Learning new behaviors is a crucial problem in behavior-based robots. This research proposes a new method of reinforcement learning, called Instance-Based Classifier Generator (IBCG), for the acquisition of reactive behaviors. In IBCG, the learning system successively memorizes a newly experienced state-action pair as an action-rule. Utility of each rule is estimated by the original temporal credit assignment procedure, which is designed so that the cooperative rules leading the system to an eventual reward should self-organize. Learning capability of IBCG is experimentally examined through a task of mobile robot navigation in both simulated and real environment. The results demonstrate that the robot with IBCG acquired behaviors such as light-seeking, collision-avoidance, and wall-following.
