Accepted Manuscript

A Goal Programming approach to solve the multiple criteria DEA

model

Ana Paulados Santos Rubem ,

Joao Carlos Correia Baptista Soares de Mello , Lidia Angulo Meza

PII: S0377-2217(16)30974-2
DOI: 10.1016/j.ejor.2016.11.049
Reference: EOR 14128

To appear in: European Journal of Operational Research

Received date: 11 June 2015

Revised date: 23 November 2016
Accepted date: 28 November 2016

Please cite this article as: Ana Paulados Santos Rubem , Joao Carlos Correia Baptista Soares de Mello ,
Lidia Angulo Meza , A Goal Programming approach to solve the multiple criteria DEA model, European
Journal of Operational Research (2016), doi: 10.1016/j.ejor.2016.11.049

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

Highlights
We revisit the goal programming approach for the Li and Reeves model
We show the inconsistencies in such approach
We present a correct goal programming approach for the Li and Reeves model
We point out some differences between the two models

T
IP
CR
US
AN
M
ED
PT
CE
AC

1
 ACCEPTED MANUSCRIPT

A GOAL PROGRAMMING APPROACH TO SOLVE THE MULTIPLE

CRITERIA DEA MODEL

Ana Paula dos Santos Rubem

Decision-Aid Department - Center for Naval Systems Analysis
Praa Baro de Ladrio, s/n, Ilha das Cobras, Centro, 20091-000, Rio de Janeiro, RJ
anarubem@id.uff.br
Joo Carlos Correia Baptista Soares de Mello*

T
Production Engineering Department - Fluminense Federal University

IP
Rua Passo da Ptria 156, So Domingos, 24210-240, Niteri, RJ

CR
jcsmello@pesquisador.cnpq.br
Lidia Angulo Meza
Production Engineering Department - Fluminense Federal University

US
Avenida dos Trabalhadores 420, 27255-125, Volta Redonda, RJ
lidia_a_meza@pq.cnpq.br
AN
* Corresponding author
M

ABSTRACT
The lack of discrimination power and the inappropriate multipliers schemes remain
major issues in data envelopment analysis (DEA). To overcome these problems, the
ED

multiple criteria DEA (MCDEA) model was introduced in the late 1990s, drawing from
a multiple objective perspective. However, because the objectives of the MCDEA
PT

model are generally conflicting, an optimal solution satisfying all objectives

simultaneously often does not exist. Within this context, goal programming (GP)
CE

approaches were proposed to solve the MCDEA model. This paper focuses specifically
on the GP formulation, known as GPDEA. However, recently, the GPDEA models were
AC

found to be invalid, and no alternative formulation, under a GP framework, was

proposed. Therefore, the aim here is to develop a formulation for adequately solving the
MCDEA model using weighted GP. In order to do so, we point out inconsistencies in
the existing GPDEA models, and we present the WGP-MCDEA model.

Keywords: Data envelopment analysis; multiple criteria analysis; goal programming;

multiple objective programming.

2
 ACCEPTED MANUSCRIPT

1. Introduction
Data envelopment analysis (DEA) is a non-parametric method based on mathematical
programming, introduced by Charnes et al. (1978), for measuring the relative efficiency
of decision-making units (DMUs). The relative efficiency of a DMU is defined as the
weighted sum of its outputs divided by the weighted sum of its inputs on a bounded
ratio scale. In some applications (e.g, when the number of DMUs in the analysis is low
comparatively to the number of inputs/outputs considered), the lack of discrimination
among DMUs as well as an inappropriate weighting scheme can both be considered as

T
DEAs disadvantages (Li and Reeves, 1999).

IP
The possibility of an inappropriate weighting scheme causes controversy among

CR
researchers. Some (e.g., Ghasemi et al., 2014) believe that all input and output should
be regarded in the computation of the DMU relative efficiency. On the other hand,
others (e.g., Angulo-Meza and Lins, 2002) believe flexibility allows the identification of

US
the DMUs that have low performance even selecting their best feasible weights.
There are several proposals meant to overcome the aforementioned problems, some of
AN
which are reviewed, for example, by Angulo-Meza and Lins (2002) and discussed by
Ghasemi et al. (2014). One of such methods was introduced by Li and Reeves (1999),
M

the multiple criteria DEA (MCDEA) model. For that, the authors used to multiple
objective linear programming (MOLP), incorporating two additional objective functions
to the original CCR model (Charnes et al., 1978). However, Li and Reeves (1999) only
ED

found a series of non-dominated solutions with no further analysis.

