Heuristic Search with Limited Memory By

HEURISTIC SEARCH WITH LIMITED MEMORY by Matthew Hatem University of New Hampshire, May, 2014 Heuristic search algorithms are commonly used for solving problems in artificial intelligence. Unfortunately, the memory requirement of A*, the most widely used heuristic search algorithm, is often proportional to its running time, making it impractical for large problems. Several techniques exist for scaling heuristic search: external memory, bounded suboptimal search, and linear-space algorithms. I address limitations in each. The thesis of this dissertation is that in order to improve scalability, memory efficient heuristic search algorithms benefit from the same techniques used in unbounded space search: best-first search order, partial expansion, bounded node generation overhead, and distance-to-go estimates. The four contributions of this dissertation make it easier to apply heuristic search to challenging problems with practical relevance. First I address limitations in external memory search with a technique for bounding overhead on problems that have real-valued costs and a another technique for reducing overhead on problems that have large branching factors. I demonstrate that these techniques achieve a new state-of-the-art on the problem of multiple sequence alignment. Second, I examine recent work in bounded suboptimal search and present a new technique for simplifying implementation and reducing run-time overhead. The third contribution addresses limitations of linear-space search with a new technique for provably bounding node regeneration overhead. Finally, I present four new algorithms that advance the state of the art in linear space bounded suboptimal search. These advances support the conclusion that search under memory limiations benefits from techniques similar to those in unbounded space search.