The application of reinforcement learning to multi-agent systems has attracted recent attention. In multi-agent systems, the state space to be handled constitutes a major problem efficiently in learning of agents. In order to cooperate agents in the same environment, it is needed to observe and evaluate the action of other agents in the multi-agent system. This case increases the dimension of state space proportional to the number of agents, exponentially. This paper presents a novel approach to overcome this problem. The approach uses together the advantages of the modular architecture, internal model and fuzzy logic in multi-agent systems. In our cooperation method, one agent estimates its action according to the internal model of the other agent. The internal model is acquired by the observation and evaluation of the other agent's actions. Fuzzy logic maps from input fuzzy sets, representing state space of each learning module to the output fuzzy sets representing the action space. The fuzzy rule base of each learning module is built through the Q-learning. Experimental results handled on pursuit domain show the effectiveness and applicability of the proposed approach.
展开▼
