Multi-Agent Based Classifier Combination

Abstract

In this paper, a classifier combination algorithm based on multi-agent system is presented at abstract level. The problem is modeled as people's tracing back to their homeland: once upon a time, people from a region left their homeland and resided in different places, decades later, their offspring set out to find their homeland according to the wide spread legends of the ancestors' origin. In this combination problem, the class label of a testing sample serves as the homeland, decisions made by classifiers serve as offspring's residual places, and legends are class creditability by classifiers acquired from combination training set. Messengers sent by offspring try to trace back to their homeland according to the legends. They act as agents and exchange information with one another, so that confidences of different places being the original place change gradually. After congruence among the messengers is achieved, combination decision is made. The co-decision matrix is used for information exchange between agents, thus relativity between classifiers is utilized, while it is rarely considered in Bayesian Rule. According to experiments on standard database, when the number of classifiers used in combination is small, this algorithm lead to less error than other methods, and its space complexity is lower than Behavior Knowledge Space (BKS) method. Experiments show that the algorithm is convergent.