As part of the colloquium of the Research Training Group Algorithmic Optimization will take place on

Monday, February 26, 2024
16:00 c.t.
Lecture hall 9

the following lecture will take place:

The Benders-by-batch algorithm to solve two-stage stochastic linear programs
Boris Detienne, University of Bordeaux

In this talk, we will introduce a new exact algorithm to solve two-stage stochastic linear programs. Based on the multicut Benders reformulation of such problems, with one subproblem for each scenario, this method relies on a partition of the subproblems into batches. The key idea is to solve at most iterations only a small proportion of the subproblems by detecting as soon as possible that a first-stage candidate solution cannot be proven optimal. We also propose a general framework to stabilize our algorithm, and show its finite convergence and exact behavior. We report an extensive computational study on large-scale instances of stochastic optimization literature that shows the efficiency of the proposed algorithm compared to nine alternative algorithms from the literature. We also obtain significant additional computational time savings using the primal stabilization schemes.