Learning is an important aspect of intelligent behavior. Unfortunately, learning rarely comes for free. Techniques developed by machine learning can improve the abilities of an agent but they often entail considerable computational expense. Furthermore, there is an inherent tradeoff between the power and efficiency of learning. More powerfulearning approaches require greater computational resources. This poses a dilemmato a learning agent that must act in the world under a variety of resource constraints. This paper investigates the issues involved in constructing a rational learning agent. Drawing on work in decision-theory we describe a framework for a rational agent that embodies learning actionsthat canmodify its own behavior. The agent must posses deliberative capabilities to assess the relative merits of these actions in the larger context of its overall behavior and resource constraints. We then sketch several algorithms that have been developed within this framework.