Jasper: the Art of Exploration in Greedy Best First Search

LAMA-2011 (Richter and Westphal 2010) is the clear winner of the sequential satisficing track in the latest International Planning Competition (IPC-2011). It finds a first solution by Greedy Best-First Search (GBFS), and then continues to improve solutions using restarting weighted A* (Richter, Thayer, and Ruml 2010). Diverse Anytime Search (DAS) (Xie, Valenzano, and Müller 2013) is a metaalgorithm designed for solution improvement. It takes an anytime planner and a post-processing system, and adds restarts and randomization for better quality search. Jasper is a satisficing planner that builds on LAMA-2011. It adds two modifications. First, it replaces the GBFS algorithm in LAMA-2011 with an improved GBFS variant, called Type Exploration based Greedy Best-First Search with Local Search (Type-GBFS-LS). GBFS always expands a node n that is closest to a goal state according to a heuristic h. GBFS’ performance strongly depends on h. Uninformative or misleading heuristics can massively increase the time and memory complexity of such searches. Type-GBFS-LS is an improved version of GBFS that is less sensitive to such flaws in heuristic functions. Second, it implements the DAS system for solution improvement, which takes the modified LAMA-2011 as the anytime planner and Aras (Nakhost and Müller 2010) as the post-processing system. A detailed description of the implementation of DAS can be found in the ICAPS paper by Xie, Valenzano and Müller (Xie, Valenzano, and Müller 2013). This paper focuses on describing the new search algorithm, Type-GBFS-LS. The remainder of this paper is organized as follows. First, we motivate this work by discussing the two potential problems of GBFS: uninformative heuristic region and misleading heuristics, followed by describing two corresponding solutions as well as their combination, Type-GBFS-LS. Later, experimental results show that the proposed algorithms improve the state of the art planner LAMA-2011 significantly.