The difficulty of a multiple objective problem lies on identifying a solution that
PT

simultaneously optimizes all objectives (Ghasemi et al., 2014), because, in most cases,
there is no such a feasible solution, thus requiring a set of non-dominated solutions.
CE

A development of the MCDEA model, aiming to obtain a full ranking, was done by
Soares de Mello et al. (2009) and a dual MCDEA formulation was proposed by Chaves
AC

et al. (2016). Other developments were carried out by Bal and rc (2007), Bal et al.
(2010) and Ghasemi et al. (2014), who incorporated some theoretical aspects into the
MCDEA model to simultaneously optimize its objectives. Bal and rc (2007)
proposed the use of a lexicographic goal programming to solve and assign priorities to
MCDEAs objective functions. Subsequently, Bal et al. (2010) proposed a weighted
goal programming (GP) approach, named GPDEA, to solve MCDEAs objectives
simultaneously.

3
 ACCEPTED MANUSCRIPT

Ghasemi et al. (2014) were the first to criticize the GPDEA models and identify some of
their flaws, also showing how these flaws lead to the invalidity of those models, as a
way of solving MCDEA. Actually, the main problem in GPDEA models derives from
the misuse of GP, which leads to a formulation that is not equivalent to that of MCDEA.
We further note the same misuse occurs in the formulation proposed in Bal and rc
(2007).
In addition, while renouncing the usage of goal programming, Ghasemi et al. (2014)
proposed a bi-objective multiple criteria DEA (BiO-MCDEA) model, which is solved

T
using a weighted sum of two MCDEAs objective functions.

IP
The first objective of this paper is to expand the work of Ghasemi et al. (2014), showing

CR
some other theoretical inconsistencies in GPDEAs formulation that have not been
addressed yet, and rectifying where Bal et al. (2010) have improperly formulated and
applied their GPDEAs weighted goal programming approach to solve the MCDEA
model.
US
Thus, the second and main objective of this paper is to introduce a weighted GP
AN
formulation to properly solve the MCDEA model; as, to the best of our knowledge, no
improved formulation has been developed as alternative to GPDEA models to enable
M

the correct solution of the MCDEA model by means of GP. We will restrain ourselves
to the Constant Returns to Scale (CRS) case
To prevent ambiguity, we use the term weights for the coefficients of weighted
ED

objective function in MOLP problems or weighted achievement function in GP

problems, while the word multipliers is used for input/output coefficients in MCDEA
PT

models. Moreover, we use the term efficient to particular situations within MCDEAs
context and refer to optimal Pareto solutions in MOLP problem as non-dominated.
CE

In Section 2, we review some of the DEA and MCDEAs basic aspects. Section 3
presents the GPDEA-CCR model and enumerates its inconsistencies. In Section 4, we
AC

present our alternative weighted GP approach for adequately solving the MCDEA
problem, henceforth referred to as WGP-MCDEA. Section 5 presents some final
comments.

2. Theoretical Background: DEA and MCDEA models

DEA is a method that is widely used for evaluating the relative efficiency of a set of
DMUs that consume multiple inputs and produce multiple outputs. The relative
efficiency is determined by optimizing the ratio between the weighted sum of outputs
4
 ACCEPTED MANUSCRIPT

and the weighted sum of inputs. The first DEA model, known as CCR, is due to
Charnes et al. (1978), and assumes constant returns-to-scale. Traditionally, DEA models
can be input- or output-oriented. In this paper, both the MCDEA (Li and Reeves, 1999)
and the GPDEA (Bal et al., 2010) models are presented in their original input-oriented
formulation.
Considering a process where each DMUk (k = 1, , n) uses r inputs xik (i = 1, , r) to
produce s outputs yjk (j = 1, , s), the single objective linear programming problem in
(1) denotes the multipliers linearized formulation for the input-oriented CCR model.

