SAT-Race 2010 Solver Description: borg-sat-10.06.07

Introduction Algorithm portfolio methods (Huberman, Lukose, and Hogg 1997) use information about solvers and problem instances to allocate computational resources among multiple solvers, attempting to maximize the time spent on those well suited to each instance. Portfolio methods such as SATzilla (Xu et al. 2008) have proved increasingly effective in satisfiability. An algorithm portfolio must decide which solvers to run and for how long to run them. These decisions rely entirely on expectations about solver behavior. The borg-sat solver attempts to to learn predictable aspects of solver behavior—such as how likely a solver is to succeed if it has previously failed—given data on the successes and failures of solvers on many problem instances. The version of this solver submitted to SAT-Race 2010, borg-sat-10.06.07, assumes a specific latent class model of solver behavior, a mixture of Dirichlet compound multinomial (DCM) distributions, which is used to identify groups of similar problem instances. This model is examined in detail by Silverthorn and Miikkulainen (2010). It captures the basic correlations between solvers, runs, and problem instances, as well as the tendency of solver outcomes to recur. Unlike the classifier employed by SATzilla, the model considers only the success or failure of each past solver run; it does not consider instance feature information. This version of borg-sat employs the DCM mixture model in computing an optimal fixed-length solver execution schedule followed for every problem instance, as described in the following section.