Collaborative learning agents with structural classifier systems

We propose a new learning agent architecture for collaborative learning. To learn any complicated task in multi-agent environment, simple reinforcement architectures have limitations on learning. Therefore, we propose splitting learning mechanism into three separate layers to learn required behavior, which are respectively organized by the Classifier Systems [Holland 86]. It can learn to communicate with other agents, to make plans, and to select actions based on the plans and other-agents' behavior. We show that these agents can select cooperative actions as a collaborative group.