Scientists Made a Huge Library of Atari Gameplay to Give AI a Power Up
Q*Bert and Ms. Pacman are featured, of course.
Q*Bert. Screengrab: www.youtube.com/watch?v=vkZhWsiHCqM
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 .
A massive database of human playthroughs could accelerate this kind of research, making researchers' lives easier and potentially producing more capable AI.
"When a person learns to play a game, you have lots of prior information about the world around you; but when an AI agent learns, it does it from scratch," said Vitaly Kurin, a Master's student at RWTH Aachen University and co-author of the paper, in an interview. "When an AI agent learns from human demonstration, it's like implicitly giving the bot all of the information we have about the world, and optimum strategies and behaviour. Everybody can use this dataset and check their ideas and models."
Kurin and his colleagues collected the database with a browser-based Atari 2600 emulator paired with an app that captures the player's input. They posted a link to the emulator to Reddit and asked people to help with the experiment by playing some video games, Kurin said. To make things faster and easier, the app only recorded the initial game state and every player input after that. The entire playthrough was then reconstructed offline using the player's recorded actions on the starting game state.
That might seem like a lot of work, but the payoff for research using the database could be huge. Machine learning using human demonstrations has already shown itself to be an impressive technique for quick learning. Researchers from the OpenAI institute recently unveiled an AI system that uses a single human demonstration in virtual reality to teach itself how to stack toy blocks.
"Let's imagine you have an autonomous car that you want to train to drive in the real world," Kurin said. "What if we gave the car a model that teaches it how to behave like a human? It suddenly becomes less dangerous because it won't do some crazy stuff after training completely from scratch."
It's enough to make me glad for my peanut human brain—when I play video games, it's just to relax. And that's fine with me.
Subscribe to Science Solved It , Motherboard's new show about the greatest mysteries that were solved by science.