Can machine learning learn a decision oracle for NP problems? A test on SAT

This note describes our experiments aiming to empirically test the ability of machine learning models to act as decision oracles for NP problems. Focusing on satisfiability testing problems, we have generated random 3-SAT instances and found out that the correct branch prediction accuracy reached levels in excess of 99%. The branching in a simple backtracking-based SAT solver has been reduced in more than 90% of the tested cases, and the average number of branching steps has reduced to between 1/5 and 1/3 of the one without the machine learning model. The percentage of SAT instances where the machine learned heuristic-enhanced algorithm solved SAT in a single pass reached levels of 80-90%, depending on the set of features used.