- Abstract: Decentralized decision makers learn to cooperate and make decisions in many domains including (but not limited to) search and rescue, drone delivery, box pushing and fire fighting problems. In these cooperative domains, a key challenge is one of sparse rewards, i.e., rewards/reinforcements are obtained only in a few situations (e.g., on extinguishing a fire, on moving a box) and in most other situations there is no reward/reinforcement. The problem of learning with sparse reinforcements is extremely challenging in cooperative Multi-Agent Reinforcement Learning (MARL) problems due to two reasons: (a) Compared to the single agent case, exploration is harder as multiple agents have to be coordinated to receive the reinforcements; and (b) Environment is not stationary as all the agents are learning at the same time (and therefore change policies) and therefore the limited (due to sparse rewards) good experiences can be quickly forgotten. One approach that is scalable, decentralized and has shown great performance in general MARL problems is Neural Fictitious Self Play (NFSP). However, since NFSP averages best response policies, a good policy can be drowned in a deluge of bad best-response policies that come about due to sparse rewards. In this paper, we provide a mechanism for imitation of good experiences within NFSP that ensures good policies do not get overwhelmed by bad policies. We then provide an intuitive justification for why self imitation within NFSP can improve performance and how imitation does not impact the fictitious play aspect of NFSP. Finally, we provide a thorough comparison (experimental or descriptive) against relevant cooperative MARL algorithms to demonstrate the utility of our approach.
- Keywords: Deep Reinforcement Learning, Cooperative Multi Agent Systems, Sparse Reward, Decentralized Decision Making