Improved Traffic Flow Optimisation in Lifelong Multi-Agent Path Finding

Peiqian Lin, Zhe Chen, David L. Dowe, Daniel Harabor

Keywords: Multi-Agent Path Finding, Heuristic Search, Combanatorial Optimisation, MAPF, Lifelong MAPF

Abstract:

Lifelong Multi-Agent Path Finding (LMAPF) addresses the fundamental problem of coordinating teams of agents continuously performing tasks in a shared environment, a core challenge in applications like automated fulfilment centres and robotics. Recent approaches that guide fast, myopic planners with high-level, congestion-aware routes have shown great success, enabling rapid coordination for thousands of agents. However, these methods estimate congestion by assuming each agent will follow one specific, pre-computed, time-independent path. This overlooks the existence of multiple cost-equivalent paths, among which a low-level planner is free to choose at runtime, leading to potentially inaccurate congestion predictions and low-quality guidance.
In this paper, we propose a novel method for computing congestion-aware guidance that reasons about this ambiguity.
Instead of committing to a single route, our approach calculates a probabilistic traffic flow by considering the likelihood of travelling across all cost-equivalent paths.
By using this probability distribution to determine expected congestion, we generate more robust and accurate guidance heuristics.
This allows planners to make more informed decisions.
Our extensive experiments on large-scale LMAPF benchmarks with up to 16,000 agents show that our approach consistently outperforms baseline algorithms in system throughput.

Area: Search, Optimization, Planning, and Scheduling (SOPS)

Generative A I: I acknowledge that I have read and will follow this policy.

Submission Number: 575