A study on reinforcement learning mechanisms with common knowledge field for heterogeneous agent systems

Abstract

We propose a new approach to realize a reinforcement learning scheme for heterogeneous multiagent systems. In our approach, we treat the collective agents systems in which there are multiple autonomous mobile robots, and given tasks are achieved based on the collective behavior approach. Also, each agent organizes and refines its knowledge for executing its own behaviors by reinforcement learning mechanisms. Thus, we discuss the reinforcement learning mechanism by which the common knowledge is effectively learned in heterogeneous-agents systems. In our approach, a common knowledge field is generated, and then the leaned rule formed knowledge is embedded in that field. The proposed reinforcement learning mechanism is constructed based on learning classifier systems. An extended model of learning classifier systems is defined to apply the model to heterogeneous-agent systems containing the common knowledge field. We perform computer simulations for multiagent escaping problems to verify our proposed method.