With reinforcement learning the learning algorithm is not given examples of optimal outputs, in contrast to supervised learning, but must instead discover them by a process of trial and error.
Christopher M. Bishop - «Pattern Recognition and Machine Learning» (2006)
In order to compare reinforcement learning with supervised learning, let's think about an agent learning to play chess. In the supervised setting, the designer has to provide the correct label on a subset of situations . Imagine what a pain in the neck it is to "give" the correct action in many situations to the agent. In fact, as a novice human player, I am unable to provide the best action in many situation. In the reinforcement learning problem, however, the designer is only expected to provide a reward signal . In the case of chess, it really is trivial: +1 for winning the game, -1 for loosing the game, and 0 otherwise. It would then be the agent's job to assign credit to actions that led to the agent winning the game or the actions that screwed it up!