The goal of reinforcement learning is to find a policy that maximizes the expected reward accumulated by an agent over time based on its interactions with the environment; to this end, a function of the state of the agent has to be learned. It is often the case that states are better characterized by a set of features. However, finding a ''good'' set of features is generally a tedious task which requires a good domain knowledge. In this paper, we propose a genetic programming based approach for feature discovery in reinforcement learning. A population of individuals, each representing a set of features is evolved, and individuals are evaluated by their average performance on short reinforcement learning trials. The results of experiments conducted on several benchmark problems demonstrate that the resulting features allow the agent to learn better policies in a reduced amount of episodes.