Model-Based Reinforcement Learning using Model Mediator in Dynamic Multi-Agent Environment

Centralised training and decentralised execution (CTDE) is one of the most effective approaches in multiagent reinforcement learning (MARL). However, these CTDE methods still require large amounts interaction with environment, even to reach same performance as very simple heuristic-based algorithms. Although modelbased RL a prominent approach improve sample efficiency, its adaptation multi-agent setting combining existing has not been well studied literature. The few studies only consider settings relaxed restrictions on number agents observable range. In this paper, we where some information about each agent’s observations (e.g. visibility, agents) are changed dynamically. such setting, fundamental challenge how train models that accurately generate complex transitions addition central state, use it for efficient policy learning. We propose model based algorithm novel architecture consisting global local prediction mediator. evaluate our model-based applied an method challenging StarCraft II micromanagement tasks show can learn fewer interactions environment.