The goal of reinforcement learning is to find a policy, directly or indirectly through a value function, that maximizes the expected re- ward accumulated by an agent over time based on its interactions with the environment; a function of the state has to be learned. In some prob- lems, it may not be feasible, or even possible, to use the state variables as they are. Instead, a set of features are computed and used as in- put. However, finding a "good" set of features is generally a tedious task which requires a good domain knowledge. In this paper, we propose a ge- netic programming based approach for feature discovery in reinforcement learning. A population of individuals each representing possibly different number of candidate features is evolved, and feature sets are evaluated by their average performance on short learning trials. The results of ex- periments conducted on several benchmark problems demonstrate that the resulting features allow the agent to learn better policies.