A Practitioner’s Guide to Reinforcement Learning | by Dr. Robert Kübler | Nov, 2023

Take your first steps in writing game-winning AI agents

In machine learning, data scientists primarily navigate the territories of supervised and unsupervised learning. However, there is a distinct and interesting subfield — reinforcement learning!

In reinforcement learning, we try to teach a so-called agent how to navigate the complexities of games, placing it within a simulated environment where it explores strategies, receives rewards for successful moves, and faces penalties for missteps.

One prominent outcome of the field of reinforcement learning is AlphaGo, a model that has beaten the world champions of Go, a game more complex than chess.

The great thing about reinforcement learning is that we do not have to tell the agent how to win. We just need to tell it what winning or losing looks like.

In chess, for example, it’s checkmating the opponent’s king, and that’s the only guidance we provide. No explicit instructions on the importance of queens or the insignificance of pawns — the agent deduces these nuances itself.

And it’s not restricted to traditional games; virtually anything can be treated as a game. Whether it’s a classic board game, a video game, or a business scenario, such as determining the most effective ad for a customer, reinforcement learning is at play. In the business scenario, the agent could gain rewards for successful customer purchases, lesser rewards for ad clicks, and face penalties when ads are ignored by a customer. It becomes a strategic game for the agent, optimizing rewards, which, in a business context, translates to revenue.

In this article, I will not go too much into the mathematical theory of reinforcement learning. I want to give you the intuition and working code to get you started. To this end, I will use the great library gymnasium that provides some neat game environments that our agents can learn to master.