Google AI Plays Video Games Better Than You
Google has once again created something that we can all marvel at and then question why it even exists -- an AI that can play games better than we can.
Unfortunately, all your years of working on your favorite game, determined to get the highest score, is now for naught and Google is to blame. NewScientist reports Google has created an artificial intelligence software that can master games just by watching them be played, much like that one friend that we all have and hate. However, this AI is way better than your friend, and has been trained to play 49 different video games from Atari 2600, 23 of which where it now holds the high score. Take that, professional gamers. The whole thing is done based on an algorithm that the software figures out called a deep neural network, which it uses to study the game, and from its new knowledge, evaluate what actions would get it the highest score.
This AI doesn't just stop with figuring out the best way to beat games, however, it has some tricks up its sleeve as well. Among the games that it's best at includes Breakout, the game where you bounce a ball against a platform to clear rows of blocks. The computer ended up figuring out the trick that many pros use of tunneling through one column of bricks and bouncing the ball off the back wall.
The reason this system works is because the designers combined the neural networks, which are often used for image recognition problems, with reinforcement learning, which deals with rewards for doing something correctly. This likens to the experience of reinforcement learning with the high score that gamers get from playing games. The great thing about this is even though neural networks have been playing games like backgammon for a while now, computer advancements allow for more data to be collected and held, which allows for better results when playing games.
Obviously this software wasn't just built for getting high scores on games. Although Google hasn't admitted what the software is really going to be used for, people close to the project have suggested that it would revolve around Google's ad placements. "You can use reinforcement learning methods to improve ad quality," Dalle Molle Institute for Artificial Intelligence Research in Manno, Switzerland's Jürgen Schmidhuber explained. "You learn to place ads that are more likely to be clicked on, which means higher rewards. This is presumably one of [Google's] motivations."
You can read more about the AI in NewScientist's post about it, but it seems that there are many things that this new software can bring to Google besides just helping out its average ad clicks.