Scalable Multi-Agent Reinforcement Learning: Adaptive Policies with Dynamic Agent and Target

Anqi Guo, Kefan Jin, Zhe Liu

Abstract: Scalability remains a significant challenge in multi-agent reinforcement learning (MARL), particularly in environments with varying numbers of agents and targets. We propose Adaptive and Dynamic Predictive Network (ADPNet), which is able to achieve the multi-agent collaborative tasks with scalable sizes. A novel motion attention mechanism is introduced into the graph attention network, which captures behavioral correlations among multiple agents to provide more accurate attention values for multi-agent relation learning. In this way, the policy can more effectively achieve scalability for collaborative tasks involving agents of varying sizes. A VAE-based motion prediction module is proposed to estimate the behavioral patterns of surrounding agents, providing prior information for the generation of collaborative policy. A variety of modules are designed and trained in an end-to-end manner, encouraging the agent to learn some implicit information that has not been discovered by humans during the training process, thus improving the cooperation performance. The extensive evaluation in both combat and MPE environments demonstrates that ADPNet significantly outperforms existing baselines, highlighting the scalability and effectiveness in complex multi-agent systems.

External IDs:dblp:conf/rcar/GuoJL25