New CNF Features and Formula Classification

Abstract

In this paper we first present three new features for classifying CNF formulas. These features are based on the structural information of the formula and consider AND-gates as well as exactly-one constraints. Next, we use these features to construct a machine learning approach to select a SAT solver configuration for CNF formulas with random decision forests. Based on this classification task we can show that our new features are useful compared to existing features. Since the computation time for these features is small, the constructed classifier improves the performance of the SAT solvers on application and hand crafted benchmarks. On the other hand, the comparison shows that the set of new features also results in a better classification.