Focused crawling aims at collecting as many Web pages relevant to a target topic as possible while avoiding irrelevant pages, reflecting limited resources available to a Web crawler. We improve on the efficiency of focused crawling by proposing an approach based on reinforcement learning. Our algorithm evaluates hyperlinks most profitable to follow over the long run, and selects the most promising link based on this estimation. To properly model the crawling environment as a Markov decision process, we propose new representations of states and actions considering both content information and the link structure. The size of the state-action space is reduced by a generalization process. Based on this generalization, we use a linear-function approximation to update value functions. We investigate the trade-off between synchronous and asynchronous methods. In experiments, we compare the performance of a crawling task with and without learning; crawlers based on reinforcement learning show better performance for various target topics.