Generalized Kalman filter based temporal difference reinforcement learning