An evolutionary algorithm for classifier and combination rule selection in multiple classifier systems

We introduce a multiple classifier system which incorporates a genetic algorithm in order to simultaneously and dynamically select not only the participating classifiers but also the combination rule to be used. In this paper we focus on exploring the efficiency of such an evolutionary algorithm with respect to the behaviour of the resulting multi-expert configurations. To this end we initially test the proposed system on an artificially generated dataset, and then on a problem drawn from the character recognition domain. Subsequently we proceed to investigate the performance of our system not only, in comparison to that of its constituent classifiers, but also in comparison to a number of alternative aggregation strategies ranging from a simple random selection scheme to the well-known "bagging" and "boosting" algorithms. Our results indicate that significant gains can be obtained by integrating an evolutionary algorithm into the multi-classifier systems design process.