Dynamic optimization with side information

We develop a tractable and flexible data-driven approach for incorporating side information into multi-stage stochastic programming. The proposed framework uses predictive machine learning methods (such as k-nearest neighbors, kernel regression, random forests) to weight the relative importance of various uncertainty sets in robust optimization formulation. Through novel measure concentration result class supervised methods, we prove that is asymptotically optimal multi-period programming with information. also describe general-purpose approximation these problems, based on overlapping linear decision rules, which computationally produces high-quality solutions dynamic problems many stages. Across variety single-stage examples inventory management, finance, shipment planning, our method achieves improvements up 15% over alternatives requires less than one minute computation time twelve