocation method proposed in the present pa-
sample Mi i i
per. We explored the factors that contribute to the relevant gains
based on these examples, which showed that the gains increased
st 1
Shi = x ( O )i , ∀ = 1, . . . , n, ∀i = 1, . . . , n. with the dependence of the marginal contributions on the order of
mst entrance by players and when players are asymmetric.
i x(O )i ∈Mi
st
186 J. Castro et al. / Computers and Operations Research 82 (2017) 180–188
where w = {w1 , . . . , w51 } = {45, 41, 27, 26, 26, 25, 21, 17, 17, 14,
Table 2
13, 13, 12,12, 12, 11, 10, . . . , 10,9, . . . , 9, 8, 8, 7, . . . , 7, 6, . . . , 6, 5,
Errors in the non-symmetric voting game.
4 times 4 times 4 times 4 times m/n 103 104 105 106 107 108
4, . . . , 4, 3, . . . , 3}.
esim
max 8.83 4.495 0.969 0.283 0.115 0.053
9 times 7 times esim 3.104 1.073 0.321 0.082 0.03 0.014
ave
3. An airport game. Let (N, c) be an airport game where N = sim
eth 31 9.8 3.1 0.98 0.31 0.098
{1, . . . , 100}. The characteristic function in this game, v, is de- est
max 8.468 2.301 0.737 0.183 0.0822 0.0239
fined as follows, ∀ S ⊆ N, 4.10% 48.80% 23.91% 35.42% 28.53% 54.87%
est
ave 2.608 0.836 0.223 0.0613 0.0208 0.00667
c (S ) = max ci , 15.97% 22.12% 30.57% 25.29% 30.6% 52.39%
i∈S st
eth 19.777 6.254 1.978 0.625 0.198 0.0625
36.19% 36.19 % 36.19% 36.19% 36.19% 36.19%
where c = {c1 , . . . , c100 } = {1, . . . , 1, 2, . . . , 2, 3, . . . , 3, 4, . . . , 4,
eMa
max 13.572 2.814 1.315 0.439 0.110 0.0553
8 times 12 times 6 times 14 times −34.94% 37.40% −26.31% −35.54% 4.35% −4.16%
5, . . . , 5, 6, . . . , 6, 7, . . . , 7, 8, . . . , 8, 9, . . . , 9, 10, . . . , 10}. eMa 2.964 0.938 0.338 0.117 0.0286 0.0118
ave
4.51% 12.58% 5.03% −29.91% 4.67% 15.71%
8 times 9 times 13 times 10 times 10 times 10 times
4. A shoes game. Let (N, v) be a shoes game where N =
{1, . . . , 100}. The characteristic function in this game is defined
as follows, ∀ S ⊆ N, Table 3
Errors in the airport game.
v(S ) = min{|Sle f t |, |Sright |}, m/n 103 104 105 106 107 108
where |Sleft | is the number of left shoes and |Sright | the num- esim
max 66.63 37.26 6.21 1.91 0.64 0.24
ber of right shoes. In this example, |Sle f t | = |Sright | = 50. In the esim
ave 14.41 5.35 1.21 0.48 0.15 0.05
sim
sampling process, without loss of generality, it is assumed that eth 310 98 31 9.8 3.1 0.98
est
max 47.370 3.450 0.920 0.430 0.0995 0.0353
the first 50 players correspond to the left shoes and the other 28.91% 90.74% 85.18% 77.48% 84.46% 85.30%
50 to the right shoes. It is easy to see that the Shapley value of est
ave 5.437 0.646 0.190 0.0570 0.0201 0.00617
1
this game is Shi = ∀i ∈ N. st
62.27% 87.93% 84.30% 88.13% 86.63% 87.77%
2 eth 83.041 26.260 8.304 2.626 0.831 0.262
5. A minimum spanning tree game. Let G = (N ∪ {0}, E ) be a val- 73.21% 73.21% 73.21% 73.21% 73.21% 73.21%
ued graph where N = {1, . . . , 100} and the cost associated with eMa
max 65.630 30.414 7.753 2.260 1.052 0.321
1.50% 18.37% −19.90% −15.49% −39.16% −25.23%
an edge (i, j) is
eMa
ave 11.340 4.458 1.315 0.493 0.147 0.0521
Ci j := 21.30% 16.67% −7.98% −2.64% 2.00% −4.03%
1, if i = j + 1, i = j − 1, i = 1 ∧ j = 100, i=100 ∧ j=1,
101, if i = 0 or j = 0,
Table 4
∞, otherwise.
Errors in the shoes game.
