P.: On universal restart strategies for backtracking search

Abstract. Constraint satisfaction and propositional satisfiability problems are often solved using backtracking search. Previous studies have shown that a technique called randomization and restarts can dramatically improve the performance of a backtracking algorithm on some instances. We consider the commonly occurring scenario where one is to solve an ensemble of instances using a backtracking algorithm and wish to learn a good restart strategy for the ensemble. In contrast to much previous work, our focus is on universal strategies. We contribute to the theoretical understanding of universal strategies and demonstrate both analytically and empirically the pitfalls of non-universal strategies. We also propose a simple approach for learning good universal restart strategies and demonstrate the effectiveness and robustness of our approach throughanextensiveempiricalevaluationonareal-worldtestbed.