While video games are already used by soldiers or surgeons to implement combat strategies or better apprehend a tricky surgery, they are also unprecedented means of training for artificial intelligences under development. Otto, the self-driving truck, for example, improved its driving skills by playing Grand Theft Auto V -a modified version of the game from which any notion of violence was removed. This virtual dojo could be fruitful for five good reasons.
Video games, real world, what’s the difference?
The graphics of video games have become realistic enough to train artificial intelligence in the same way as in the real world. Interviewed by Motherboard, Mark Schmidt, professor of computer science at the University of British Columbia in Canada, praises this mode of training. Video games allow researchers to teach AI what the real world looks like much faster than with photos. In addition, AI can learn by itself. It doesn’t need to be told what it is looking at.
Minecraft, a training ground for personal assistants
Less realistic video games also do the trick. In an investigation about this new way of training AI, The Economist came to know about an astounding project: Project Malmo, which aims to develop an AI capable of integrating into the human environment, in order to solve their problems. To do this, Katja Hofmann, the project manager, trains it on Minecraft, the famous survival game where the player finds himself alone on an island and has to hunt and build a home directly from his environment, to live as long as possible. According to Hofmann, this game, with its unrealistic graphics, is complex enough to be relevant. For instance, he and his team used it to teach AI to help a human player to capture a fugitive pig. The algorithm must then learn by itself the notion of cooperation, by looking at how humans play.
Deepmind, king of the Atari
In early 2017, AlphaGo, Google’s Deepmind AI, destroyed the best go players in the world one after another. This prowess was undoubtedly made possible by a relentless training on a panel of 49 Atari games, about which Wired wrote an article in 2015. The most formative game was Breakout, a brick-breaker derived from Pong. After spending sleepless nights playing, the AI became “better than any human player because it has perfectly integrated the mechanics of the game”, says Demis Hassabis, co-founder of Deepmind. In short, if it takes a human player several days to grasp the rules of the game and succeed in moving on to the next level, artificial intelligence takes only a few hours to become the best player in the world, says Cinema Blend.
It’s all in the reflexes
Early 2017, MIT researchers developed an experimental artificial intelligence, which was also very good at Super Smash Bros. They have thus developed a neural network capable of coordinating its movements in game and of interacting with the different objects that stand in front of it. If playing against AI serves generally as a way of getting used to the game for newbies, the MIT’s AI has given the best players a hard time, although showing some limitations, especially in the face of unannounced gestures. In any case, it represents a sparring partner of choice, with constantly evolving talents.