Escaping Local Optima: Constraint Weights vs. Value Penalties

Constraint Satisfaction Problems can be solved using either iterative improvement or constructive search approaches. Iterative improvement techniques converge quicker than the constructive search techniques on large problems, but they have a propensity to converge to local optima. Therefore, a key research topic on iterative improvement search is the development of effective techniques for escaping local optima, most of which are based on increasing the weights attached to violated constraints. An alternative approach is to attach penalties to the individual variable values participating in a constraint violation. We compare both approaches and show that the penalty-based technique has a more dramatic effect on the cost landscape, leading to a higher ability to escape local optima. We present an improved version of an existing penalty-based algorithm where penalty resets are driven by the amount of distortion to the cost landscape caused by penalties. We compare this algorithm with an algorithm based on constraint weights and justify the difference in their performance.