Optimal Scenario-Tree Selection for Multistage Stochastic Programming

摘要

We propose an algorithmic strategy for Multistage Stochastic Optimization, to learn a decision policy able to provide feasible and optimal decisions for every possible value of the random variables of the problem. The decision policy is built using a scenario-tree based solution combined with a regression model able to provide a decision also for those scenarios not included in the tree. For building an optimal policy, an iterative scenario generation procedure is used which selects through a Bayesian Optimization process the more informative scenario-tree. Some preliminary numerical tests show the validity of such an approach.