Analysis and Generation of Pseudo-Industrial MaxSAT Instances

We propose two random generation models for MaxSAT and Partial MaxSAT in order to produce instances more similar to the industrial benchmarks used in the MaxSAT evaluation. Following the work of [4] and [2], we analyze properties of industrial instances and use a non-uniform (powerlaw) distribution to select the variables. We also study empirically the optimum (minim number of unsatisfiable clauses) that we obtain with these models, and the relative performance of some MaxSAT solvers. We observe that industrial specialized MaxSAT solvers are better on these random formulas than random specialized solvers. We conclude that instances generated with these new models are more similar to industrial instances than the generated with the classical random models based on uniform probability distributions.