Learning a Reactive Restart Strategy to Improve Stochastic Search

Building on the recent success of bet-and-run approaches for restarted local search solvers, we introduce the idea of learning online adaptive restart strategies. Universal restart strategies deploy a fixed schedule that runs with utter disregard of the characteristics that each individual run exhibits. Whether a run looks promising or abysmal, it gets run exactly until the predetermined limit is reached. Bet-and-run strategies are at least slightly less ignorant as they decide which trial to use for a long run based on the performance achieved so far. We introduce the idea of learning fully adaptive restart strategies for blackbox solvers, whereby the learning is performed by a parameter tuner. Numerical results show that adaptive strategies can be learned effectively and that these significantly outperform bet-and-run strategies.