Reinforcement Learning

3 Credits · Level: Graduate

The field of machine learning is typically divided into three fundamental sub-paradigms. These include supervised learning, unsupervised learning, and reinforcement learning (RL). The discipline of reinforcement learning focuses on how intelligent agents learn to perform actions, inside a specified environment, to maximize a cumulative reward function. Over the past several decades, there has been a push to incorporate concepts from the field of deep-learning into the agents used in RL algorithms. This has spawned the field of deep reinforcement learning. To date, the field of deep RL has yielded stunning results in a wide range of technological applications. These include, but are not limited to, self-driving cars, autonomous game play, robotics, trading and finance, and Natural Language Processing. This course will begin with an introduction to the fundamentals of traditional, i.e. non-deep, reinforcement learning. After reviewing fundamental deep learning topics the course will transition to deep RL by incorporating artificial neural networks into the models. The course includes a coding emphasis to showcase applied implementations of RL within the python ecosystem. Topics include multi-armed bandits, contextual bandits, function approximation via neural networks, Markov decision processes, Monte Carlo methods, temporal difference learning, deep Q-learning, actor-critic methods, and policy gradient methods. Students must have intermediate coding experience in Python as well as knowledge of introductory statistics and multivariable calculus.