Improving SAT Solving Using Monte Carlo Tree Search-Based Clause Learning

Abstract

Most modern state-of-the-art Boolean Satisfiability (SAT) solvers are based on the Davis–Putnam–Logemann–Loveland algorithm and exploit techniques like unit propagation and Conflict-Driven Clause Learning (CDCL). Even though this approach proved to be successful in practice, the success of the Monte Carlo Tree Search (MCTS) algorithm in other domains led to research in using it to solve SAT problems. A recent approach extended the attempt to solve SAT using an MCTS-based algorithm by adding CDCL. We further analyzed the behavior of that approach by using visualizations of the produced search trees and based on the analysis introduced multiple heuristics improving different aspects of the quality of the learned clauses. While the resulting solver can be used to solve SAT on its own, the real strength lies in the learned clauses. By using the MCTS solver as a preprocessor, the learned clauses can be used to improve the performance of existing SAT solvers.