GENETIC ALGORITHMS WITH DIVERSITY MEASURES TO BUILD CLASSIFIER SYSTEMS

Abstract

The combination of classifiers is an active research area of the machine learning and pattern recognition communities. Many t heoretical and empirical studies have been published demonstrating the advantages of the paradigm of combination of classifiers over the individual classifiers. When combining classifiers it is important to guarantee the diversity among them [16]. Some statistical measures can be used to estimate how diverse the ensembles of classifiers are, they are called diversity measures. On the other hand the number of individual classifiers are very big and just a litter group of classifiers can be generated a large number of combinations, so emerge the idea to use one of the meta-heuristics: the genetic algorithms (gas). Genetic algorithms play a significant role as search technique for handling complex spaces in many fields. They are based on the underlying genetic process in biological organisms and on the natural evolution principles of populations. These algorithms process a population of chromosomes, which represent search space solutions, with three operations: selection, crossover and mutation. In this paper some diversity measures are presented and one modulation of genetic algorithm using diversity measures is enunciated and implemented in order to obtain, from all possible combinations of a large number of base classifiers, a combination that ensures greater diversity among the chosen classifiers and the best accuracy in multi classifier system. We also present and discuss the results of applying the implemented system in two fields of application. Finally, the general conclusions are exposed.