Generalization in Reinforcement Learning: Safely Approximating the Value Function