Hardness-Aware Restart Policies

Recent work has demonstrated that it is possible to boost the efficiency of combinatorial search procedures via the use of principled restart policies. We present a coupling of machine learning and dynamic programming that extends prior efforts by endowing restart policies with knowledge of the hardness of the specific instance being solved. This ability allows a restart policy to take into consideration an updated probability distribution over hardness as a previously unseen instance is being solved. We discuss the methods, highlighting their importance for real-world applications of combinatorial search. Finally, we present the empirical results.