Reinforcement Learning Approach to Acquisition of Motion Patterns for Autonomous Robotic Systems

摘要

A learning classifier system, defined in the continuous state and action spaces, is outlined in this paper. The system adopts the following distinctive mechanisms: instance-based generation of action-rules; introduction of generalized rules; combined policy of rule evaluation that include profit sharing and non-conservative bucket brigade scheme. Feasibility of the system is tested under a simulation example where two planar robots are coupled via a common object. The simulation results show that motion patterns of cooperative behavior can be evolved using the classifier system.