Research on Attention-Based Predictive Model in Multi-Agent Reinforcement Learning

Agent actions planning is a challenging problem in multi-agent reinforcement learning. Recent methods typically build their predictive models by full connection layers, but the shortage of utilizing action and observation information in complex multi-agent environment limits the performance of those methods. To address this issue, we propose a novel predictive model based on self-attention mechanism which can effectively use information in the environment. Specifically, we use the MADDPG algorithm to make a dataset where the input is the observations and actions and the output is the predicted observations and rewards. We also introduce the multi-head attention mechanism into the predictive model so that the predictive model can better combine different information for prediction. We perform extensive experiments to assess the proposed method and compare it with the representative baseline approach. Results demonstrate that our method outperforms competitors with higher accuracy in different tasks. Our implementation will be publicly available.