# Ensembling Prioritized Hybrid Policies for Multi-agent Pathfinding

**Authors:** Huijie Tang, Federico Berto, Jinkyoo Park  
**arXiv:** [2403.07559](https://arxiv.org/abs/2403.07559)  
**Date:** 2024-03-12  
**Summary page:** https://www.summarizepaper.com/en/paper/2403.07559v2/

## Key points

- Recent advancements in Multi-Agent Reinforcement Learning (MARL) for Multi-Agent Pathfinding (MAPF) emphasize the efficacy and scalability of communication-based approaches.
- A novel method called EPH has been introduced to address challenges in navigating structured environments with dense obstacles and numerous agents.
- EPH incorporates a selective communication block to enhance agent coordination by gathering more comprehensive information within multi-agent settings.
- The model is trained using a Q learning-based algorithm, which supports three advanced inference strategies aimed at optimizing performance during execution:
- Integration of neural policies with single-agent expert guidance for efficient navigation through conflict-free zones.
- Utilization of Q value-based methods to prioritize conflict resolution and handle deadlock situations effectively.
- Introduction of an ensemble method to select the most optimal solution from multiple possibilities.
- Empirical evaluations demonstrate that EPH performs competitively against state-of-the-art neural methods for MAPF in complex multi-agent environments.
- The research has been accepted for presentation at the 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2024).
- The code for EPH is open-source and available at https://github.com/ai4co/eph-mapf.

## Summary

In the realm of <keyword1>(MARL) for <keyword2>(MAPF), recent advancements have highlighted the efficacy and scalability of communication-based approaches. To address challenges in navigating structured environments with dense obstacles and numerous agents, a novel method called <keyword3>(EPH) has been introduced. EPH incorporates a selective communication block to enhance agent coordination by gathering more comprehensive information within multi-agent settings. The model is trained using a Q learning-based algorithm, which forms the foundation for three advanced inference strategies aimed at optimizing performance during execution. Firstly, EPH integrates neural policies with single-agent expert guidance to navigate conflict-free zones efficiently. Secondly, it employs Q value-based methods to prioritize conflict resolution and handle deadlock situations effectively. Lastly, an ensemble method is introduced to select the most optimal solution from multiple possibilities. Empirical evaluations of EPH in complex multi-agent environments demonstrate its competitive performance against state-of-the-art neural methods for MAPF. This research has been accepted for presentation at the 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2024). The code for EPH is open-source and available at https://github.com/ai4co/eph-mapf. Authored by Huijie Tang, Federico Berto, and Jinkyoo Park, the study "Ensembling Prioritized Hybrid Policies for Multi-agent Pathfinding" showcases a promising approach towards enhancing <keyword1>-<keyword2> solvers in challenging scenarios characterized by intricate environmental structures and high agent densities.