T
Max

IP
s. t.
(1)

CR
US
where yjk and xik are the values of output j and input i for DMUk, respectively; uj and vi
are the decision variables that denote the multipliers (weights) assigned to the output j
and input i, respectively; and ho is the relative efficiency of DMUo, the DMU under
AN
evaluation. In (1), the DMUo is efficient if and only if ho = 1, meaning that the
inequality constraint associated to DMUo is active (i.e., the slack is null).
M

Li and Reeves (1999) stated that it is possible to define the deviation do= (1 - ho) and
use it, alternatively, as a measure of inefficiency in place of ho, so that DMUo is
ED

efficient if and only if do = 1. Otherwise, the lower do is, the less inefficient is the DMU.
Including the deviation variable, the second constraint in (1) becomes the second
constraint in model (2). In addition, they proposed the inclusion of two additional
PT

objective functions in the linearized CCR multipliers formulation. The first additional
MCDEAs objective minimizes the maximum deviation within the set of evaluated
CE

DMUs (i.e., Min Max dk , k=1,...,n), and it is referred to as minimax. The second
additional MCDEAs objective minimizes the sum of the deviations (i.e., Min ),
AC

and it is called minisum. The linearized MCDEA model is presented in (2).

5
 ACCEPTED MANUSCRIPT

Min (or Max )

Min
Min
s. t. (2)

where dk is the deviation for DMUk (k = 1, , n) and M is the variable corresponding

T
of the linearization of the minimax objective function. Moreover, only the deviation for

IP
DMUo (i.e., do) is limited to the interval [0, 1]. Deviations for the other DMUs (i.e.,
dko) can be greater than unity.

CR
In MCDEA, DMUo is minimax efficient if and only if the do value corresponding to the
solution that minimizes the second objective function in (2) is null. Analogously, DMUo

US
is minisum efficient if and only if the do value corresponding to the solution that
minimizes the third objective function in (2) is null. Thus, to be minimax or minisum
AN
efficient, the DMU must necessarily be efficient in traditional DEA sense.

3. A critical analysis of the GPDEA model

M

GP provides means of striving towards the achievement of several objectives

simultaneously, it can be thought of as a way for dealing with MCDEA. Its basic
ED

approach consists of setting a specific numeric goal (or aspiration level) for each
objective function and then seeking a solution that minimizes the weighted sum of
PT

unwanted deviations of these objective functions from their respective goals.

Bal and rc (2007) and Bal et al. (2010) proposed the use of GP to solve MCDEA
CE

model. In the former, the authors applied the lexicographic method to solve the
achievement function, while the weighted sum is used in the latter.
AC

In this paper, we demonstrate that, conversely to the statement of Bal et al. (2010), their
weighted approach solution design is not an equivalent composite objective form to
MCDEAs problem. The GPDEA-CCR formulation of Bal et al. (2010) is shown in (3).
{ }
s. t.
(3)

6
 ACCEPTED MANUSCRIPT

where, for DMUo, and are the unwanted deviations for the goal that constraints
the weighted sum of inputs to unity; and are the unwanted deviations for the
goal that makes the weighted sum of outputs less than or equal to unity; and are
the unwanted deviations for the goal that makes M as the maximum deviation, where all
dk deviation variables are also unwanted. The achievement function {

T
} implies that equal weights are assigned to all

IP
unwanted deviations.

CR
In the following, we highlight the problems of formulation in (3). In fact, Ghasemi et al.
(2014) were those who first noticed and highlighted some of those problems.

