Google acquired the London-based artificial intelligence (AI) firm DeepMind for £400m last year. Since the acquisition there hasn’t been a lot of information about what the team is working on.
However, new research published today in the science magazine Nature revealed that DeepMind’s AI software has learned to play Atari video games – and it’s good at playing them too.
The AI software, dubbed the “Deep Q network agent”, has learned how to play 49 video games and can play half of the games better than an experienced human player.
Deep Q managed to get 75 percent of the score of a professional tester in 29 games. The software performed particularly well in Video Pinball.
TheDeep Q software is based on two AI techniques; deep learning and deep reinforcement learning.
It uses an algorithm that helps it improve at performing specific tasks over time. This means that the system is not pre-programmed, but rather learns from experience – strictly using raw pixels as data input.
By playing the same game over and over again the AI software can eventually learn the optimal way of achieving high scores.
The team has developed a “a general-learning algorithm that should be applicable to many other tasks,” said Koray Kavukcuoglu, a Google researcher.
A huge advancement in self-learning systems
Dr. Demis Hassabis, a former gamer, chess champion, neuroscientist, and VP of DeepMind Engineering, said:
“These are the full results complete with a whole bunch of careful controls and benchmarks.”
“This work is the first time anyone has built a single, general learning system that can learn directly from experience to master a wide range of challenging tasks, in this case a set of Atari games, and to perform at or better than human level,”
“Up until now, self-learning systems have only been used for relatively simple problems. For the first time, we have used it in a perceptually rich environment to complete tasks that are very challenging to humans.”
“The ultimate goal is to build smart general purpose machines. We’re many decades from doing that, but this is the first significant rung of the ladder,”
