CLASSIFIER COMBINATION: THE ROLE OF A-PRIORI KNOWLEDGE V.DI LECCE 1 , G.DIMAURO 2 , A.GUERRIERO 1 , S.IMPEDOVO 2 , G.PIRLO 2 , A.SALZO 2 (1) Dipartimento di Ing. Elettronica -Politecnico di Bari- via Re David -70126 Bari- Italy (2) Dipartimento di Informatica - Università di Bari - Via Orabona, 4 - 70126 Bari – Italy The aim of this paper is to investigate the role of the a-priori knowledge in the process of classifier combination. For this purpose three combination methods are compared which use different levels of a-priori knowledge. The performance of the methods are measured under different working conditions by simulating sets of classifier with different characteristics. For this purpose, a random variable is used to simulate each classifier and an estimator of stochastic correlation is used to measure the agreement among classifiers. The experimental results, which clarify the conditions under which each combination method provides better performance, show to what extend the a-priori knowledge on the characteristics of the set of classifiers can improve the effectiveness of the process of classifier combination. 1 Introduction Classifier combination is a diffuse strategy that has been widely used in complex classification problems for which very high performance is required [1]. For classifier combination, many methods have been proposed so far which are generally classified into three categories depending on the amount of information they combine [2,3]. Abstract-level combination methods use the top candidate provided by each classifier [4,5,6] ; Ranked-level combination methods use the entire ranked list of candidates [7,8] ; Measurement-level combination methods use also the confidence value of each candidate in the ranked list [9,10]. Among the others, classifier combination at abstract-level is the most general approach since every classifier is able at least to provide results at abstract level. In the process of classifier combination, some kind of a-priori knowledge can also be used in order to achieve better performance. On the basis of the kind of a-priori knowledge the combination methods use, they can be classified into three categories. Methods of the first category do not require any kind of a-priori information on the combined classifiers [4,7,8]. Methods of the second category use information at the level of individual classifiers as a- priori knowledge [5]. Methods of the third category require information at the level of the entire set of combined classifiers [6,9,10]. In this paper, the role of a-priori knowledge in the process of classifier combination is investigated by comparing three combination methods: the Majority Vote Method (MV) which is of the first category [4]; the Dempster-Shafer Method (DS) which is of the second 143 In: L.R.B. Schomaker and L.G. Vuurpijl (Eds.), Proceedings of the Seventh International Workshop on Frontiers in Handwriting Recognition, September 11-13 2000, Amsterdam, ISBN 90-76942-01-3, Nijmegen: International Unipen Foundation, pp 143-152