A leader-following paradigm based deep reinforcement learning method for multi-agent cooperation games.

Abstract

Multi-agent deep reinforcement learning algorithms with centralized training with decentralized execution (CTDE) paradigm has attracted growing attention in both industry and research community. However, the existing CTDE methods follow the action selection paradigm that all agents choose actions at the same time, which ignores the heterogeneous roles of different agents. Motivated by the human wisdom in cooperative behaviors, we present a novel leader-following paradigm based deep multi-agent cooperation method (LFMCO) for multi-agent cooperative games. Specifically, we define a leader as someone who broadcasts a message representing the selected action to all subordinates. After that, the followers choose their individual action based on the received message from the leader. To measure the influence of leader's action on followers, we introduced a concept of information gain, i.e., the change of followers' value function entropy, which is positively correlated with the influence of leader's action. We evaluate the LFMCO on several cooperation scenarios of StarCraft2. Simulation results confirm the significant performance improvements of LFMCO compared with four state-of-the-art benchmarks on the challenging cooperative environment.