Reproduce “Cooperative Reward Shaping for Multi-Agent Pathfinding” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhenyu Song, Ronghao Zheng, Senlin Zhang, Meiqin Liu
A new cooperative reward shaping (CoRS) method is introduced to enhance Multi-Agent Pathfinding (MAPF) in distributed reinforcement learning. It encourages agents to consider neighbors' potential rewards, leading to improved pathfinding success rates by over 20% and reduced makespan in high-density scenarios compared to a baseline, while maintaining computational simplicity.
Key insights
- CoRS promotes active cooperation by shaping an agent's reward based on the maximum possible average reward its action enables for its neighbors, moving beyond selfish behaviors.
- The theoretical analysis demonstrates that CoRS, under certain conditions, aligns individual optimal policies with the Individual-Global-Max (IGM) condition, ensuring agents maximize their own shaped reward while also contributing to overall system efficiency.
- Approximating the cooperativeness metric by considering only local neighbors and decoupling the maximization over their actions ensures the reward shaping remains stable during training and computationally feasible for large-scale systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.