Reinforcement learning for autonomous mobile robots by forming approximate classificatory concepts

摘要

This paper presents a method for an autonomous robot's behavior-acquisition using a reinforcement learning (RL) method-concept intensive Q-learning. We first attempt to form a robot's classificatory class-concepts on the objects in the world using a concept formation technique and construct its decision models at multiple abstraction levels for determining its action sequence to behave appropriately in the world. Instead of establishing direct mapping between perceptual states and actions as the conventional RL, we develop a method for acquiring behavior via those models and show that our method can improve the performance as well as the transparency of learning.