Multi-Armed Bandits for Adaptive Constraint Propagation