Robustness and Evolution in an Adaptive System Application on Classification Task

Abstract

In this paper, we proposed an approach to a single-step ClassifierSystem, in which the useful population is built by progressivelyspecializing classifiers. It has been applied to a classificationtask in a medical domain. To permit the system to explore alternativeswithout making decisions earlier in learning stages, all the classifiersthat might be selected are triggered and receive the resulting rewardcorresponding to their action. The payoff function involves the classifier'sperformance, its specificity and the system's performance (its robustness).Genetic operators are activated with a probability which dependson the system's robustness. During the test stages, no further learningtakes place and the system's performance is measured by the percentageof correct classification made on the second set of examples. Whenthe measure of performance is the highest, the population is stabilizedand contains the correct classifiers (the payoff function and geneticoperators have no more effect on classifiers). This approach achievesconvergency more quickly and makes it possible to have a final accuratepopulation without over-specializing.