US
Notwithstanding, while Ghasemi et al. (2014) emphasized GPDEAs weak
discrimination and poor multipliers distribution, we focus on the problems related to the
formulation of the model itself. Thus, herein, we perform a distinct type of critical
AN
analysis, in which we seek to exhaustively enumerating each fault comprised in the
misconceived formulation presented in (3).
M

The GPDEA-CCR model in (3) includes five inconsistencies:

(i) converting MCDEAs the first constraint (i.e., ) in goal;
ED

(ii) formulating a goal that suits a minimization objective while converting

MCDEAs first objective (considering Max ) into goal;
PT

(iii) not segregating MCDEAs third constraint (i.e., from the

aspiration level assigned in the conversion of its minimax objective in goal;
CE

(iv) use of excessive number of deviation variables in the conversion of minimax

objective in goal; and
(v) inserting the sum of deviations (decision variables) of the MCDEA model -
AC

which represents the minisum objective - in the achievement function.

Although Ghasemi et al. (2014) previously outlined the first aforementioned
inconsistency, here we analyse it in further detail. The other four inconsistencies are
first addressed in this work.

3.1. First inconsistency

7
 ACCEPTED MANUSCRIPT

Recall that the main purpose of GP is to transform conflicting multiple objectives in

specific numerical goals. Thus, the first inconsistencies refers to the assignment of
unwanted deviations and to the first constraint of MCDEA model in (2) (i.e.,
) and not to an objective, as preconized by the fundamentals of GP. This
constraint actually comes from traditional DEA models and allows the conversion of the
original fractional programming problem into the linearized formulation (1).
As demonstrated by Ghasemi et al. (2014), such a constraint must be satisfied to render
correctness. In fact, within GP terminology, this type of constraint, which must

T
necessarily be fulfilled, is referred to as a hard or technical constraint, and thus, it shall

IP
not be transformed into a goal (for further insight, see, e.g., Caballero et al., 1997).

CR
Ghasemi et al. (2014) also remark that in an input-oriented model (as those presented
here), it is absolutely necessary to set the constraint , and then seek to
achieve output levels that are as high as possible. This process is a fundamental aspect

US
of scaling and benchmarking, where it is imperative to fix either the weighted sum of
inputs (as should be done here) or the weighted sum of outputs (in case of an output-
AN
oriented model not addressed herein) to be unitary, before proceeding to determine the
other.
Nonetheless, in (3), the weighted sum of inputs for DMUo (i.e.,
M

) can be null
(when and , because, through its first goal, given by
ED

, when the minimization of the unwanted deviations leads to and

, then , a condition that is highly
problematic. Indeed, this problem occurs because, according to the fundamentals of the
PT

CCR model that stem from its fractional form, in which the relative efficiency is given
by , if the weighted sum of inputs is null, then relative
CE

efficiency will be undetermined.

Conversely, in the linearized formulation (1), the aforementioned condition is
AC

unachievable due to the insertion of the constraint .

As shown here above, in the GPDEA-CCR model, due to the first inconsistency of its
formulation, if the unwanted deviations and are unitary and null, respectively, then
the weighted sum of inputs will be null. Notwithstanding, inappropriately, this fact by
itself does not lead to undetermined relative efficiency scores, as it should, to render
consistency with DEAs fundamentals.

8
 ACCEPTED MANUSCRIPT

This inconsistency (i.e., the incorrectness of relative efficiency results) occurs because
if, in addition, in the second goal of GPDEA-CCR model given by

, when, beyond and , the deviations and are also unitary

and null, respectively, then, both the weighted sum of outputs and the weighted sum of
inputs will be null.
This occurs because, when and , then, the second goal turns into
. In this case, although the relative

T
efficiency score should be undetermined (to be consistent with the fundamentals of
DEA, as aforementioned), for the GPDEA-CCR formulation in (3), as

IP
, then, from the third goal, given by

CR
(
then, for DMUo, do (inefficiency) will be null, and thus ho (efficiency) will be unitary (a

US
completely inconsistent result, later illustrated in Table 2).
To illustrate the implications of the first inconsistency in the formulation of the
AN
GPDEA-CCR model (including the inconsistency in the relative efficiency results), we
resort to the hypothetical dataset consisting of 10 DMUs with four inputs and four
outputs, as exhibited in Table 1, originally reported by Bal et al. (2010), and herein
M

reproduced for ease of reference. To avoid distortions arising from different input-
output ranges, we normalize the data of Table 1, dividing each inputoutput by their
ED

maximum value, as recommended by Bal et al. (2010).

