REINFORCEMENT LEARNING AND STOCHASTIC OPTIMIZATION: A UNIFIED FRAMEWORK FOR SEQUENTIAL DECISIONS