In machine learning, in parallel to algorithms themselves, the representation of data is a point of utmost importance. Efforts on data pre-processing in general are a key ingredient to success. An algorithm that performs poorly on a particular form of given data may perform much better, both in terms of efficiency and the quality of the solution, when the same data is represented in another form. Despite the amount of literature on the subject, the issue of how to enrich a representation to suit the underlying mechanism is clearly still pending. In this paper, we approach this problem within the context of reinforcement learning, and in particular, interested in discovery of a ``good'' representation of data for the LSPI algorithm. To this end, we use the cascade-correlation learning architecture to automatically generate a set of basis functions which would lead to a better approximation of the value function, and consequently improve the performance of the resulting policies. We also show the effectiveness of the idea on some benchmark problems.