Investigating Arbitration Strategies in an Animat Navigation System

This paper reports on recent experiments applying classifier systems to the problem of supporting both local and global navigation in a simulated animat. The basis of this research is a hybrid learning system that extends the classifier representation to enable environmental feedback to impinge directly upon the classifier population. The system applies a connectionist representation to the condition sets of classifiers which enables the direct encoding of classifier condition/fitness values onto network nodes. The goal of the system is to achieve domain objectives by calibrating classifier behaviour during the exploration of the domain and evolving new classifiers to exploit the domain by discovering goal states.