Strategic Interaction Multi-Agent Deep Reinforcement Learning

Despite the proliferation of multi-agent deep reinforcement learning (MADRL), most existing typical methods do not scale well to the dynamics of agent populations. And as the population increases, the dimensional explosion of joint state-action and the complex interaction between agents make learning extremely cumbersome, which poses the scalability challenge for MADRL. This paper focuses on the scalability issue of MADRL with homogeneous agents. In a natural population, local interaction is a more feasible mode of interplay rather than global interaction. And inspired by the strategic interaction model in economics, we decompose the value function of each agent into the sum of the expected cumulative rewards of the interaction between the agent and each neighbor. This novel value function is decentralized and decomposable, which enables it to scale well to the dynamic changes in the number of large-scale agents. Hereby, the corresponding strategic interaction reinforcement learning algorithm (SIQ), is proposed to learn the optimal policy of each agent, wherein a neural network is employed to estimate the expected cumulative reward for the interaction between the agent and one of its neighbors. We test the validity of the proposed method in a mixed cooperative-competitive confrontation game through numerical experiments. Furthermore, the scalability comparison experiments illustrate that the scalability of the SIQ algorithm outperforms the independent learning and mean field reinforcement learning algorithms in multiple scenarios with different and dynamically changing numbers.