Table 1 Hypothetical dataset
PT

Inputs Outputs
DMU
x1 x2 x3 x4 y1 y2 y3 y4
1 32 50 82 46 47 93 54 65
CE

2 61 56 68 37 88 56 92 80
3 42 58 45 34 94 65 80 80
4 73 39 88 81 50 53 93 97
AC

5 45 38 68 41 47 42 70 52
6 86 62 44 32 86 45 100 47
7 38 74 71 74 83 91 62 74
8 61 54 70 62 79 60 72 98
9 84 52 38 47 85 68 51 41
10 87 47 31 52 78 95 70 92

Table 2 presents some of the results derived for the GPDEA-CCR model in (3) applied
to the data in Table 1, after the normalization procedure. From Table 2, we note that, for
such data set, the GPDEA-CCR model results in unitary values for and and null

9
 ACCEPTED MANUSCRIPT

values for and . As previously explained in this section, these results for the
deviation variables of the GP procedure lead to null weighted sum of inputs and null
weighted sum of outputs, and thus in null multipliers for all input-output variables.
However, mistakenly, in such cases, GPDEA-CCR model assigns unitary relative
efficiency scores, as those exhibited in Table 2, for DMUs 1 and 5. In this situation, the
correctly derived scores for those DMUs should be undetermined.
Table 2 GPDEA-CCR results based on the normalized data
DMU v1 v2 v3 v4 u1 u2 u3 u4 d1- d1+ d2- d2+ d0 h0

T
1 0 0 0 0 0 0 0 0 1 0 1 0 0 1

IP
2 0.350 0.997 0 0 0.298 0.412 0.464 0 0 0 0.052 0 0.052 0.948
3 0.369 1 0 0 0.313 0.433 0.488 0 0 0 0 0 0 1

CR
4 0.293 0.875 0.052 0.241 0 0.464 0.797 0 0 0 0 0 0 1
5 0 0 0 0 0 0 0 0 1 0 1 0 0 1
6 0.296 0.844 0 0 0.252 0.349 0.392 0 0 0 0.212 0 0.212 0.788
7
8
9
0.213
0.257
0.337
0.733
0.883
0.960
0
0
0
0.190
0.229
0
0.065
0.079
0.287
0.352
0.424
0.396
US0.564
0.679
0.446
0
0
0
0
0
0
0
0
0
0.255
0.177
0.229
0
0
0
0.255
0.177
0.229
0.745
0.823
0.771
AN
10 0.356 1.014 0 0 0.303 0.419 0.471 0 0 0 0 0 0 1

Note that, except for DMUs 1 and 5, the relative efficiency scores exhibited in Tables 2
are equal to those reported by Ghasemi et al. (2014, p. 643) as true values because
M

those provided by Bal et al. (2010) were not accurate. For DMUs 1 and 5, although
Ghasemi et al. (2014, p. 642) acknowledge the relative efficiency scores obtained from
ED

GPDEA-CCR can be unitary in case of null multipliers for all input-output variables,
they reported null values instead. Contrariwise, we insisted in reporting the inconsistent
unitary values derived from the application of the GPDEA-CCR model, to emphasize
PT

the incorrectness of the model, and because null scores are not correct either. As
previously discussed in this section, the scores for DMUs 1 and 5 should undetermined
CE

to comply with the fundamentals of the CCR model in its original fractional form.
In addition, also note the multipliers herein displayed differ from those of Ghasemi et
AC

al. (2014), because we used normalized data as was claimed to be the case in Bal et al.
(2010), who did not present the optimal multipliers derived for their GPDEA-CCR
model. Nevertheless, null multipliers in Tables 2 match those reported in Ghasemi et al.
(2014).

