Google DeepMind AI Now Plays Atari Games

After deploying AlphaGo to win games against the best human Go players, Google’s DeepMind set their focus on another type of game.

DeepMind is now working on AI that’s able to play video games and the approach to this training is the same as what humans usually do to get achievements in games – they let the AI watch YouTube videos.

While AlphaGo was designed to win a game with specific rules, classic video games pose a huge challenge to AI, as they require a lot of exploration and decision making. To overcome this challenge, the DeepMind team explored ways in which their AI could take images from videos and convert them into the type of data it can learn from:

“Here we propose a two-stage method that overcomes these limitations by relying on noisy, unaligned footage without access to such data. First, we learn to map unaligned videos from multiple sources to a common representation using self-supervised objectives constructed over both time and modality (i.e. vision and sound). Second, we embed a single YouTube video in this representation to construct a reward function that encourages an agent to imitate human gameplay. This method of one-shot imitation allows our agent to convincingly exceed human-level performance on the infamously hard exploration games.”

You can read their paper here for more insights into the project. #machinemagic

While DeepMind won’t be beating Dark Souls this year, the company’s track record so far makes the idea of watching AI speedruns on Twitch into something that’s just around the corner, perhaps in a year or two.