A Hybrid Evolutionary Algorithm for Bayesian Networks Learning: an Application to Classifier Combination C. De Stefano, F. Fontanella, C. Marrocco and A. Scotto di Freca Universit` a di Cassino Via G. Di Biasio, 43 02043 Cassino (FR) – Italy {destefano,fontanella,cristina.marrocco,a.scotto}@unicas.it Abstract. Classifier combination methods have shown their effectiveness in a number of applications. Nonetheless, using simultaneously multiple classifiers may result in some cases in a reduction of the overall performance, since the responses provided by some of the experts may generate consensus on a wrong decision even if other experts provided the correct one. To reduce these undesired effects, in a previous paper, we proposed a combining method based on the use of a Bayesian Network. The structure of the Bayesian Network was learned by using an Evolutionary Algorithm which uses a specifically devised data structure to encode Direct Acyclic Graphs. In this paper we presents a further improvement along this direction, in that we have developed a new hybrid evolutionary algo- rithm in which the exploration of the search space has been improved by using a measure of the statistical dependencies among the experts. Moreover, new ge- netic operators have been defined that allow a more effective exploitation of the solutions in the evolving population. The experimental results, obtained by using two standard databases, confirmed the effectiveness of the method. 1 Introduction The idea of combining the results provided by different experts for improving the over- all classification rate has been widely investigated in the literature and it is now an active area of research in the fields of Machine Learning and Pattern Recognition [1–3]. The rationale behind this idea is that the weakness of each single expert may be compensated without losing the strength of each of them, thus obtaining an overall performance that can be better than that of any single expert. Even if many studies have been published in the literature, which demonstrate, theoretically or empirically, the effectiveness of com- bining methods and their advantages over individual classifier models [4], their use may result in some cases in a reduction of the overall performance. This effect is mainly due to the fact that the responses provided by some of the experts may generate consensus on a wrong decision, even if other classifiers in the combining pool provided the correct class. Thus, the main problem to be solved is that of defining a combining rule able to solve these conflicts and to take the right classification decision even when the experts disagree. We believe that an effective way to overcome the above drawbacks is that of con- sidering the combined effects of the whole set of responses provided by the experts on