Sample-based Learning Methods

This week you will learn how to estimate value functions and optimal policies, using only sampled experience from the environment. This module represents our first step toward incremental learning methods that learn from the agent’s own interaction with the world, rather than a model of the world. You will learn about on-policy and off-policy methods for prediction and control, using Monte Carlo methods---methods that use sampled returns. You will also be reintroduced to the exploration problem, but more generally in RL, beyond bandits.

This week, you will learn about one of the most fundamental concepts in reinforcement learning: temporal difference (TD) learning. TD learning combines some of the features of both Monte Carlo and Dynamic Programming (DP) methods. TD methods are similar to Monte Carlo methods in that they can learn from the agent’s interaction with the world, and do not require knowledge of the model. TD methods are similar to DP methods in that they bootstrap, and thus can learn online---no waiting until the end of an episode. You will see how TD can learn more efficiently than Monte Carlo, due to bootstrapping. For this module, we first focus on TD for prediction, and discuss TD for control in the next module. This week, you will implement TD to estimate the value function for a fixed policy, in a simulated domain.

This week, you will learn about using temporal difference learning for control, as a generalized policy iteration strategy. You will see three different algorithms based on bootstrapping and Bellman equations for control: Sarsa, Q-learning and Expected Sarsa. You will see some of the differences between the methods for on-policy and off-policy control, and that Expected Sarsa is a unified algorithm for both. You will implement Expected Sarsa and Q-learning, on Cliff World.