Anytime Multi-Agent Path Finding using Operation Parallelism in Large Neighborhood Search

TL;DR

This work tackles the scalability of anytime MAPF by introducing DROP-LNS, a parallel destroy-repair framework that executes multiple destroy-repair operations concurrently and updates the best-known solution asynchronously. By employing a main thread to dispatch tasks to worker threads, along with synchronization strategies that balance productivity and exploitation, DROP-LNS achieves higher-quality solutions within fixed time budgets compared to state-of-the-art baselines like MAPF-LNS and LaCAM*. The study demonstrates that asynchronous best-known updates and parallel operation significantly improve solution quality on diverse MAPF benchmarks, while highlighting how synchronization overhead and problem scale influence gains. These results suggest practical benefits for large-scale, real-time MAPF applications and point to avenues for further optimization, including refined synchronization and GPU-based parallelism.

Abstract

Multi-Agent Path Finding (MAPF) is the problem of finding a set of collision-free paths for multiple agents in a shared environment while minimizing the sum of travel time. Since solving the MAPF problem optimally is NP-hard, anytime algorithms based on Large Neighborhood Search (LNS) are promising to find good-quality solutions in a scalable way by iteratively destroying and repairing the paths. We propose Destroy-Repair Operation Parallelism for LNS (DROP-LNS), a parallel framework that performs multiple destroy and repair operations concurrently to explore more regions of the search space within a limited time budget. Unlike classic MAPF approaches, DROP-LNS can exploit parallelized hardware to improve the solution quality. We also formulate two variants of parallelism and conduct experimental evaluations. The results show that DROP-LNS significantly outperforms the state-of-the-art and the variants.

Figures (7)

• Figure 1: Conceptual timelines of different parallelism variants with a main thread "M" and two worker threads "1" and "2". Black blocks indicate the productivity, i.e., the time when threads are processing, and red lines indicate the idle time. Blue arrows indicate events where a worker thread receives tasks, and purple dotted arrows indicate events where a worker thread returns a solution.
• Figure 2: Illustrative example of the DROP-LNS framework with a main thread "M" and two worker threads "1" and "2". Arrows are the actions from each thread.
• Figure 3: Average solution quality among all instances with the highest number of agents on each map solved by SYNC-LNS, DETA-LNS, and DROP-LNS with 2, 4, 8, and 16 threads, respectively. The lower the average suboptimality ratio, the better the solution quality.
• Figure 4: Solution quality of instances solved by LaCAM*, MAPF-LNS, and DROP-LNS. Instances are grouped by the number of agents. The lower the average suboptimality ratio, the better the solution quality.
• Figure 5: Sum of delays (in thousands) versus runtime (in seconds) in two instances with 300 agents in Room map and with 3000 agents in Den520d respectively solved by LaCAM*, MAPF-LNS, and DROP-LNS. The pink region indicates the AUC of DROP-LNS.