You might think you're good at Atari games, but a new artificial intelligence system from Alphabet subsidiary DeepMind called Agent57 is probably better. The company claims its A.I. can beat the average person on all 57 Atari 2600 games.
Agent57 uses a type of machine learning called deep reinforcement learning to learn from its mistakes and get better at playing the games. A research paper that was published by DeepMind on Tuesday explains why games are a great way to test out how A.I.
"Games are an excellent testing ground for building adaptive algorithms: they provide a rich suite of tasks which players must develop sophisticated behavioral strategies to master, but they also provide an easy progress metric – game score – to optimize against," the paper reads. "The ultimate goal is not to develop systems that excel at games, but rather to use games as a stepping stone for developing systems that learn to excel at a broad set of challenges."
DeepMind used the same type of machine learning to develop its A.I. system AlphaGo, which beat the 33-year-old grandmaster of the ancient Chinese game Go, Lee Sedol, at the game four out of five times in 2016. When AlphaGo won the first round against Sedol, Elon Musk commented that experts believe it would be a decade before A.I. could achieve such a feat.
Some of the most challenging games Agent57 had to tackle were Montezuma’s Revenge, Pitfall, Solaris, and Skiing. Other A.I. systems have had a difficult time with those games, but Agent57 did better than any A.I. has been able to before and exceeded the performance of the average person for the first time.
Pitfall and Montezuma's Revenge are difficult for A.I. because they require a lot of strategy. The games Solaris and Skiing are difficult for A.I. because it takes time to determine how your actions affected your overall performance, so it's hard for A.I. to learn from its mistakes. Agent57 was able to outperform humans despite these challenges.
"With Agent57, we have succeeded in building a more generally intelligent agent that has above-human performance on all tasks in the Atari57 benchmark," the paper reads. "Agent57 was able to scale with increasing amounts of computation: the longer it trained, the higher its score got. While this enabled Agent57 to achieve strong general performance, it takes a lot of computation and time; the data efficiency can certainly be improved."
The Atari 2600 was released in 1977, and millions of consoles were sold by 1980. The Atari changed gaming forever, and its games maintain a large fanbase to this day. Iconic games like Pitfall!, Missile Command, Space Invaders, Asteroids and more are still played by people around the world, but they're usually playing them on their computer. If you want to buy an actual Atari 2600, you'll have to shell out around $60 on eBay.
A.I. is developing quickly, and it's cathartic to see it advancing by playing old video games we all know and love. When A.I. can beat us at Super Smash Bros., then we'll start worrying.
The Inverse analysis
It's pretty incredible how much DeepMind's A.I. has developed in a relatively short time. We're curious what games this A.I. might master next and how it will be applied when they move on from dominating video games. As we've reported, A.I. is already capable of diagnosing cancer and predicting the weather, so there's no telling what it'll be able to do in the years to come. Hopefully, Elon Musk's nightmares don't come true and it ends up killing all of us.