Residual Algorithms: Reinforcement Learning with Function Approximation