SMAPPO: Stein Variational Gradient Descent-Based Multi-Agent Proximal Policy Optimization

With the development of multi-agent reinforcement learning (MARL), advancements have been achieved in robotics, communications, and unmanned aerial vehicle (UAV) systems and et al. A critical challenge in MARL research lies in the limited diversity of agent policies. This paper proposes a Stein Variational Gradient Descent-based Multi-Agent Proximal Policy Optimization (SMAPPO) that quantifies the discrepancy between historical and current policies, thereby guiding policy exploration to enhance behavioral diversity. Experiment shows the effectiveness of work in promoting diverse policy adoption through coordinated exploration.