Games and Reinforcement Learning

Video games are an excellent simulation environment for training machine learning algorithms. Because of games’ increasing complexity, games can be viewed as a model for our own reality. Learning how to play games is the first step to learn how to operate in real life. We value play as a way to learn for our children, and play is equally good for machines.

Reinforcement Learning (RL) techniques seem particularly well suited for games. The main focus of RL is to reward an algorithm when it completes a sub-task or moves in a good direction, and give it a penalty when it doesn’t. This is the closest to raising a child with a set of rules on which it builds its world-view model. Reinforcement Learning agents learn tasks by trial and error. They must balance exploration (new behaviors) with exploitation (repeating past behaviors).

Experiments in Reinforcement Learning within games like Go, DOTA 2, or Quake III Capture the Flag show that...