A minimum spanning tree game (N, c) is a cost game, m/n 103 104 105 106 107 108
where for a given S ⊆ N, c(S) is the sum of the
esim 46 12.85 2.85 1.265 0.355 0.1255
edge cost of the minimum spanning tree, i.e., c (S ) = max
esim 10.95 3.85 0.95 0.31 0.115 0.035
ave
minimum spanning tree of the graph G|S∪{0} , which is the par- sim
eth 310 98 31 9.8 3.1 0.98
tial graph restricted to the players in coalition S and the source est
max 32.277 12.061 3.653 1.266 0.336 0.103
node 0. It is easy to see that the Shapley value for this game is 29.83% 5.92% −21.98% −0.08% 5.35% 18.19%
est
Shi = 2, ∀i ∈ N. ave 10.632 4.237 1.218 0.380 0.100 0.0356
2.90% −9.13% −22.00% 18.42% 13.04% −1.44%
st
eth 30.371 9.604 3.037 0.960 0.304 0.0960
We compared the estimations obtained by stratified random 2.00% 2.00% 2.00% 2.00% 2.00% 2.00%
sampling with the optimum allocation, the results obtained us- eMa
max 44.612 18.812 4.525 1.620 0.121 0.549
ing the original algorithm based on simple random sampling, and 3.02% −31.69% −37.02% −21.91% −35.34% 3.59%
those obtained by employing the allocation method of Maleki eMa
ave 12.777 4.813 1.505 0.546 0.155 0.0390
−14.30% −20.01% −36.88% −43.22% −25.81% −10.26%
et al. Tables 1–5 show the maximum error, average error, and theo-
retical error bound for these estimates, those based on simple ran-
J. Castro et al. / Computers and Operations Research 82 (2017) 180–188 187
the worst performance in this case. For the last game, as men- player’s arrival position (which does not necessarily increase with
tioned earlier, the method of Maleki et al. performed in a simi- the size of coalitions, as found with the algorithm proposed by
lar manner to our method, although both obtained improvements Maleki et al.). In situations where the symmetric algorithms or that
compared with the original method based on simple random sam- proposed by Maleki et al. perform very well, our algorithm does
pling (see Table 5). not obtain significantly worse results because we make no assump-
tions regarding the variability of the marginal contributions.
7. Conclusions In this paper, we focused on the Shapley value. It is important
to note that the method used to estimate the Shapley value can be
Cooperative game theory is a very important topic and it has extended in a very similar manner to estimate any semivalue.
applications in many disciplines, such as economics, politics, bi- Finally, in conclusion, we not that the proposed procedure can
ology, sociology, operations research, and social network analysis. be computed in a parallel manner because the random estima-
One of the main topics in cooperative game theory is the allocation tion process in each stratum does not require information from
of the worth of the grand coalition among the players involved in the other strata. It would not be difficult to obtain a pseudo-code
the game, where the Shapley value is the solution concept that is that works in parallel in order to reduce the complexity in a linear
used most often for this purpose. In this sense, the Shapley value is manner.
useful when it is necessary to allocate the value that a set of play-
ers can achieve if they agree to cooperate. Nevertheless, calculating Acknowledgement
the Shapley value has been an open topic in game theory, where
the complexity of its calculation is generally exponential, although This paper was partially supported by the Plan Nacional de I+D+i
its calculus has been solved well in some particular cases. This is of the Spanish Government under projects MTM2011-27892 and
one of the reasons why the Shapley value has not been applied MTM2015-70550.
extensively to real and large-scale problems.
As mentioned in the Introduction, the problem of computing References
Shapley value’s has been studied from two alternative perspectives
Bilbao, JM, Fernández, JR, Jiménez Losada, A, López, JJ, 20 0 0. Generating functions
in recent years: algorithms to compute the Shapley value exactly for computing power indices efficiently. TOP 8 (2), 191–213.
(e.g., see Bilbao et al., 20 0 0; Granot et al., 2002) and algorithms Bolus, S, 2011. Power indices of simple games and vector-weighted majority games
that focus on its approximation, where the main aim is to “esti- by means of binary decision diagrams. Eur. J. Oper. Res. 210 (2), 258–272.
Castro, J, Gómez, D, Tejada, J, 2008. A polynomial rule for the problem of sharing
mate” the Shapley value within a reasonable computational time. delay costs in PERT networks. Comput. Oper. Res. 35 (7), 2376–2387.
The second approach has been investigated significantly less often Castro, J, Gómez, D, Tejada, J, 2009. Polynomial calculation of the Shapley value
than the first. However, considering the widespread application of based on sampling. Comput. Oper. Res. 36 (5), 1726–1730.
