REINFORCEMENT LEARNING
Reinforcement Learning is one of the most fascinating fields of machine learning. Unlike supervised learning, reinforcement learning models can learn complex processes independently, even without beautifully tabulated data.
For me, it is most fun to see AI agents win video games, but you can also use reinforcement learning to solve business problems. Just phrase it as a game, and off you go! You only have to define…
- the environment your agent lives in,
- what decisions your agent can take, and
- what success and failure look like.
Before you continue, please read my introductory article about reinforcement learning. It gives you some more context and shows you how to conduct a simple, yet effective form of reinforcement learning yourself. It also serves as a basis for this article.
In this article, you will learn about deep Q-learning, why we need it, and how to implement it yourself to master a game that looks much more difficult than the ones in my other article.
You can find the code in my Github.
In the article linked above, we conducted Q-learning to make an agent play some simple games with small discrete observation spaces. In the Frozen Lake game, as an example, you have 16 fields (=states or observations, I use these terms interchangeably from now on.) you can stand on in the 4×4 map. In the gymnasium version of the card game Blackjack, there are 32 · 11 · 2 = 704 states.