Interactions between learning and decision making

Abstract

We quantify the effects of learning and decision making on each other in three parts. In the first part, we look at how knowledge about decision making can influence learning. Let the decision cost be the amount spent by the practitioner in executing a policy. If we have prior knowledge about this cost, for instance that it should be low, then this knowledge can help restrict the hypothesis space for learning, which can help with its generalization. We derive a suite of theoretical generalization bounds and an algorithm for this setting. In the second part, we look at how knowledge about learning can influence decision making. We study this in the context of robust optimization. Taking the uncertainty of learning the right model into account, we derive multiple probabilistic guarantees on the robustness of the resulting policy. In the last part, we explore the interactions between learning and decision making in depth for two applications. The first application is in the area of power grid maintenance and the second is in the area of professional racing. We provide tailored solutions for modeling, predicting and making decisions in each context.