Anytime Multi-Agent Path Finding with an Adaptive Delay-Based Heuristic

Authors

  • Thomy Phan University of Southern California
  • Benran Zhang University of Southern California
  • Shao-Hung Chan University of Southern California
  • Sven Koenig University of California, Irvine

DOI:

https://doi.org/10.1609/aaai.v39i22.34495

Abstract

Anytime multi-agent path finding (MAPF) is a promising approach to scalable and collision-free path optimization in multi-agent systems. MAPF-LNS, based on Large Neighborhood Search (LNS), is the current state-of-the-art approach where a fast initial solution is iteratively optimized by destroying and repairing selected paths of the solution. Current MAPF-LNS variants commonly use an adaptive selection mechanism to choose among multiple destroy heuristics. However, to determine promising destroy heuristics, MAPF-LNS requires a considerable amount of exploration time. As common destroy heuristics are stationary, i.e., non-adaptive, any performance bottleneck caused by them cannot be overcome by adaptive heuristic selection alone, thus limiting the overall effectiveness of MAPF-LNS. In this paper, we propose Adaptive Delay-based Destroy-and-Repair Enhanced with Success-based Self-learning (ADDRESS) as a single-destroy-heuristic variant of MAPF-LNS. ADDRESS applies restricted Thompson Sampling to the top-K set of the most delayed agents to select a seed agent for adaptive LNS neighborhood generation. We evaluate ADDRESS in multiple maps from the MAPF benchmark set and demonstrate cost improvements by at least 50% in large-scale scenarios with up to a thousand agents, compared with the original MAPF-LNS and other state-of-the-art methods.

Downloads

Published

2025-04-11

How to Cite

Phan, T., Zhang, B., Chan, S.-H., & Koenig, S. (2025). Anytime Multi-Agent Path Finding with an Adaptive Delay-Based Heuristic. Proceedings of the AAAI Conference on Artificial Intelligence, 39(22), 23286–23294. https://doi.org/10.1609/aaai.v39i22.34495

Issue

Section

AAAI Technical Track on Multiagent Systems