Iterative state-space reduction for flexible computation

Abstract

Flexible computation is a general framework for decision making under limited computational resources. It enables an agent to allocate limited computational resources to maximize its overall performance or utility. In this paper, we present a strategy for flexible computation, which we call iterative state-space reduction. The main ideas are to reduce a problem space that is difficult to search to one that is relatively easy to explore, to use the optimal solution from the reduced space as an approximate solution to the original problem, and to iteratively apply multiple reductions to progressively find better solutions. The basic operation for state-space reduction is heuristic node pruning, which excludes the nodes that do not seem to lead to high-quality solutions from further examination. Iterative state-space reduction can be combined with a state-space search, such as best-first search or depth-first branch-and-bound (DFBnB). The resulting algorithm can run as an anytime algorithm, which provides a tradeoff between solution quality and computation. Based on an analytical modal, we analyze the probability that one iteration of the iterative process finds a solution. Furthermore, by combining it within DFBnB we apply iterative state-space reduction to three combinatorial problems, the maximum Boolean satisfiability, the symmetric TSP and asymmetric TSP. Our experimental results show that iterative state-space reduction is effective and efficient, making DFBnB find better solutions with less computation.