Inferring Variable Conflicts for Local Search

For efficiency reasons, neighbourhoods in local search are often shrunk by only considering moves modifying variables that actually contribute to the overall penalty. These are known as conflicting variables. We propose a new definition for measuring the conflict of a variable in a model and apply it to the set variables of models expressed in existential second-order logic extended with counting (∃SOL). Such a variable conflict can be automatically and incrementally evaluated. Furthermore, this measure is lower-bounded by an intuitive conflict measure, and upper-bounded by the penalty of the model. We also demonstrate the usefulness of the approach by replacing a built-in global constraint by an ∃SOL version thereof, while still obtaining competitive results.