Explorations in fuzzy classifier system architectures

The Fuzzy Classifier System paradigm is an elegant and versatile combination of evolutionary and lifetime reinforcement learning based on an underlying Fuzzy Logic structure. It possesses a powerful potential to be a general-purpose linguistically interpretable problem-solver for continuous real-valued problem domains. We present a new description and analysis of a sequence of experiments that we have conducted over the past two years to investigate Fuzzy Classifier System architectures. These experiments have been carried out in the context of a mobile robot control problem. Although some of the individual stages of this sequence of work have already been reported on as the work has proceeded, this paper contains new discussion and analysis of the work in a wider context. Classifier Systems fall into two main categories, the "Pittsburgh" and "Michigan" approaches. We have found that, despite progressive modifications that improved performance of the Michigan approach, the relatively simple Pittsburgh architecture is still able to achieve comparable performance in this application. However, there are three important caveats to this statement. First, the Pittsburgh approach typically requires more fitness evaluations to achieve a similar level of performance. Dependent on the application, this could be a serious disadvantage. Second, there is much more work to be done on tuning the performance of the Michigan approach. Third, there are far more complex problem domains in which these algorithms can be tested and compared, many of which can be identified within the current, mobile robot, application. The outcome of this future work is, at present, unknown.