Video Games by Reinforcement Learning

Contrary to supervised learning, where an algorithm has to associate an input with an output, in reinforcement learning you have another kind of maximization task. You are given an environment (that is, a situation) and you are required to find a solution that will act (something that may require to interact with or even change the environment itself) with the clear purpose of maximizing a resulting reward. Reinforcement learning algorithms, then, are not given any clear, explicit goal but to get the maximum result possible in the end. They are free to find the way to achieve the result by trial and error. This resembles the experience of a toddler who experiments freely in a new environment and analyzes the feedback in order to find out how to get the best from their experience. It also resembles the experience we have with a new video game...