Dissecting “Reinforcement Learning” by Richard S. Sutton with Custom Python Implementations, Episode III
We continue our deep dive into Sutton’s great book about RL [1] and here focus on Monte Carlo (MC) methods. These are able to learn from experience alone, i.e. do not require any kind of model of the environment, as e.g. required by the Dynamic programming (DP) methods we introduced in the previous post.
This is extremely tempting — as often the model is not known, or it is hard to model the transition probabilities. Consider the game of Blackjack: even though we fully understand the game and the rules, solving it via DP methods would be very tedious — we would have to compute all kinds of probabilities, e.g. given the currently played cards, how likely is a “blackjack”, how likely is it that another seven is dealt … Via MC methods, we don’t have to deal with any of this, and simply play and learn from experience.
Due to not using a model, MC methods are unbiased. They are conceptually simple and easy to understand, but exhibit a high variance and cannot be solved in iterative fashion (bootstrapping).
As mentioned, here we will introduce these methods following Chapter 5 of Sutton’s book…