Multi-Agent Distributed Deep Deterministic Policy Gradient for Partially Observable Tracking

In many existing multi-agent reinforcement learning tasks, each agent observes all the other agents from its own perspective. addition, training process is centralized, namely critic of can access policies agents. This scheme has certain limitations since every single only obtain information neighbor due to communication range in practical applications. Therefore, this paper, a distributed deep deterministic policy gradient (MAD3PG) approach presented with decentralized actors and critics realize tracking. The distinguishing feature proposed framework that we adopted execution, where takes agent’s agents’ into account. Experiments were conducted tracking tasks based on particle environments N(N=3,N=5) track target partial observation. results showed method achieves higher reward shorter time compared methods, including MADDPG, DDPG, PPO, DQN. novel leads more efficient effective