Applying Deep Reinforcement Learning to Poker

We will cover the subject of Deep Reinforcement Learning, more specifically the Deep Q Learning algorithm introduced by DeepMind, and then we’ll apply a version of this algorithm to the game of Poker.

Reinforcement learning

Machine Learning and Deep Learning have become a hot topic in the past years. With the recent improvements in parallel computing, we have witnessed in the last decades some major breakthroughs. Algorithms are consistently solving very complex tasks such as Image/Video recognition and generation. These algorithms generally require huge datasets to achieve reasonable performances.

Reinforcement Learning is a type of Machine Learning where an algorithm doesn’t have training data at the beginning. The goal is for an agent to evolve in an environment and learn from its own experience. In order for a Reinforcement Learning algorithm to work, the environment (state based on actions taken) must be computable and have some kind of a reward function that evaluates how good an agent is. This is why it can be easily applied to games, as the states are clearly defined by the program, the rules are generally simple, and there is a clear goal (generally a score metric). However, Reinforcement Learning can also be applied to simulate real-world problems like robotics simulation.