Predicting Tactical Solutions to Operational Planning Problems Under Imperfect Information

This paper offers a methodological contribution at the intersection of machine learning and operations research. Namely, we propose methodology to quickly predict expected tactical descriptions operational solutions (TDOSs). The problem address occurs in context two-stage stochastic programming where second stage is demanding computationally. We aim high speed TDOS associated with problem, conditionally on first variables. may be used support solution overall by avoiding online generation multiple scenarios solutions. formulate prediction as optimal program, whose approximate supervised learning. training dataset consists large number deterministic problems generated controlled probabilistic sampling. labels are computed based these (solved independently offline), employing appropriate aggregation subselection methods uncertainty. Results our motivating application load planning for rail transportation show that deep models produce accurate predictions very short computing time (milliseconds or less). predictive accuracy close lower bounds calculated sample average approximation programs.