Improving Multi-agent Planning with Unsolvability and Independent Plan Detection

Multi-Agent Planning (MAP) is a challenging issue in the field of Artificial Intelligence that has gained increasing attention in the recent years. Because of the computational complexity and demanding interaction among agents, the high execution times and the volume of communication are open problems. To address these problems, we propose a Lightweight Coordination Multi-Agent Planning (LCMAP) approach that includes modules to compute unsolvability and independent plans detection, before the actual planning phase. The unsolvability detection is employed to avoid searching for planning tasks without solution, whereas the independent plans detection identifies goals that can be carried out independently by certain agents. The latter minimizes the interaction among agents, resulting in a low communication load and lightweight coordination during the actual planning phase. With the proposed modules, LCMAP can perform the coordination process and transform the MAP into multiple single-agent planning problems. The experimental results, initially focused on loosely coupled domains, show that LCMAP is up to 1.48× faster than the Distributed Cooperative Multi Agent Planning (FMAP) approach. Compared to the FMAP performance, LCMAP attained a reduction of up to 777× in the number of messages exchanged among agents during the planning process. In addition, using the benchmarks available at Unsolvability International Competition, LCMAP detected four instances as unsolvable problems. Therefore, the LCMAP unsolvability module proved its efficiency considering different number of actions and predicates.