TRANSFER LEARNING FOR MULTI-AGENT COORDINATION
Peter Vrancx, Yann-Michaël De Hauwere, Ann Nowé · 2011
Transfer learning leverages an agent’s experience in a source task in order to improve its performance in a related target task. Recently, this technique has received attention in reinforcement learning settings. Training a reinforcement learning agent on a suitable source task allows the agent to reuse this experience to significantly improve performance on more complex target problems. Currently, reinforcement learning transfer approaches focus almost exclusively on speeding up learning in single agent systems. In this paper we investigate the potential of applying transfer learning to the problem of agent coordination in multi-agent systems. The idea underlying our approach is that agents can determine how to deal with the presence of other agents in a relatively simple training setting. By then generalizing this knowledge, the agents can use this experience to speed up learning in more complex multi-agent learning tasks.