Generalization in Wilson's classifier system

Abstract

We analyze generalization with the XCS classifier system when the system is applied to animat problems in grid-worlds. Our aim is to give a unified view of generalization with XCS, in order to explain some of the phenomena reported in the literature. Initially, we apply XCS to two environments. Our results show that there are situations in which the generalization mechanism of XCS may prevent the system from converging to optimum. Accordingly, we study XCS's generalization mechanism analyzing the conditions under which the system may fail to evolve an optimal solution. We draw a hypothesis in order to explain the results reported so far. Our hypothesis suggests that XCS fails to learn an optimal solution when, due to the environment structure and to the exploration strategy employed, the system does not explore all the areas of the environment frequently. We thus introduce a meta exploration strategy that is used as theoretical tool to validate our hypothesis experimentally.