Scenario grouping in a progressive hedging-based meta-heuristic for stochastic network design

We propose a methodological approach to build strategies for grouping scenarios as defined by the type of scenario decomposition, type of grouping, and the measures specifying scenario similarity. We evaluate these strategies in the context of stochastic network design by analyzing the behavior and performance of a new progressive hedging-based meta-heuristic for stochastic network design that solves subproblems comprising multiple scenarios. We compare the proposed strategies not only among themselves, but also against the strategy of grouping scenarios randomly and the lower bound provided by a state-of-the-art MIP solver. The results show that, by solving multi-scenario subproblems generated by the strategies we propose, the meta-heuristic produces better results in terms of solution quality and computing efficiency than when either single-scenario subproblems or multiple-scenario subproblems that are generated by picking scenarios at random are solved. The results also show that, considering all the strategies tested, the covering strategy with respect to commodity demands leads to the highest quality solutions and the quickest convergence.