One of the project I done is to analyze the John Moody's "recurrent reinforcement learning for FX trading" in summer 06. You may take a look at that to have some feel.