A Reinforcement Learning Approach to Interval Constraint Propagation

When solving systems of nonlinear equations with interval constraint methods, it has often been observed that many calls to contracting operators do not participate actively to the reduction of the search space. Attempts to statically select a subset of efficient contracting operators fail to offer reliable performance speed-ups. By embedding the recency-weighted average Reinforcement Learning method into a constraint propagation algorithm to dynamically learn the best operators, we show that it is possible to obtain robust algorithms with reliable performances on a range of sparse problems. Using a simple heuristic to compute initial weights, we also achieve significant performance speed-ups for dense problems.