We address the problem of selecting the best configuration out of a set of configurations for DaoOpt - an AND/OR Branch-and-Bound search based solver for combinatorial optimization problems expressed as MPE queries over graphical models. DaoOpt takes different parameters for solving a problem, the running time of a problem depends on the configuration chosen for it. We want to learn for a given problem instance which configuration is the fastest to solve it and predict the configuration likely to solve a new problem in the least amount of time. The results indicate that our predictor is able to improve the running time of the problems.