Band Search: An Efficient Alternative to Guided Depth-First Search

In this paper, we propose a novel search algorithm called band search that generalizes guided depth-first and best-first searches. The search allocates a band of at most W nodes in each level of the search tree for storing active nodes in that level, D priority lists, one for each level, for storing overflow nodes, and D counters, one for each level, for keeping track of and to limit the degree of backtracking allowed. The algorithm has three major features: a) it selects for expansion in a best-first fashion from all nodes in the bands, b) it moves nodes from an overflow list into the corresponding band in a depth-first fashion, and c) it restricts backtracking so that at most W nodes in a level are fully searched before allowing new nodes in this level into the band. The algorithm specializes to be a guided depth-first search (GDFS) when W is one, and a best-first search (BFS) when W is unlimited. By choosing W appropriately, we show empirically that the algorithm can often outperform GDFS and has performance close to that of BFS. Moreover, the algorithm can be used instead of GDFS in iterative searches, such as IDA*, MIDA*, and DFS*. Finally, we identify conditions upon which the algorithm behaves like BFS, and the anomalous behavior when the algorithm performs worse than GDFS.