Publication

W. Chen, Z. Wang, J. Li, S. Koenig and B. Dilkina. No Panacea in Planning: Algorithm Selection for Suboptimal Multi-Agent Path Finding. In AAAI-23 Workshop on Multi-Agent Path Finding, 2023.

Abstract: Since more and more algorithms are proposed for multi-agent path finding (MAPF) and each of them has many hyperparameters to be specified, choosing one for a specific scenario that fulfills some requirements is a very important task. Previous research in algorithm selection for MAPF mostly focused on optimal algorithms and showed that machine learning could help. In this paper, we study algorithm selection for general solvers for MAPF, which further includes choosing between different suboptimal algorithms and choosing between solvers from different hyperparameters of the same algorithm. We formulate the problem as a group of prediction problems with different optimization objectives, which handle the new tradeoff between runtime and solution quality introduced by suboptimal algorithms, and different metrics to evaluate the learning model. Then we propose a group of learning tasks to solve these production problems. We identify the issue of always using resize for inputs. We use extensive experiments to show how different learning tasks should be used for different problems. We also discuss how to choose machine learning models for MAPF algorithm selection to balance the model size and the final performance.

Download the paper in pdf.