3.2. Second inconsistency

10
 ACCEPTED MANUSCRIPT

The second inconsistency is related to the conversion of the first objective function of
the MCDEA model in (2) (traditional CCR objective function) into a goal.
Recall this objective comprises the maximization of the relative efficiency
(i.e., Max or, alternatively, the minimization of inefficiency
(i.e., Min . Bal et al. (2010) use the first approach.
According to the fundamentals of GP (see, e.g., Caballero et al., 1997), when converting
an objective of maximization into goal, the resulting goal should be of the type
objective function aspiration level, which, after the inclusion of the deviation

T
variables, should turn into objective function + d- - d+ = aspiration level, where d- is

IP
the unwanted deviation to be minimized.

CR
However, Bal et al.s (2010) conversion of the first MCDEAs objective denoted by the
maximization of the relative efficiency ( ) resulted in the goal given by

deviation US
, with the subsequent minimization of the unwanted
in the achievement function. By doing so, what is actually (and
improperly) being performed by the GPDEA-CCR model is the minimization of the
AN
relative efficiency (i.e., ) instead of its maximization. In other words,
minimizing in the aforementioned goal is equivalent to imposing the restriction that
M

, as and . Both the proper formulation of the goals and

setting of the unwanted deviations to be inserted on the corresponding achievement
ED

function will be addressed in Section 4.

PT

3.3. Third inconsistency

The third inconsistency refers to the conversion of the second MCDEAs objective
CE

(minimax) in goal. Recall that, as explained in Section 2, the variable M is introduced

only to linearize the minimax objective, that is,
Min
AC

Min Max

Thus, the third inconsistency in the formulation of the GPDEA-CCR model consists of
not separating the hard constraints , which linearize the minimax objective
in the MCDEA model and must be fulfilled, from the aspiration level to be assigned to
the goal that converts the minimax objective (denoted by the variable M) into goal. This
occurs because, while formulating the GPDEA-CCR model in (3), Bal et al. (2010)

11
 ACCEPTED MANUSCRIPT

inserted the deviation variables related to the conversion of minimax objective in goal
(i.e., and ) directly (and improperly) in those hard constraints.

3.4. Fourth inconsistency

The fourth inconsistency is derived from the third inconsistency and refers to the fact
that when fixing an aspiration level to M, it is not necessary to assign k different pairs of
deviations to the hard constraints (but only one pair to M itself) to convert
the minimax objective in goal. Thus, only two deviation variables and would be

T
necessary, and in the achievement function should be replaced by . In fact,

IP
the use of unnecessary unwanted deviation variables in the GPDEA-CCR formulation

CR
may eventually make it more difficult to achieve the corresponding goal.

3.5. Fifth inconsistency

US
The fifth inconsistency is related to the third MCDEAs objective (minisum) because
the corresponding sum of the deviation decision variables ( ),which individually
AN
represent the deviation from efficiency of each DMU and whose sum denotes the
minisum objective, is directly (and improperly) inserted in the achievement function of
M

the GPDEA-CCR model in (3).

There are two problems with this procedure: (i) the minisum objective should have been
ED

converted in goal, with the insertion of two deviation variables, and (ii) at least one or
even both of these deviation variables should be inserted in the achievement function of
the GPDEA-CCR model, instead of the MCDEAs decision variables themselves.
PT

Due to this inconsistency, the relative efficiencies derived from the GPDEA-CCR
model are equivalent to those obtained by the individual optimization of the minisum
CE

objective in the MCDEA model in (2). Although Ghasemi et al. (2014) have observed
this equivalence; they have not discussed or explained the reasons behind it.
AC

Note that all of the aforementioned inconsistency also affect the formulation presented
by Bal and rc (2007). Thus, due to the serious problems of the GPDEA-CCRs
formulation just highlighted, we are compelled to derive a proper goal programming
formulation that seeks to simultaneously optimizing the three objectives of the MCDEA
model in (2).

