Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Reevaluation of Large Neighborhood Search for MAPF: Findings and Opportunities

Jiaqi Tan, Yudong Luo, Jiaoyang Li, Hang Ma

TL;DR

This work reevaluates large neighborhood search for MAPF (MAPF-LNS) under a unified evaluation framework, exposing critical inconsistencies in prior assessments and implementing baselines to enable fair comparisons. Across six benchmark maps and diverse agent counts, rule-based neighborhood strategies, particularly RandomWalk and RandomWalkProb, consistently deliver strong time efficiency and near-optimal delays, while learning-based methods (SVM-LNS, Neural-LNS, Bandit-LNS) incur substantial overheads and offer no clear advantage. The study also clarifies that replan solvers like PP outperform PBS in most scenarios, and that initial solution quality is less decisive than iteration dynamics. The authors propose concrete future directions, including targeting high-delayed agents, contextual bandits, learned replanning orders, and dynamic neighborhood sizing, which can guide more effective integration of machine learning into MAPF-LNS while adhering to rigorous benchmarking standards.

Abstract

Multi-Agent Path Finding (MAPF) aims to arrange collision-free goal-reaching paths for a group of agents. Anytime MAPF solvers based on large neighborhood search (LNS) have gained prominence recently due to their flexibility and scalability, leading to a surge of methods, especially those leveraging machine learning, to enhance neighborhood selection. However, several pitfalls exist and hinder a comprehensive evaluation of these new methods, which mainly include: 1) Lower than actual or incorrect baseline performance; 2) Lack of a unified evaluation setting and criterion; 3) Lack of a codebase or executable model for supervised learning methods. To address these challenges, we introduce a unified evaluation framework, implement prior methods, and conduct an extensive comparison of prominent methods. Our evaluation reveals that rule-based heuristics serve as strong baselines, while current learning-based methods show no clear advantage on time efficiency or improvement capacity. Our extensive analysis also opens up new research opportunities for improving MAPF-LNS, such as targeting high-delayed agents, applying contextual algorithms, optimizing replan order and neighborhood size, where machine learning can potentially be integrated. Code and data are available at https://github.com/ChristinaTan0704/mapf-lns-unified.

Reevaluation of Large Neighborhood Search for MAPF: Findings and Opportunities

TL;DR

This work reevaluates large neighborhood search for MAPF (MAPF-LNS) under a unified evaluation framework, exposing critical inconsistencies in prior assessments and implementing baselines to enable fair comparisons. Across six benchmark maps and diverse agent counts, rule-based neighborhood strategies, particularly RandomWalk and RandomWalkProb, consistently deliver strong time efficiency and near-optimal delays, while learning-based methods (SVM-LNS, Neural-LNS, Bandit-LNS) incur substantial overheads and offer no clear advantage. The study also clarifies that replan solvers like PP outperform PBS in most scenarios, and that initial solution quality is less decisive than iteration dynamics. The authors propose concrete future directions, including targeting high-delayed agents, contextual bandits, learned replanning orders, and dynamic neighborhood sizing, which can guide more effective integration of machine learning into MAPF-LNS while adhering to rigorous benchmarking standards.

Abstract

Multi-Agent Path Finding (MAPF) aims to arrange collision-free goal-reaching paths for a group of agents. Anytime MAPF solvers based on large neighborhood search (LNS) have gained prominence recently due to their flexibility and scalability, leading to a surge of methods, especially those leveraging machine learning, to enhance neighborhood selection. However, several pitfalls exist and hinder a comprehensive evaluation of these new methods, which mainly include: 1) Lower than actual or incorrect baseline performance; 2) Lack of a unified evaluation setting and criterion; 3) Lack of a codebase or executable model for supervised learning methods. To address these challenges, we introduce a unified evaluation framework, implement prior methods, and conduct an extensive comparison of prominent methods. Our evaluation reveals that rule-based heuristics serve as strong baselines, while current learning-based methods show no clear advantage on time efficiency or improvement capacity. Our extensive analysis also opens up new research opportunities for improving MAPF-LNS, such as targeting high-delayed agents, applying contextual algorithms, optimizing replan order and neighborhood size, where machine learning can potentially be integrated. Code and data are available at https://github.com/ChristinaTan0704/mapf-lns-unified.
Paper Structure (36 sections, 4 figures, 17 tables, 1 algorithm)

This paper contains 36 sections, 4 figures, 17 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Background: MAPF and MAPF-LNS
  4. Neighborhood Selection
  5. Discussion on Selection Strategies
  6. A Unified Setting for Evaluation
  7. Environments
  8. Initial Solution
  9. Replan Solver
  10. Neighborhood Size and Number of Agents
  11. Evaluation Criterion
  12. Implementation Details
  13. Findings and Future Directions
  14. Key Insights and Analyses
  15. Outlooks on Improving MAPF-LNS
  16. ...and 21 more sections

Figures (4)

  • Figure 1: Delay vs. Time in maps with a Medium number of agents. Error bars represent the variance of delay across $25$ different scenes. The best rule-based strategies are Adaptive for empty and warehouse, RandomWalk for random, and RandomWalkProb for ost003d, den520d, and Paris.
  • Figure 2: Delay vs. Iteration in maps with a Medium number of agents. Error bars represent the variance of delay across $25$ different scenes. The best rule-based strategies are Adaptive for empty and warehouse, RandomWalk for random, and RandomWalkProb for ost003d, den520d, and Paris.
  • Figure 3: Delay vs. Iteration when the replan solver is PBS or PP. The neighborhood selection strategy is RandomWalk with a neighborhood size of 25 in all cases.
  • Figure 4: Final Delays using the Best and Worst neighborhood size within $300$s. RW, INT, RAND, ADP, and RWP stand for RandomWalk, Intersection, Random, Adaptive, and RandomWalkProb. The blue and pink columns indicate the highest and lowest delays. The neighborhood sizes are labeled at the top of each subfigure.