Professional Documents
Culture Documents
net/publication/235916564
CITATIONS READS
10 1,541
6 authors, including:
All content following this page was uploaded by Véronique Delcroix on 15 January 2015.
Abstract
Multi-Criteria Decision Making (M CDM ) methods involve the selection of the
best (adequate) solutions from a set of available alternatives taking into account
multiple, often conflicting, criteria. The aim of this paper is to propose a partic-
ular structure of Bayesian networks to model M CDM problems. The problem
that we address aims to provide assistance to a person to select an appropriate
object among a large set of alternatives. The person and the alternatives are each
characterized by several attributes that influence both the importance of the cri-
teria and the quality of the alternatives according to the criteria. Our framework
is illustrated on a real example which concerns the problem of choosing a manual
wheelchair.
Keywords: bayesian network, multi-criteria, decision making, importance, satis-
faction, person, quality of alternative
1 Introduction
In many situations humans have to make a decision to choose an alternative among
several; think for example of buying a flight ticket, dinner in restaurant, etc.
Some of the major problems in the decision process are that the factors and the
criteria to be considered are multiple, sometimes complex and often conflicting,
and that the number of available alternatives may be too large to be explicitly
handled. The objective of Multi-Criteria Decision Making (M CDM ) Roy (1985);
Roy and Bouyssou (1993); Bouyssou et al. (2006) is to model and solve such
complex decision problems. An M CDM problem can be defined by: a set of
decision alternatives (objects, actions or choice) A={a1 , a2 , ..., an }, a set of criteria
C={c1 , c2 ,..., cm } according to which desirability of an alternative is judged, and
a set of couples (ai , cj ) or cj (ai ) which evaluate an alternative ai for criterion cj .
The objective is to choose the “best” alternative with is often defined as the one
with the highest score (utility).
In this paper, we present a model based on a particular structure of Bayesian
network for modeling MCDM problems. Our objective is to provide assistance to a
person to select an appropriate alternative among a great set of alternatives. The
general idea of the proposed approach is to take into account important pieces of
286 Karima Sedki et al.
Characteristics: concern the parameters (or attributes) that identify the person
or alternatives which are considered in the decision process. The characteristics
to be considered depend on the concerned decision problem.
Criteria: they are the specific elements that we consider to determine the de-
sirability of each alternative. The set C = {c1 , c2 , . . . cm } denotes the set of
criteria.
Index: For each criterion, we associate an index which is a numerical function
on the set of alternatives. The index of a criterion for a given alternative
represents the quality of this alternative regarding this criterion. Each index
depends on a subset of characteristics of alternatives.
An alternative is well characterized by its characteristics’ values. However, any
combination of characteristics’ values does not specify an alternative. For example,
assume that for the transport problem, there are two characteristics: Country ∈
{Italy,China,N ewZealand} and T ransportM ode ∈ {Train,Car,Plane}. It is clear
that the combination (N ewZeland,Car) does not represent a possible alternative.
In this sense, we define the following concepts:
Real alternative: Concerns the alternatives that can be chosen (possible choices).
They can be saved in a database. In the previous example, the combination
(N ewZeland, Car) does not constitute a real alternative.
Silhouette of choice: It is obtained by performing inference in the model (i.e
the constructed Bayesian network) from the known characteristics of the per-
son. So, each characteristic of alternatives is associated with a probability
distribution. The combination of characteristics’ values which have highest
probabilities in the silhouette of choice represents an Ideal alternative. It is
possible that the ideal alternative does not exist among the set of possible
alternatives, namely there is no real alternative corresponding to this ideal
alternative. An Example of silhouette of choice is given in Table 1.
If the alternative (Italy, P lane) exists, it is an ideal alternative for the person
concerned by the choice.
Table 2: Meaning of the values of importance, satisfaction and index of a given criterion
of these two variables requires the intervention of domain experts. The variable
satisfaction of a criterion is to maximize. It reflects the satisfaction of the person
for a given alternative regarding a criterion. The satisfaction variable for each
criteria c can be defined by the following function:
Sat(c) = Index(c)Imp(c) (1)
This function has been chosen since it presents the following characteristics
(cf. figure 2): (1) if a given criterion is indifferent for the person, she/he will be
satisfied for this criterion whatever the index of a given alternative. This means
that if the importance is close to 0, the satisfaction is close to 1; (2) if the alternative
presents a good index for a given criterion, the person will be satisfied whatever
the importance of this alternative for this criterion. Namely, if the index is close
to 1, the satisfaction is close to 1; (3) whatever the degree of importance of a
criterion, the higher is the index, the higher is the satisfaction, but the lower is
the importance, the more easily the person will be satisfied. This function is used
to generate the conditional probability table for satisfaction variable 2 .
The graph of the BN may include arcs between several characteristics of the
person, for example knowing the size of the person gives an idea on the length of
its legs. The same is true between attributes of the alternative.
The model may also include direct arcs from a characteristic of the person
to a characteristic of the alternative to represent a constraint such as “the total
width of the FRM must be less than the width of the smallest passage used by
the person”.
