Deep Skill Chaining with Diversity for Multi-agent Systems*

Abstract

AbstractMulti-agent reinforcement learning requires the reward signals given by the environment to guide the convergence of individual agents’ policy networks. However, in a high-dimensional continuous space, the non-stationary environment may provide outdated experiences that lead to the inability to converge. The existing methods can be ineffective in achieving a satisfactory training performance due to the inherent non-stationary property of the multi-agent system. We propose a novel reinforcement learning scheme, MADSC, to generate an optimized cooperative policy. Our scheme utilizes mutual information to evaluate the intrinsic reward function that can generate a cooperative policy based on the option framework. In addition, by linking the learned skills to form a skill chain, the convergence speed of agent learning can be significantly accelerated. Hence, multi-agent systems can benefit from MADSC to achieve strategic advantages by significantly reducing the learning steps. Experiments are performed on the SMAC multi-agent tasks with varying difficulties. Experimental results demonstrate that our proposed scheme can effectively outperform the state-of-the-art methods, including IQL, QMIX, and hDQN, with a single layer of temporal abstraction.KeywordsReinforcement learningMulti-agent systemsTemporal abstractionMutual informationSkill discovery