No Panacea in Planning: Algorithm Selection for Suboptimal Multi-Agent Path Finding

TL;DR

The paper tackles selecting solvers for suboptimal MAPF to balance runtime and solution quality rather than optimizing runtime alone. It builds a large, image-based dataset from a standard MAPF benchmark, defines two objective families (score-based and bound-based), and trains multiple computer vision models to predict the best solver or its hyperparameters. The findings show that no single loss or model excels across all objectives; regression approaches can better capture gap-based performance, while classification with traditional CV models like ViT performs robustly, leading to practical guidance on rescaling strategies and model choice. The work highlights the practical impact of objective-aligned learning and demonstrates that hyperparameter selection can be integrated within the same predictive framework.

Abstract

Since more and more algorithms are proposed for multi-agent path finding (MAPF) and each of them has its strengths, choosing the correct one for a specific scenario that fulfills some specified requirements is an important task. Previous research in algorithm selection for MAPF built a standard workflow and showed that machine learning can help. In this paper, we study general solvers for MAPF, which further include suboptimal algorithms. We propose different groups of optimization objectives and learning tasks to handle the new tradeoff between runtime and solution quality. We conduct extensive experiments to show that the same loss can not be used for different groups of optimization objectives, and that standard computer vision models are no worse than customized architecture. We also provide insightful discussions on how feature-sensitive pre-processing is needed for learning for MAPF, and how different learning metrics are correlated to different learning tasks.

Figures (6)

• Figure 1: The frequency of each algorithm being the best one on a given map when $w=0.001$ in Eq. \ref{['eq:score_def']}. Different maps have different numbers of total scenarios because different maps are different in size and obstacles so the total capacity is naturally different.
• Figure 2: Overall performance of different algorithm selection models in different tasks in terms of accuracy (higher is better) and VBS-SBS gap (lower is better). Blue points are actual samples we collected in different models.
• Figure 3: Dataset result for the case of $score=time+1 \times cost$ as score definition.
• Figure 4: Dataset result for the case of $score=time+0.1 \times cost$ as score definition.
• Figure 5: Dataset result for the case of $score=time+0.001 \times cost$ as score definition on hyperparameter selection (EECBS) dataset.