Reinforcement Learning - 705.741

This course will focus on both the theoretical and the practical aspects of designing, training, and testing reinforcement learning systems. The course begins with an examination of Markov decision processes (MDPs), which provide a sound mathematical basis for modeling and solving complex sequential decision problems. The more traditional analytical method for solving MDPs, dynamic programming, will be reviewed. We will then examine the major reinforcement learning approaches, such as Monte Carlo methods, temporal difference methods, policy gradient methods, and deep learning methods, comparing them as appropriate to dynamic programming techniques. Fundamental issues and limitations on the performance of reinforcement learning algorithms (e.g., the credit assignment problem, the exploration / exploitation tradeoff, on-policy learning versus off-policy learning, partial observability, and algorithm convergence properties) will be examined for each approach. Weekly exercises and discussion topics will reinforce and expand on the classroom material. In addition, students will gain practical experience during a semester-long project by programming, training, and testing various reinforcement learning algorithms.

Course Prerequisite(s)

EN.625.638/EN.605.647 – Neural Networks or experience programming artificial neural networks in a high-level language.