The complete model is built from the association of all the Bayesian networks
(BN) associated with each criterion represented in Figure 1. The association con-
sists simply to identify all common variables between these BN. An example of a
complete model for choosing a manual wheelchair is given in Section 3.
then perform inference to get the satisfaction of the criteria. Formally, the objec-
P erson
tive of this step is to compute the probabilities P (Sat(ci )|Cobs , C Alternative ) for
P erson
each criterion ci where Cobs is the sub-set of person’s characteristics that are
observed.
Figure 3: The graph of a Bayesian network for choosing a manual wheelchair regarding
three criteria.
observed variables and the nodes importance are the target ones. The result of
the first inference (see Figure 4) shows that the most important criterion for this
person is the maneuverability of the MWC, which is explained by his active profile.
Whereas, the comfort of rest is considered irrelevant for him. Indeed, these two
criteria are usually conflicting. The comfort of loading the MWC in car is consid-
ered as relatively important for this person because he loads himself the MWC in
a small car.
The next step allows to obtain a silhouette of choice. Hence, we perform a
second inference after having fixed the values of the satisfaction nodes to their
maximals values (i.e the satisfaction of each criterion is very important). This
provides a probability distribution for characteristics of the MWC which are tar-
get nodes. We can observe that several variables have a preferred value, namely
with a high probability. For example, it is recommended for this person to
choose an MWC with a folding frame and removable brackets (in Figure 5, we
have P (T ypeF rame=F olding)=0.559 and P (RemovableBrackets=Y es)=0.757).
Theses values induce a good index of the criterion CLC (see Figure 5). Recall
that to maximize the satisfaction when the importance is high, it is necessary that
the index will be relatively good. This is observed on the Figure 5.
The third step consists to choose the best real MWC from the silhouette of
choice. Considering only characteristics that affect the three treated criteria, there
exists effectively an MWC where values of these characteristics correspond to those
294 Karima Sedki et al.
Figure 4: Result of the first inference: importance of criteria for the person.
concerns the decision nodes. In case several decisions are considered, this model
deals with a sequence of decisions. In our proposition, the decision concerns the
attributes of the alternatives. In that representation, each attribute of the alterna-
tive could be considered as a decision, but they do not constitute a chronological
sequence. For example, while choosing a MWC, the order of the characteristics of
the MWC considered by the specialist who helps the person to choose is highly
dependent of the specialist. Determining a sequence of the characteristics of the
alternative is not feasible. In order to consider a single decision as “the choice of
an alternative”, it would be necessary to add a node “FRM” (or “alternative”) as
a parent of all the attributes of the alternatives, whose values are the list of real
MWC. In that eventuality, the definition of this node and of all its children should
be updated to take into account each new real alternative. On the contrary, the
model that we propose is not dependent of the set of real alternatives. It only
depends of the list of attributes that must be taken in account. The same kind of
representation is choosen in Fenton and Neil (2001) that also deals with MCDM
problem. The authors use a BN, and not an ID, in which the decision noded (such
as “mode of transport” or “start time” in a travel example) are represented by
chance node.
296 Karima Sedki et al.
Another remark about influence diagram concerns the nodes of utility. There
is no probability on these nodes. A utility function is supposed to be known with
certainty. In our model, the importance and index nodes can be either determinis-
tic or probabilistic. That is to say that our model can manage uncertainty on the
quality of alternatives for a criteria, and also the uncertainty on the importance
of a criteria for a person. The distinction between certain and uncertain criteria
is explained in Fenton and Neil (2001). In their proposition, the BN only includes
uncertain criteria. Next, a traditionnal MCDM technique is used to combine all
criteria.
Finally, utility nodes does not have children, except another utility function.
But in that case, the inference is not possible. In our proposition, we expect that
the value of the importance influences the index, and thus the characteristics of
the FRM.
Among the well known MCDM methods, one can cite (Cho (2003)) the Analytic
Hierarchy Process (AHP), Bayesian Analysis, multi-attribute utility/value theory,
and Outranking Methods such as ELECTRE Figueira et al. (2005). In those
method, the person concerned by the choice is not really taken in account. This
person can define the pairwise comparison matrix for criteria, but this may be
a difficult task. In some cases, the person is not able to evaluate the relative
importance between two criteria in her/his personal situation. In our method,
the person has to give personal pieces of information, but nothing else. The
knowledge of the experts about the importance of the criteria depending on their
characteristics is include in the BN, both in the structure and in the function of
the nodes of importance.
Another advantage of our method is that it clearly distinguish between the
attributes of the alternative and the criteria. The function used to evaluate the
quality of an alternative regarding a criteria allows a real flexibility with the at-
tributes. For example, it allows to deal with alternatives that do not all have
exactly the same characteristics. In the classical methods mentioned above, the
criteria are directly associated with an attribute of the alternative, or a simple
combination (sum) of some attributes. In the model proposed in Fenton and Neil
(2001), the uncertain criteria are defined by a function of some attributes of the
alternatives. However, the distintion between criterion and attribute is not very
clear (a node such as “start time” seems to be considered as a criteria in some
pages and not in others).
