Multi-Agent Reinforcement Learning Policy Transfer by Buffer

The learning complexity of learning from scratch in a new environment is getting higher because of the existence of multi-agent, which costs enormous calculating resources and time. In our work, to accelerate the training of multi-agent reinforcement learning in the new environment, we propose a method to transfer knowledge by measuring the differences between environments. The method mainly includes two aspects: (1) policy and transition model learning in the source environment;(2) Accelerate training by using knowledge from the source environment. We validated our approach in The StarCraft Multi-Agent Challenge (SMAC) [1]. Experimental results show that this method can effectively transfer knowledge and speed up the training algorithm.