Multi-Agent Reinforcement Learning With Decentralized Distribution Correction

This work considers decentralized multi-agent reinforcement learning (MARL), where the global states and rewards are assumed to be fully observable, while the local behavior policy is preserved locally for resisting adversarial attack. In order to cooperatively accumulate more rewards, the agents exchange messages among a time-varying communication network to reach consensus. For these cooperative tasks, we propose a decentralized actor-critic algorithm, where the agents make individual decisions, but the joint behavior policy is optimized towards more cumulative rewards. We provide the theoretical analysis towards the convergence under the tabular setting and then expand it to nonlinear function approximations. Furthermore, by incorporating decentralized distribution correction, the agents are trained in an off-policy manner for higher sample efficiency. Finally, we conduct experiments to evaluate the algorithms, where the proposed algorithm performs competitively in both stability and asymptotic performance. Note to Practitioners —Fully decentralized MARL algorithms are widely applied in multi-agent systems for generating cooperative behaviors, e.g., multiple unmanned aerial vehicles (UAV) cooperatively performing search and rescue tasks, multiple vehicles efficiently passing a crowded intersection, and multiple robots cooperatively handling cargo or obstacles. Focusing on these potential applications, this work is motivated to improve the sample efficiency of recent decentralized MARL algorithms by incorporating off-policy training approaches. In this work, we reweight historical trajectories via a decentralized average consensus step and develop corresponding policy-optimization procedures, with which previous trajectories could be used to stabilize later iterations. Since the training materials are augmented by historical samples, the sample efficiency is significantly improved, and the training process is stabilized. With the fully decentralized training approach, the proposed algorithms are expected to be applied in large-scale systems, e.g., vehicle teams and UAV groups, for effective real-time control.