Understanding and Improving Local Exploration for GBFS

Greedy Best First Search (GBFS) is a powerful algorithm at the heart of many state-of-the-art satisficing planners. The Greedy Best First Search with Local Search (GBFS-LS) algorithm adds exploration using a local GBFS to a global GBFS. This substantially improves performance for domains that contain large uninformative heuristic regions (UHR), such as plateaus or local minima. This paper analyzes, quantifies and improves the performance of GBFS-LS. Planning problems with a mix of small and large UHRs are shown to be hard for GBFS but easy for GBFS-LS. In three standard IPC planning instances analyzed in detail, adding exploration using local GBFS gives more than three orders of magnitude speedup. As a second contribution, the detailed analysis leads to an improved GBFSLS algorithm, which replaces larger-scale local GBFS explorations with a greater number of smaller explorations.