Scientists Made a Huge Library of Atari Gameplay to Give AI a Power Up

Artificial intelligence is slowly proving that that video games aren't a total waste of time, at least for machines: It's through learning to play games that AI algorithms can acquire all sorts of generalizable skills, like problem-solving. Now, computer scientists from RWTH Aachen University in Germany and Microsoft Research have released the largest-ever database of human playthroughs for some of the most popular games for the Atari 2600. Artificial agents using deep learning techniques will be able to pull patterns out of these playthroughs and learn from them. According to a paper posted to the arXiv preprint server this week, which is undergoing peer review, the database contains more than 45 hours of gameplay from five games: Q*Bert, Ms. Pacman, Space Invaders, Video Pinball, and Montezuma's Revenge. Video games are an increasingly popular training ground for AI to solve general problems, like how to quickly arrive at a course of action, or how to effectively learn in an environment where the rewards for learning are sparse, which is notoriously the case for Montezuma's Revenge .