We present a computational study of several strategies to solve two-stage stochastic linear programs by integrating the adaptive partition-based approach with level decomposition. A partition-based formulation is a relaxation of the original stochastic program, obtained by aggregating variables and constraints according to a scenario partition. Partition refinements are guided by the optimal second-stage dual vectors computed at certain first-stage solutions. The proposed approaches rely on the level decomposition with on-demand accuracy to dynamically adjust partitions until an optimal solution is found. Numerical experiments on a large set of test problems including instances with up to one hundred thousand scenarios show the effectiveness of the proposed approaches.
Citation
Submitted for publication