CSCI 3386 Reinforcement Learning

Today, computer programs like AlphaZero can teach themselves from scratch in less than a day to play games like chess, go, and shogi better than human experts. How? Reinforcement learning (RL), the branch of machine learning concerned with how a machine can teach itself, through trial and error, to optimize the reward it obtains over time from an uncertain and dynamic environment. RL applies to many problems besides board games, including flying drones, steering cars, landing rockets, puppeteering robots, improving chatbots, folding proteins, containing plasma, trading stocks, and even proving theorems. Students will learn the fundamental ideas and algorithms at the heart of RL in its many incarnations, ranging from dynamic programming to model-free RL. And they will see how recent breakthroughs in deep neural networks have propelled RL forward. Students will learn these things not only in concept, but by creating and experimenting with their own RL agents, including one which teaches itself to play an Atari 2600 video game.