The paper proposes a novel approach called LayeredMAPF to decompose a MAPF instance involving a large number of agents into multiple isolated subproblems with fewer agents. The key highlights are:
The decomposition process involves three steps:
a. Identifying initial clusters of agents based on their dependence paths.
b. Refining the clusters by bipartitioning them into smaller clusters.
c. Further decomposing the clusters into levels with a specific solving order.
The decomposition ensures the solvability of each subproblem by verifying the existence of a solution under a simplified scenario where agents only need to avoid the targets of previous subproblems and the starts of later subproblems.
A framework is presented to enable general MAPF algorithms to solve each subproblem independently and merge their solutions into a conflict-free final solution. This framework handles the differences between serial and parallel MAPF methods.
Extensive experiments are conducted on classic MAPF benchmarks to evaluate the impact of decomposition on time cost, memory usage, success rate, and solution quality. The results demonstrate that the decomposition can significantly reduce the memory usage and time cost, particularly for serial MAPF methods.
The source code of the proposed algorithm is made publicly available to facilitate further research within the community.
To Another Language
from source content
arxiv.org
Ключові висновки, отримані з
by Zhuo Yao,Wei... о arxiv.org 04-22-2024
https://arxiv.org/pdf/2404.12773.pdfГлибші Запити