Stochastic Node Caching for Memory-Bounded Search

With linear-storage search, the same nodes are eventually revisited many times because only the search paths are stored to memory. Some algorithms such as MREC have been proposed to solve this problem by storing nodes as well. MREC is an algorithm that reduces the number of nodes revisited by storing a certain number of nodes located near the root node. Proposed in this paper is stochastic node caching, which involves storing nodes on a probability basis. In so doing, only those nodes are stored that are visited frequently, so that the number of nodes revisited can be reduced efficiently while using limited memory re-sources. To prove the efficiency of stochastic node caching, this method was compared with MREC while pursuing the same goal. Experiments were performed using the 15-puzzle problem, a typical problem for linear-storage search, and a more complicated problem of gene alignment. The experiments illustrated the properties of the two algorithms, and proved that stochastic node caching is efficient in reducing the number of nodes revisited.