MiniSAT with Classification-based Preprocessing

We present a classification-based approach to selecting preprocessors of CNF formulas for the complete solver MiniSAT. Three different preprocessors are considered prior to running MiniSAT. To obtain training data for classification, each preprocessor is run, followed by running MiniSAT on its resulting CNF, on instances from the last competition. On each instance, the preprocessor leading to the smallest solving time is considered the “best” and used as the classification for the instance. A decision tree is trained to select the best preprocessor according to features computed from the CNF formulas. The decision tree can therefore be used to predict the best preprocessor for new problem instances.