Chalkiadakis, G, Elkind, E, Wooldridge, M, 2012. Computational aspects of coop-
game theory in real-world problems, where exact solutions are of-
erative game theory. Synthesis Lectures on Artificial Intelligence and Machine
ten not possible, it is necessary to develop algorithms that can fa- Learning. Morgan & Claypool.
cilitate this approximation. It should be noted (as mentioned by Cochran, WG, 1977. Sampling Techniques. Wiley.
Conitzer, V, Sandholm, T, 2006. Complexity of constructing solutions in the core
Maleki et al., 2014) that the first attempt to estimate the Shap-
based on synergies among coalitions. Artif. Intel. J. 170, 607–619.
ley value for a large class of games was the algorithm proposed Crama, Y, Leruth, L, 2007. Control and voting power in corporate networks: concepts
by Castro et al. (2009). Other attempts to estimate the Shapley and computational aspects. Eur. J. Oper. Res. 178, 879–893.
value were mostly for a specific class of problems. Thus, for a real Deng, X, Papadimitriou, CH, 1994. On the complexity of cooperative solution con-
cepts. Math.Oper. Res. 19 (2), 257–266.
situation and real cooperative game, the Shapley value must be Faigle, U, Kern, W, 1992. The Shapley value for cooperative games under precedence
estimated using a sampling approach. For example, Fatima et al. constraints. Int. J. Game Theory 21 (3), 249–266.
(2006) proposed a randomized polynomial method for approxi- Fatima, SS, Wooldridge, M, Jenniggs, NR, 2006. An analysis of the Shapley value and
its uncertainty for the voting game. Lect.NotesArtif.Intell. 3937, 85–98.
mating the Shapley value of weighted voting and k-majority voting Fernández, JR, Algaba, E, Bilbao, JM, Jiménez, A, Jiménez, N, López, JJ, 2002. Gener-
games, while Crama and Leruth (2007) used a Monte Carlo simula- ating functions for computing the Myerson value. Ann Oper Res 109, 143–158.
tion to estimate the Shapley value for a complex voting game. The Granot, D, Kuipers, J, Chopra, S, 2002. Cost allocation for a tree network with het-
erogeneous customers. Math. Oper. Res. 27 (4), 647–661.
second well-known approach for estimating the Shapley value for Hoeffding, W, 1963. Probability inequalities for sums of bounded random variables.
a large class of games was proposed by Maleki et al. (2014), who J. Am. Stat. Assoc. 58 (301), 13–30.
proposed the use of stratified random sampling to reduce the error Lohr, H, 1999. Sampling: Design and Analysis. Duxbury.
Lucchetti, R, Moretti, S, Patrone, F, Radrizzani, P, 2010. The Shapley and Banzhaf
estimation in the original method described by Castro et al. They values in microarray games. Comput. Oper. Res. 37, 1406–1412.
also considered the same strata treated in the present paper. Maleki, S., Tran-Thanh, L., Hines, G., Rahwan, T., Rogers, A., 2014. Bounding the esti-
In order to compare the three algorithms for Shapley value es- mation error of sampling-based Shapley value approximation arXiv:1306.4265.
Mann, I, Shapley, LS, 1960. Values for Large Games IV: Evaluating the Electoral Col-
timation in general games, the symmetric algorithm defined by
lege by Monte Carlo Techniques. Technical report. The RAND Corporation, Santa
Castro et al. (2009) and the algorithm proposed by Maleki et al. Monica.
(2014) require (to obtain better results) that the variabilities of Moretti, S, 2010. Statistical analysis of the Shapley value for microarray games. Com-
the marginal contributions of all players are similar. Moreover, the put. Oper. Res. 37, 1413–1418.
Neyman, J, 1934. On the two different aspects of the representative method: the
symmetric algorithm requires that the marginal contributions do method of stratified sampling and the method of purposive selection. J. R. Stat.
not depend greatly on the positions in which the players arrive. Soc. 97, 558–606.
The algorithm proposed by Maleki et al. also requires that the vari- Owen, G, 1995. Game Theory. Academic press.
Shapley, LS, 1953. A value for n-person games. In: Kuhn, HW, Tucker, AW (Eds.),
ability of the marginal contributions increases with the size of the Contributions to the Theory of Games II (Annals of Mathematics Studies 28).
coalitions. Our proposed algorithm clearly obtains improvements Princeton University Press, pp. 307–317.
in situations where the variabilities in the marginal contributions Weber, RJ, 1988. Probabilistic values for games. In: Roth, A (Ed.), The Shapley Value:
Essays in Honor of Lloyd S. Shapley. Cambridge University Press, pp. 101–119.
of each player are very different, or in those situations where the
variability of the marginal contributions depends greatly on each