This letter addresses the issue of learning shortest paths in complex networks, which is of utmost importance in real-life navigation. The approach has been partially motivated by recent progress in characterizing navigation problems in networks, having as extreme situations the completely ignorant (random) walker and the rich directed walker, which can pay for information that will guide to the target node along the shortest path. A learning framework based on a first-visit Monte Carlo algorithm is implemented, together with four independent measures that characterize the learning process. The methodology is applied to a number of network classes, as well as to networks constructed from actual data. The results indicate that the navigation difficulty and learning velocity are strongly related to the network topology.