Probabilistic Loss Functions for Self-Supervised SAT Solvers

Nikolaos Efthymiou, Nikolaos Karalias

Keywords: Neural Combinatorial Optimization, Unsupervised Learning, Boolean Satisfiability

Abstract: In this study, we design novel probabilistic loss functions for training Graph Neural Networks in an unsupervised way to tackle the CNF-SAT problem, which is an important NP-complete problem. In particular, we investigate the power of the Lovász Local Lemma (LLL) in obtaining satisfiability certificates in a differentiable manner. Given that the LLL provides provable discretization procedures, such as the Moser-Tardos algorithm, our approach offers an end-to-end hybrid SAT solver.

Submission Number: 50