Learning Operational Decisions with Intertemporal Dependence and Moderate Non-stationarities

This study investigates the machine learning tools that are employed to learn data-to-decision mappings. A data-to-decision mapping, once learnt, adopts relevant operational information as input and effectively outputs a reliable decision variable. The operational data used by the machine learning tools in numerous operational decision-making environments are sequentially generated from real operations, and therefore, exhibit an inter-temporal dependence structure. Standard machine learning tools and their associated performance guarantees frequently rely on the assumption that the data are independent and identically distributed (i.i.d.), which can be easily violated for operational data. We develop performance guarantees in the form of out-of-sample generalization bounds for machine learning tools under comparatively more realistic assumptions accommodating non-i.i.d. phenomena, such as inter-temporal dependence and moderate non-stationarities. Our findings yield reliable performance guarantees when machine learning tools are used in an operational decision-making environment