Admissible Landmark Heuristic for Multi-Agent Planning

Heuristics are a crucial component in modern planning systems. In optimal multiagent planning the state of the art is to compute the heuristic locally using only information available to a single agent. This approach has a major deficiency as the local shortest path can arbitrarily underestimate the true shortest path cost in the global problem. As a solution, we propose a distributed version of a state-of-the-art LMCut heuristic. We show that our distributed algorithm provides estimates provably equal to estimates of the centralized version computed on the global problem. We also evaluate the algorithm experimentally and show that on a number of domains, the distributed algorithm can significantly improve performance of a multiagent planner.