4. Proposed WGP-MCDEA-CCR model

12
 ACCEPTED MANUSCRIPT

Due to the inconsistencies of the GPDEA presented in the previous section, we will
present a GP formulation to properly solve the MCDEA model. For that, we use the
weighted sum approach for a p-dimensional GP problem as in (4):
Min (
s.t. (
(4)

(F is a feasible set and Z is unrestricted in sign)
where i is the weight attached to the ith goal; fi(Z) is the linear function of decision

T
variables z1, z2, , zp for the ith goal; gi is the aspiration level of the ith goal; and

IP
are positive and negative unwanted deviations from the aspiration levels of the ith

CR
goal, respectively.
The functional form in (4) is the one based on the development of the GPDEA model of
Bal et al. (2010) in (3), but the authors considered the particular case in which 1=2= 3

US
=1. Therefore, as our aim here is to introduce a corrected version for their original (but
incorrectly formulated) proposal, we also base our improved formulation on the
AN
achievement function form shown in (4). Although the use of equal weights is
consistent with Li and Reeves (1999) procedure of setting no preference order among
the three objectives of the MCDEA model, we will make further considerations on this
M

issue later.
Taking into account the triple-objective MCDEA problem in (2), we (or the DMs) first
ED

assign to each of the three objective functions the values g1, g2 and g3 that denote their
respective aspiration levels. The aspiration levels are those values wished to be achieved
PT

as a minimum, or to not be overcome, and even, in some cases, to be exactly achieved,

for the corresponding objective (Caballero et al., 1997).
CE

Then, by combining the objective to the aspiration level, we obtain the so-called goal,
which, as indicated by Caballero et al. (1997), must satisfy the following.
AC

For the objective of maximization, when assigning the aspiration level, the
objective must be greater or equal to the aspiration level.
For the objective of minimization, when assigning the aspiration level, the
objective must be less than or equal to the aspiration level.
For equality, that is, when it is desirable that the objective be equal the aspiration
level, the objective may be originally either of maximization or of minimization.

13
 ACCEPTED MANUSCRIPT

Provided that do is used as a measure of inefficiency in the first objective, in place of ho,
the three objectives of MCDEA model in (2) are of minimization. If ho is used in the
first objective function, this objective turns to be of maximization.
After setting the goals, we shall introduce in each goal two positive variables, the so-
called deviations, denoted by d- and d+, which represent the difference between the
aspiration level and the result achieved for the corresponding objective. According to
Caballero et al. (1997), the insertion of such variables shall be performed as follows.
If the objective is of maximization, then the goal shall be greater or equal the

T
aspiration level. This goal implies that the unwanted deviation variable must be

IP
the one that reduces the aspiration level (i.e., d-).

CR
If the objective is of minimization, then the goal shall be less than or equal to the
aspiration level. This goal implies that the unwanted deviation variable must be
the one that increases the aspiration level (i.e., d+).

US
For the equality case, the goal shall be equal to the aspiration level. This implies
that the sum of both deviation variables (i.e., d- + d+) must be considered
AN
unwanted, and, thus, minimized in the achievement function. This process is, in
fact, what is done for every goal in the achievement function form shown in (4).
M

Table 3 shows the three MCDEAs objective functions, the two possible alternatives of
goals derived to each objective, their respective representation after the inclusion of
deviation variables, and unwanted deviations to be minimized, depending on the goal
ED

selected (equality or non-strict inequality).

Table 3 Structural framework for the application of GP to MCDEA problems
PT

MCDEAs objectives Initial goals Transformed goals Unwanted deviations

CE

or or
AC

or or

or or

Following the GP methodology, the achievement function shall be the minimization of

the weighted sum of the unwanted deviations of Table 3, which may vary depending on
the initial goals set. First, in an attempt of following the initial idea of Bal et al. (2010),

14
 ACCEPTED MANUSCRIPT

let us consider the case where 1=2= 3 =1. Thus, taking into account the less restrictive
goal for each objective function, the achievement function could be represented as
Min { }. Notwithstanding, in our proposal, we prefer to let the setting of
such values be performed by the DMs, as later explained.
Next, to complete the GP formulation for the MCDEA problem, as remarked by
Caballero et al. (1997), we shall first list the set of MCDEAs original constraints,
which comprises the hard restrictions (i.e., those that must be fulfilled). Subsequently,
we list the constraints derived from the setting of the goals, referred to as soft

T
restrictions, as they are not compulsive.

IP
Thereby, our GP formulation for the MCDEA model presented in (2) is given by (5).

CR
WGP-MCDEA-CCR model:
{ }
s. t.

