Véronique Delcroix

A Bayesian network for recurrent multi-criteria and multi-attribute decision problems: Choosing a manual wheelchair

This paper discusses recurrent multi-criteria, multi-attribute decision problems. Because of the possibility of decision-maker ignorance or low decision-maker involvement the decision problem structuring is done once for all by a group of experts and does not involve the implication of the decision makers. We propose an original model based on Bayesian networks, which provides a decision process that helps the decision-maker to select an appropriate alternative among a set of alternatives, taking into account multiple criteria that are often conflicting. Our model makes it possible to represent in the same model the decision case (i.e., the decision-maker characteristics, contextual characteristics, their needs and preferences), the set of alternatives with the different attributes, and the choice criteria. The model allows us to compute the value of three essential elements: the importance of each criterion, which is based on the decision-case characteristics; each criterions evaluation index in terms of the alternative; and each criterions satisfaction index. The recurrent problem of choosing a manual wheelchair (MWC) illustrates the construction and use of our model.

Fuzzy Evidence in Bayesian Network

The Bayesian Networks are graphical models that are easy to interpret and update. These models are useful if the knowledge is uncertain, but they lack some means to express ambiguity. To face this problem, we propose Fuzzy Evidence in Bayesian Networks and combine the Fuzzy Logic and Bayesian Network. This has allowed to benefit from mutual advantages of these two approaches, and to overcome the problem of data and observation ambiguity. This paper proposes an inference algorithm which uses the Bayesian Network and Fuzzy Logic reliability. This solution has been implemented, tested and evaluated in comparison with the existing methods.

Comparative study of supervised classification algorithms for the detection of atmospheric pollution

Abstract The management of atmospheric pollution using video is not yet widespread. However it is an efficient way to evaluate the polluting rejects coming from large industrial facilities when traditional captors are not usable. This paper presents a comparison of different classifiers for a monitoring system of polluting smokes. The data used in this work are stemming from a system of video analysis and signal processing. The database includes the pollution level of puffs of smoke defined by an expert. Six machine learning techniques are tested and compared to classify the puffs of smoke: k-nearest neighbour, naive Bayes classifier, artificial neural network, decision tree, support vector machine and a fuzzy model. The parameters of each type of classifier are split into three categories: learned parameters, parameters determined by a first step of the experimentation, and parameters set by the programmer. We compare the results of the best classifier of each type depending on the size of the learning set. A part of the discussion concerns the robustness of the classifier facing the case where classes of interest are under-represented, as the high level of pollution in our data.

On the Effectiveness of Diversity When Training Multiple Classifier Systems

Discussions about the trade-off between accuracy and diversity when designing Multiple Classifier Systems is an active topic in Machine Learning. One possible way of considering the design of Multiple Classifier Systems is to select the ensemble members from a large pool of classifiers focusing on predefined criteria, which is known as the Overproduce and Choose paradigm. In this paper, a genetic algorithm is proposed to design Multiple Classifier Systems under this paradigm while controlling the trade-off between accuracy and diversity of the ensemble members. The proposed algorithm is compared with several classifier selection methods from the literature on different UCI Repository datasets. This paper specifies several conditions for which it is worth using diversity during the design stage of Multiple Classifier Systems.

Uncertain Evidence in Bayesian Networks: Presentation and Comparison on a Simple Example

We consider the problem of reasoning with uncertain evidence in Bayesian networks (BN). There are two main cases: the first one, known as virtual evidence, is evidence with uncertainty, the second, called soft evidence, is evidence of uncertainty. The initial inference algorithms in BNs are designed to deal with one or several hard evidence or virtual evidence. Several recent propositions about BN deal with soft evidence, but also with ambiguity and vagueness of the evidence. One of the proposals so far advanced is based on the fuzzy theory and called fuzzy evidence. The original contribution of this paper is to describe the different types of uncertain evidence with the help of a simple example, to explain the difference between them and to clarify the appropriate context of use.

Understanding soft evidence as probabilistic evidence: Illustration with several use cases

This paper aims to get a better understanding of the notions of evidence, probabilistic evidence and likelihood evidence in Bayesian Networks. Evidence comes from an observation of one or several variables. Soft evidence is probabilistic evidence, since the observation consists in a local probability distribution on a subset of variables that has to replace any former belief on these variables. It has to be clearly distinguished from likelihood evidence, also called virtual evidence, for which the evidence is specified as a likelihood ratio. Since the notion of soft evidence is not yet widely understood, most of the Bayesian Networks engines do not propose related propagation functions and the terms used to describe such evidence are not stabilised. First, we present the different types of evidence on a simple example with an illustrative context. Then, we discuss the understanding of both notions in terms of knowledge and observation. Next, we propose to use soft evidence to represent certain evidence on a continuous variable, after fuzzy discretization.

Bayesian network based on the method of AHP for making decision

The multi-criteria decision problem is to find the most satisfactory solution among many alternatives, taking into account several criteria that may be conflicting. The proposed method focuses on the person since the importance (weight) given to each criterion is defined according to the characteristics of the person. We use a special structure of Bayesian network (BN) based on the method AHP. The graph of the BN and the probabilities associated with nodes are designed to translate the knowledge of experts on the selection of an alternative.

From Information to Evidence in a Bayesian Network

Evidence in a Bayesian network comes from information based on the observation of one or more variables. A review of the terminology leads to the assessment that two main types of non-deterministic evidence have been defined, namely likelihood evidence and probabilistic evidence but the distinction between fixed probabilistic evidence and not fixed probabilistic evidence is not clear, and neither terminology nor concepts have been clearly defined. In particular, the term soft evidence is confusing. The article presents definitions and concepts related to the use of non-deterministic evidence in Bayesian networks, in terms of specification and propagation. Several examples help to understand how an initial piece of information can be specified as a finding in a Bayesian network.

Comparison of Several Classifiers for the Detection of Polluting Smokes

This paper addresses the pollution detection problem by using a camera and analyzing the pictures. A camera is used to record visual scenes around complex plants. Then several signals are computed to describe the pictures. Our aim is to detect among the various clouds if there are polluting smokes. We assume in this paper that the signals are useful to classify the clouds and that we do not need other data. In this paper two types of classifiers are studied: Bayesian networks and a k-nearest neighbour classifier.

Supervised Classification Algorithms Applied to the Detection of Atmospheric Pollution

The management of atmospheric polluting reject is based on the ability to measure this pollution. This paper deals with the case where no local sensor can be used, inducing the use of video to detect and evaluate the atmospheric pollution coming from large industrial facilities. This paper presents a comparison of different classifiers used in a monitoring system of polluting smokes detected by cameras. The data used in this work are stemming from a system of video analysis and signal processing. The database also includes the pollution level of plumes of smoke defined by an expert. Several Machine Learning techniques are tested and compared. The experimental results are obtained from a real world database of polluting rejects. The parameters of each type of classifier are split in three categories: learned parameters, parameters determined by a first step of the experimentation, and parameters set by the programmer. The comparison of the results of the best classifier of each type indicates that all of them provide good results.