ABSTRACT
When faced with a decision problem, it is customary to formulate a model, learn parameter values from data, and then to optimize decisions based on the fitted model. The separate treatment of learning and optimization is appropriate if the formulated model accurately captures the generating process. On the other hand, if the model is misspecified, significant gains can be realized through better coordination of learning and optimization. I will discuss directed learning, an approach to learning that results in improved decisions when models are misspecified, and provide an example involving the learning of a forecasting model for use in dynamic programming.