In order to compare the above methods with ours, it would be interesting to
use the BN constructed as explain in the previous section to obtain the index of
each criteria for all the real alternatives. This values, associated with the degrees
of importance of each criterion (step 1 of our proposition), constitute the base
of classical MCDM methods. This approach is very near of the one proposed
by Fenton and Neil (2001). They propose to use a particuler BN to evaluate all
the alternatives on uncertain criteria, before applying a traditional technique to
combine the values of all criteria for each alternative. The fact to evaluate one
by one all the alternatives may be very long if there is a lots of alternatives, since
an inference is needed each time. Our proposition require only three inferences.
Another comparison would be interesting about the method that we propose:
first, use our BN to compute the satisfaction on each criteria, for all the real
Bayesian Network Model for Decision Problems 297
alternative, and next, select the alternative producing the best average satisfaction.
We suppose that this naı̈ve and expensive method would give he same result as
the one we propose, but this is to be proven! This is one of the next step of our
work!
4.2 Conclusion
We have proposed in this paper an M CDM framework based on a Bayesian net-
work and centered on the person. This approach takes into account important
pieces of information for the decision process. These information are related to
the person concerned by the choice regarding his/her characteristics and to the
alternatives described by several parameters. The satisfaction associated to an
alternative is determined jointly by two components: the analysis of the quality of
the alternative for each criterion (i.e index of a criterion) and the analysis of the
level of importance required for each criterion (i.e importance of a criterion). The
proposed method requires three inferences in the Bayesian network: an inference
to compute the importance of the criteria; a second to propose a silhouette of
choice that gives preferred values for each characteristic of the alternatives, via
a probability distribution. The third inference computes the level of satisfaction
for each criterion about the real alternative selected according to the silhouette of
choice. This method allows to obtain directly the best alternative, without having
to evaluate each alternative.
As the number of possible alternatives can be very large (in the case of our
application, the database contains over 400 M W C without forgetting the possible
settings). Several techniques can be exploited to reduce the number of alternatives
to be considered for computing distances. We can, for example, firstly consider
the strong constraints that will discard some alternatives. We can also discard the
real alternatives where the value of one of its characteristics takes zero probability
(or low) in the silhouette of choice.
The proposed method based on Bayesian network has many advantages. The
most important is certainly the fact that we consider the person’s characteristics
concerned by the choice to define importance of the criteria. The part of the
model that concerns the person may be more or less developed according to the
information that we dispose about the person. It is not asked to inform all the
characteristics of the complete model. The information that are not provided
are replaced by prior probabilities. Furthermore, the particular structure of the
proposed model can take into account the very specific criteria, such as the capacity
of disabled people. It is not embarrassing that a criterion concerns only certain
people. It is sufficient to correctly define the importance of this criterion: if the
importance of a criterion is equal to zero for a person, then this means that this
criterion is indifferent to him and will not affect the choice. Another important
point is that dependency between criteria is treated because of some criteria are
some characteristics in common.
Another advantage of this method is that it allows to evaluate the alternatives
for each criterion with respect to the person. This is important because many
M CDA methods require that these evaluation are known with certainty (and re-
gardless of the person). The proposed method allows us to combine many criteria,
298 Karima Sedki et al.
References
Darwiche Adnan (2009), Modeling and Reasoning with Bayesian Networks, Cambridge
University Press, New York, NY, USA, ISBN 0521884381, 9780521884389.
Denis Bouyssou, Thierry Marchant, Marc Pirlot, Alexis Tsoukis, and Philippe
Vincke (2006), Evaluation and decision models with multiple criteria: Stepping stones
for the analyst, International Series in Operations Research and Management Science,
Volume 86, Boston, 1st edition, ISBN 0-387-31098-3.
Keun Tae Cho (2003), Multicriteria decision methods: An attempt to evaluate and unify,
Mathematical and computer modelling, 37(9-10):1099–1119.
Norman Fenton and Martin Neil (2001), Making Decisions: Using Bayesian Nets and
MCDA, Knowledge-Based Systems, 14:307–325.
J. Figueira, V. Mousseau, and B. Roy (2005), ELECTRE methods, in Multiple Cri-
teria Decision Analysis: State of the Art Surveys, pp. 133–162, Springer, London.
Finn V. Jensen (1996), An Introduction to Bayesian Networks, UCL Press.
Norsys (1998), Netica Application, Norsys Software Corp., http://www.norsys.com.
R. M. Oliver and J. Q. Smith (1990), Influence Diagrams, Belief Nets and Decision
Analysis, John Wiley.
Judea Pearl (1988), Probabilistic reasoning in intelligent systems: networks of plau-
sible inference, Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, ISBN
0-934613-73-7.
B. Roy (1985), Méthodologie multicritère d’aide à la décision, Economica, Paris.
B. Roy and D. Bouyssou (1993), Aide Multicritere a la Decision: Methodes et Cas,
Edition Economica.