Reinforcement Learning and Reactive Search: an adaptive MAX-SAT solver

This paper investigates Reinforcement Learning (RL) applied to online parameter tuning in Stochastic Local Search (SLS) methods. In particular, a novel application of RL is proposed in the Reactive Tabu Search (RTS) scheme, where the appropriate amount of diversification in prohibition-based local search is adapted in a fast online manner to the characteristics of a task and of the local configuration. The experimental tests demonstrate promising results on Maximum Satisfiability (MAX-SAT) instances when compared with state-of-the-art SLS SAT solvers, such us AdaptNovelty, rSAPS and gNovelty.