US
AN
(5)

M
ED

As mentioned in Section 2, do is the only deviation limited to the interval [0, 1]. For that
reason, we could have directly assigned the aspiration level g1=1 to the fourth constraint
PT

in (5). However, because the deviations for the other DMUs (i.e., dko) can be greater
than unity, we cannot set the maximum values for aspiration levels g2 and g3 a priori, as
CE

these values depend on the dataset and shall be calibrated by the DMs during the
analysis. For the sake of uniformity in the procedure, we allow all aspiration levels be
AC

set a posteriori.
Formulation in (5) improves the GPDEA-CCR model, and will be henceforth referred to
as WGP-MCDEA-CCR. The model proposed differs from that of Bal et al. (2010)
essentially by the four aspects: (i) achievement function, in terms of unwanted
deviations and usage of unequal weights; (ii) first and third constraints of MCDEA
model that remain unaltered; (iii) fourth, fifth and sixth constraints that comprise the
goals; and (iv) the setting of aspiration levels that are left to the DMs.

15
 ACCEPTED MANUSCRIPT

5. Final Comments
The main objective of this paper was to develop a formulation to properly solving the
MCDEA model using a weighted goal programming approach, as the GPDEA-CCR
model of Bal et al. (2010) notably fails in that purpose. For that, we enumerated the five
inconsistencies incorporated in the formulation of Bal et al. (2010) for their GPDEA-
CCR model, amongst which only one was previously addressed by Ghasemi et al.
(2014).
As, to the best of our knowledge, no alternative formulation has been proposed in place

T
of the GPDEA-CCR model, we derived a new formulation, named as the WGP-

IP
MCDEA-CCR model, which adequately solves the MCDEA problem using weighted

CR
GP.
From a theoretical point of view, this paper has succeeded in developing and presenting
a proper GP approach for the MCDEA.

US
In future research, we intend to numerically test the model and use it to some real world
application. Moreover, we intend to verify the possibility of extending the model to
AN
variable returns to scale (VRS).
M

Acknowledgements
We would like to thank FAPERj, CAPES and CNPq for their financial support.
ED

References
Angulo-Meza, L., & Lins, M.P.E. (2002). Review of methods for increasing
PT

discrimination in data envelopment analysis. Annals of Operations Research, 116, 225

242.
CE

Bal, H., & rkc, H.H. (2007). A Goal programming approach to weight dispersion in
data envelopment analysis. Gazi University Journal of Science, 20, 117125.
AC

Bal, H., rkc, H.H., & Celebioglu, S. (2010). Improving the discrimination power and
weights dispersion in the data envelopment analysis. Computers and Operations
Research, 37, 99107.
Caballero, R., Gmez, T., Gonzlez, M., Muoz, M.M., Rey, L., & Ruiz, F. (1997).
Programacin matemtica para economistas (Mathematical programming for
economists). (1st ed.). Mlaga: Servicio de Publicaciones y Divulgacin Cientfica de la
Universidad de Mlaga, (Chapter 4).

16
 ACCEPTED MANUSCRIPT

Charnes, A., Cooper, W.W., & Rhodes, E. (1978). Measuring the efficiency of decision
making units. European Journal of Operational Research, 2, 429444.
Chaves, M.C.C, Soares de Mello, J.C.C.B., Angulo-Meza, L. (2016). Studies of some
duality properties in the Li and Reeves model. Journal of the Operational Research
Society, 67, 474482.
Ghasemi, M.R., Ignatius, J., & Emrouznejad, A. (2014). A bi-objective weighted model
for improving the discrimination power in MCDEA. European Journal of Operational
Research, 233, 640650.

T
Li X.-B., & Reeves G.R. (1999). A multiple criteria approach to data envelopment

IP
analysis. European Journal of Operational Research, 115, 507517.

CR
Soares de Mello, J.C.C.B., Clmaco, J.C.N., & Angulo-Meza, L. (2009). Efficiency
evaluation of a small number of DMUs: an approach based on Li and Reeves's model.
Pesquisa Operacional, 29, 97110.

US
AN
M
ED
PT
CE
AC

17