Reinforcement Learning (436-0-20)

Instructors

Bradly Stadie

Meeting Info

2001 Sheridan GPC 2447: Mon, Wed 5:00PM - 6:20PM

Overview of class

Reinforcement Learning (RL) is the problem of learning to align agent behavior with human desires. This alignment is done with the introduction of a reward signal, which tells the agent the desirability of its current state. The agent then interacts with its environments, looking for actions that can maximally align its state with the provided reward. RL appears most naturally in the context of robotics, but also has applications to industrial design, biology, medicine, optimization, and machine learning. Recently, deep RL has been a crucial component in large language models (LLMS), where reinforcement learning from human feedback allows us to better align LLMs with human sensibilities. The key algorithms in RL are policy gradients and TD methods such as Q-learning. We will cover both classes of methods extensively in the class.

Registration Requirements

Linear algebra (MATH 240-0), Stat 415-0 (Introduction to Machine Learning) or similar

Learning Objectives

Students will be able to read current literature in reinforcement learning, and will understand how to code basic RL algorithms in Python.

Teaching Method

Lecture

Evaluation Method

40% Three Assignments
25% Midterm
5% Final Project Proposal
30% Final Project