An extension to the XCS classifier system for stochastic environments

We analyze XCS learning capabilities in stochastic environments where the result of agent actions can be uncertain. We show that XCS can cope when the degree of uncertainty is limited. We propose an extension to XCS, called XCSm, that can learn optimal solutions for higher degrees of uncertainty. We test XCSm when the uncertainty affects the whole environment and when the uncertainty is limited to some areas. Finally, we show that XCSm is a proper extension of XCS, in that it coincides with it when it is applied to deterministic environments.