Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Lightweight and Effective Preference Construction in PIBT for Large-Scale Multi-Agent Pathfinding

Keisuke Okumura, Hiroki Nagai

TL;DR

This work targets large-scale MAPF by enhancing PIBT with two lightweight tiebreakers, Hindrance and Regret, to steer action choices without increasing asymptotic complexity or sacrificing completeness. Hindrance estimates one-step avoidance costs to better dodge other agents, while Regret learns how actions affect others’ future choices, with multiple PIBT runs to update expectations. Across one-shot and lifelong settings on diverse maps, HR (the combination of Hindrance and Regret) consistently improves solution quality (notably 10–20% reduction in Sum-of-Costs in dense cases and ≥40% throughput gains in lifelong MAPF) with only modest runtime overhead. The findings suggest that simple, well-tuned local tiebreakers can markedly enhance the practicality of PIBT-based MAPF in real-time, large-scale applications, with adaptive or hybrid extensions proposed for future work.

Abstract

PIBT is a computationally lightweight algorithm that can be applied to a variety of multi-agent pathfinding (MAPF) problems, generating the next collision-free locations of agents given another. Because of its simplicity and scalability, it is becoming a popular underlying scheme for recent large-scale MAPF methods involving several hundreds or thousands of agents. Vanilla PIBT makes agents behave greedily towards their assigned goals, while agents typically have multiple best actions, since the graph shortest path is not always unique. Consequently, tiebreaking about how to choose between these actions significantly affects resulting solutions. This paper studies two simple yet effective techniques for tiebreaking in PIBT, without compromising its computational advantage. The first technique allows an agent to intelligently dodge another, taking into account whether each action will hinder the progress of the next timestep. The second technique is to learn, through multiple PIBT runs, how an action causes regret in others and to use this information to minimise regret collectively. Our empirical results demonstrate that these techniques can reduce the solution cost of one-shot MAPF and improve the throughput of lifelong MAPF. For instance, in densely populated one-shot cases, the combined use of these tiebreaks achieves improvements of around 10-20% in sum-of-costs, without significantly compromising the speed of a PIBT-based planner.

Lightweight and Effective Preference Construction in PIBT for Large-Scale Multi-Agent Pathfinding

TL;DR

This work targets large-scale MAPF by enhancing PIBT with two lightweight tiebreakers, Hindrance and Regret, to steer action choices without increasing asymptotic complexity or sacrificing completeness. Hindrance estimates one-step avoidance costs to better dodge other agents, while Regret learns how actions affect others’ future choices, with multiple PIBT runs to update expectations. Across one-shot and lifelong settings on diverse maps, HR (the combination of Hindrance and Regret) consistently improves solution quality (notably 10–20% reduction in Sum-of-Costs in dense cases and ≥40% throughput gains in lifelong MAPF) with only modest runtime overhead. The findings suggest that simple, well-tuned local tiebreakers can markedly enhance the practicality of PIBT-based MAPF in real-time, large-scale applications, with adaptive or hybrid extensions proposed for future work.

Abstract

PIBT is a computationally lightweight algorithm that can be applied to a variety of multi-agent pathfinding (MAPF) problems, generating the next collision-free locations of agents given another. Because of its simplicity and scalability, it is becoming a popular underlying scheme for recent large-scale MAPF methods involving several hundreds or thousands of agents. Vanilla PIBT makes agents behave greedily towards their assigned goals, while agents typically have multiple best actions, since the graph shortest path is not always unique. Consequently, tiebreaking about how to choose between these actions significantly affects resulting solutions. This paper studies two simple yet effective techniques for tiebreaking in PIBT, without compromising its computational advantage. The first technique allows an agent to intelligently dodge another, taking into account whether each action will hinder the progress of the next timestep. The second technique is to learn, through multiple PIBT runs, how an action causes regret in others and to use this information to minimise regret collectively. Our empirical results demonstrate that these techniques can reduce the solution cost of one-shot MAPF and improve the throughput of lifelong MAPF. For instance, in densely populated one-shot cases, the combined use of these tiebreaks achieves improvements of around 10-20% in sum-of-costs, without significantly compromising the speed of a PIBT-based planner.
Paper Structure (17 sections, 2 equations, 5 figures, 2 tables, 3 algorithms)

This paper contains 17 sections, 2 equations, 5 figures, 2 tables, 3 algorithms.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Problem Definition
  4. PIBT
  5. Related Work
  6. Hindrance
  7. Time Complexity.
  8. Regret Learning
  9. Time Complexity.
  10. Experiments
  11. Benchmarks.
  12. One-shot MAPF
  13. Effect on Hyperparameters in Regret Learning.
  14. Lifelong MAPF
  15. Extremely Dense Situations
  16. ...and 2 more sections

Figures (5)

  • Figure 1: Motivating example for calculating the $\mathit{hindrance}$ term. Goals for the agents are marked with coloured boxes.
  • Figure 2: Motivating example for calculating the $\mathit{regret}$ term. Grey arrows represent backtracking in PIBT.
  • Figure 3: Results for one-shot MAPF. The success rate of planning within 1s (top), average runtime (middle) and SoC normalised by lower bound ($1.0$ is minimum; bottom) for successful cases are shown.
  • Figure 4: Results for lifelong MAPF. Average throughput (top) and response time (bottom) are shown.
  • Figure 5: One-shot MAPF on empty in extremely dense situations.