A parallel Branch-and-Fix Coordination based matheuristic algorithm for solving large sized multistage stochastic mixed 0–1 problems

Abstract

A parallel matheuristic algorithm is presented as a spin-off from the exact Branch-and-Fix Coordination (BFC) algorithm for solving multistage stochastic mixed 0–1 problems. Some steps to guarantee the solution’s optimality are relaxed in the BFC algorithm, such that an incomplete backward branching scheme is considered for solving large sized problems. Additionally, a new branching criterion is considered, based on dynamically-guided and stage-wise ordering schemes, such that fewer Twin Node Families are expected to be visited during the execution of the so-called H-DBFC algorithm. The inner parallelization IH-DBFC of the new approach, allows to solve in parallel scenario clusters MIP submodels at different steps of the algorithm. The outer parallel version, OH-DBFC, considers independent paths and allows iterative incumbent solution values exchanges to obtain tighter bounds of the solution value of the original problem. A broad computational experience is reported for assessing the quality of the matheuristic solution for large sized instances. The instances dimensions that are considered are up to two orders of magnitude larger than in some other works that we are aware of. The optimality gap of the H-DBFC solution value versus the one obtained by a state-of-the-art MIP solver is very small, if any. The new approach frequently outperforms it in terms of solution’s quality and computing time. A comparison with our Stochastic Dynamic Programming algorithm is also reported. The use of parallel computing provides, on one hand, a perspective for solving very large sized instances and, on the other hand, an expected large reduction in elapsed time.