On-Demand Bound Computation for Best-First Constraint Optimization

Many problems in Artificial Intelligence, such as diagnosis, control, and planning, can be framed as constraint optimization problems where a limited number of leading solutions are needed. An important class of optimization algorithms use A*, a variant of best first search, to guide the search effectively, while employing a heuristic evaluation function that is computed using dynamic programming. A key bottleneck, however, is that significant effort can be wasted precomputing bounds that are not used to generate the leading solutions. This paper introduces a method for solving semi-ring CSPs, based on a variant of A* that generates on demand, only those bounds that are specifically required in order to generate a next best solution. On demand bound computation is performed using “lazy”, best-first variants of constraint projection and combination operators, and a scheme that coordinates the computation of these operators by exploiting a decomposition of the optimization problem into a tree structure. We demonstrate a significant performance improvement over bound precomputation on randomly generated, semiring-based optimization problems.