Exploration Among and Within Plateaus in Greedy Best-First Search

Recent enhancements to greedy best-first search (GBFS) such as DBFS, epsilon-GBFS, Type-GBFS improve performance by occasionally introducing exploratory behavior which occasionally expands non-greedy nodes. However, most of these exploratory mechanisms do not address exploration within the space sharing the same heuristic estimate (plateau). In this paper, we show these two modes of exploration, which work across (inter-) and within (intra-) plateau, are complementary, and can be combined to yield superior performance. We then introduces a new fractal-inspired scheme called Invasion-Percolation diversification, which addresses "breadth"-bias instead of the "depth"-bias addressed by the existing diversification methods. We evaluate IP-diversification for both intra- and inter-plateau exploration, and show that it significantly improves performance in several domains. Finally, we show that combining diversification methods results in a planner which is competitive to the state-of-the-art for satisficing planning.