Structure features for SAT instances classification

The success of portfolio approaches in SAT solving relies on the observation that different SAT solvers may dramatically change their performance depending on the class of SAT instances they are trying to solve. In these approaches, a set of features of the problem is used to build a prediction model, which classifies instances into classes, and computes the fastest algorithm to solve each of them. Therefore, the set of features used to build these classifiers plays a crucial role. Traditionally, portfolio SAT solvers include features about the structure of the problem and its hardness. Recently, there have been some attempts to better characterize the structure of industrial SAT instances. In this paper, we use some structure features of industrial SAT instances to build some classifiers of industrial SAT families of instances. Namely, they are the scale-free structure, the community structure and the self similar structure. First, we measure the effectiveness of these classifiers by comparing them to other sets of SAT features commonly used in portfolio SAT solving approaches. Then, we evaluate the performance of this set of structure features when used in a real portfolio SAT solver. Finally, we analyze the relevance of these features on the analyzed classifiers.
QC 20170821