Intrinsic Motivation for Deep Deterministic Policy Gradient in Multi-Agent Environments

Deep reinforcement learning (DRL) has been increasingly applied to multi-agent domains in recent years. However, the agent explores the environment randomly, resulting in low exploration efficiency and learning performance, which limits the development of multi-agent reinforcement learning. In this paper, we introduce a new intrinsic motivation mechanism, named Group Intrinsic Curiosity Module(GICM), into Multi-Agent Deep Deterministic Policy Gradient (MAdDpG). By using GICM, the policy encourages agents to pursue novel and surprising states for searching the current scenarios more comprehensively. We evaluate our approach on three different kinds of setting tasks in simulation platform provided by OpenAI. Empirical results show that our proposed approach achieves the better performance in policy convergence speed and higher success rate in completing the tasks than vanilla reinforcement learning algorithms.