Informed Backtracking Beam Search

Beam searches are a very effective algorithm for solving problems that prove intractable to complete algorithms like weighted A* and greedy search. Unfortunately, the inadmissible pruning that is the hallmark of a beam search makes the algorithm incomplete. Existing complete extensions to beam search expand pruned nodes systematically according to generation order. Selecting pruned nodes to expand based upon generation order often causes the expansion of the most promising nodes to be delayed. This can cause the complete backtracking beam searches to expend unnecessary effort exploring less promising parts of the search space leading to delayed solutions, lower quality solutions, or both. In this paper, we show that considering heuristic information when deciding where to backtrack can significantly improve the performance of backtracking beam searches. Although the beam searches using informed backtracking outperform the other kinds of restarting beam searches, none of the backtracking beam search algorithms are competitive with weighted A* and other best-first searches.