Authors:
Max Frommknecht
and
Pavel Surynek
Affiliation:
Faculty of Information Technology, Czech Technical University in Prague, Thákurova 9, Prague, Czechia
Keyword(s):
Multi Agent Path Finding, SAT, SAT Solver, Local Search, Initial Assignment, Conflict-Driven Clause Learning.
Abstract:
This paper explores the suitability of Stochastic Local Search (SLS) solvers for Multi-Agent Pathfinding (MAPF) translated into the SAT domain. Traditionally, SAT encodings of MAPF have been tackled using Conflict-Driven Clause Learning (CDCL) solvers, but this work investigates the potential of SLS solvers, particularly ProbSAT, in solving MAPF under the makespan objective. By employing the MDD-SAT approach and comparing the performance of ProbSAT against the Glucose 4 CDCL solver, the effects of eager and lazy encodings are evaluated, as well as the benefit of providing initial variable assignments. Results show that ProbSAT can effectively solve MAPF instances, especially when initial assignments based on agents’ shortest paths are provided. This study suggests that SLS solvers can compete with CDCL solvers in specific MAPF scenarios and highlights future research directions for optimizing SLS performance in MAPF.