This course is an introduction to sequential decision making and reinforcement learning. We start with a discussion of utility theory to learn how preferences can be represented and modeled for decision making. We first model simple decision problems as multi-armed bandit problems in and discuss several approaches to evaluate feedback. We will then model decision problems as finite Markov decision processes (MDPs), and discuss their solutions via dynamic programming algorithms. We touch on the notion of partial observability in real problems, modeled by POMDPs and then solved by online planning methods. Finally, we introduce the reinforcement learning problem and discuss two paradigms: Monte Carlo methods and temporal difference learning. We conclude the course by noting how the two paradigms lie on a spectrum of n-step temporal difference methods. An emphasis on algorithms and examples will be a key part of this course.
Overview
Syllabus
- Decision Making and Utility Theory
- Welcome to Decision Making and Reinforcement Learning! During this week, Professor Tony Dear provides an overview of the course. You will also view guidelines to support your learning journey towards modeling sequential decision problems and implementing reinforcement learning algorithms.
- Bandit Problems
- Welcome to week 2! This week, we will learn about multi-armed bandit problems, a type of optimization problem in which the algorithm balances exploration and exploitation to maximize rewards. Topics include action values and sample averaging estimation, 𝜀-greedy action selection, and the upper confidence bound. You could post in the discussion forum if you need assistance on the quiz and assignment.
- Markov Decision Processes
- Welcome to week 3! This week, we will focus on the basics of the Markov decision process, including rewards, utilities, discounting, policies, value functions, and Bellman equations. You will model sequential decision problems, understand the impact of rewards and discount factors on outcomes, define policies and value functions, and write Bellman equations for optimal solutions. You could post in the discussion forum if you need assistance on the quiz and assignment.
- Dynamic Programming
- Welcome to week 4! This week, we will cover dynamic programming algorithms for solving Markov decision processes (MDPs). Topics include value iteration and policy iteration, nonlinear Bellman equations, complexity and convergence, and a comparison of the two approaches.You could post in the discussion forum if you need assistance on the quiz and assignment.
- Partially Observable Markov Decision Processes
- Welcome to week 5! This week, we will go through topics on partial observability and POMDPs, belief states, representation as belief MDPs, and online planning in MDPs and POMDPs. You will also apply your knowledge to update the belief state and employ a belief transition function to calculate state values. You could post in the discussion forum if you need assistance on the quiz and assignment.
- Monte Carlo Methods
- Welcome to week 6! This week, we will introduce Monte Carlo methods, and cover topics related to state value estimation using sample averaging and Monte Carlo prediction, state-action values and epsilon-greedy policies, and importance sampling for off-policy vs on-policy Monte Carlo control. You will learn to estimate state values, state-action values, use importance sampling, and implement off-policy Monte Carlo control for optimal policy learning. You could post in the discussion forum if you need assistance on the quiz and assignment.
- Temporal-Difference Learning
- Welcome to week 7! This week, we will cover topics related to temporal difference learning for prediction, TD batch methods, SARSA for on-policy control, and Q-learning for off-policy control. You will learn to implement TD prediction, TD batch and offline methods, SARSA and Q-learning, and compare on-policy vs off-policy TD learning. You will then apply your knowledge in solving a Tic-tac-toe programming assignment.You could post in the discussion forum if you need assistance on the quiz and assignment.
- Reinforcement Learning - Generalization
- Welcome to week 8! This module covers n-step temporal difference prediction, n-step SARSA (on-policy and off-policy), model-based RL with Dyna-Q, and function approximation. You will be prepared to implement n-step TD learning, n-step SARSA, Dyna-Q for model-based learning, and use function approximation for reinforcement learning. You will apply your knowledge in the Frozen Lake programming environment. You could post in the discussion forum if you need assistance on the quiz and assignment.
Taught by
Tony Dear
