A Generalized Framework for Lifelong Planning A*

Maxim Likhachev* and Sven Koenig**

*Carnegie Mellon University, **University of Southern California

 

Abstract

Recently, it has been suggested that Lifelong Planning A* (LPA*), an incremental version of A*, might be a good heuristic search-based replanning method for HSP-type planners. LPA* uses consistent heuristics and breaks ties among states with the same f-values in favor of states with smaller g-values. However, HSP-type planners use inconsistent heuristics to trade off plan-execution cost and planning time. In this paper, we therefore develop a general framework that allows one to develop more capable versions of LPA* and its nondeterministic version Minimax LPA*, including versions of LPA* and Minimax LPA* that use inconsistent heuristics and break ties among states with the same f-values in favor of states with larger g-values. We then show experimentally that the new versions of LPA* indeed speed it up on grids and thus promise to provide a good foundation for building heuristic search-based replanners.