Rule Extraction from Neural Network Robot Controller using Classifier Systems

Biological inspired learning systems have proven to be very powerful techniques for learning robot controllers. However, in some cases the learned concepts are unreadable for humans. This prevents a deep semantic analysis of what has been really learned by those systems. We present in this paper an alternative to obtaining rule representation of a neural network robot controller. In the first step we acquire a data set resulting from the execution of an NN controller in different environments. Then, a Classifier System is used for automatically generating the symbolic rules. Classifier Systems are special production systems where conditions and actions are codified in order to learn new rules by means of Genetic Algorithms (GA). These systems combine the execution capabilities of symbolic systems and the learning capabilities of Genetic Algorithms. The results show that the obtained symbolic rules system is very accurate modeling the NN controller.