A Perturbation-based Approach for Multi-classifier System Design

This paper presents a perturbation-based approach useful to select the best combination method for a multi-classifier system. The basic idea is to simulate small variations in the performance of the set of classifiers and to evaluate to what extent they influence the performance of the combined classifier. In the experimental phase, the Behavioural Knowledge Space and the Dempster-Shafer combination methods have been considered. The experimental results, carried out in the field of hand-written numeral recognition, demonstrate the effectiveness of the new approach.