The most celebrated corners of machine learning over the past decades are those successful at predicting - e.g., spam classification, medical diagnoses, or cat faces. But machine learning as actually used in practice is commonly prescriptive rather than predictive: decisions must be made in order to maximize a reward. The misuse of predictive approaches for prescriptive policy needs is as old as multivariate regression itself. Such problems are common in health, commerce, and engineering. These problems broadly fall under the umbrella of reinforcement learning. I will motivate and illustrate some of these applications of reinforcement learning, then show how methods from statistical physics, particularly variational and Monte Carlo methods, can be used to extend and improve modeling approaches.