While computer systems can be programmed to perform specific pre-defined actions based on the information they're given, they've yet to achieve something that resembles human learning. Google is now trying to teach artificial intelligence the intricacies of trial and error by making computer programmes "curious", and it once again turning to the Atari to do so.
Deep Q is no stranger to Atari games – in fact, it's quite the pro. Google's AI is capable of beating at least 49 titles thanks to the time it spent watching these games being played, and has even managed to beat professional human players' top scores in 23 of these.
Montezuma's Revenge, however, offers a unique set of challenges. The notoriously difficult 1984 platformer tasks players with navigating a maze comprised of 72 trap-filled rooms. The game requires careful planning and forethought to complete, something current AIs aren't particularly adept at. Indeed, previous attempts by Deep Q at the game have ended badly, with the computer managing to clear just two of the game's rooms.
Yet thanks to a new algorithm DeepMind calls "intrinsic motivation," Google's AI has been able to improve its game significantly. This works by rewarding the AI for exploring its environment and testing out different approaches to the game, which is delivered on top of the overall reward incentive for completing the game. As a result, Deep Q learns a sort of curiosity for the world around it and begins testing new approaches on its own accord.
Thanks to intrinsic motivation, Deep Q was able to beat 15 of 24 rooms presented to it in DeepMind's latest experiment. Granted that was after 50 million frames, but it's still a massive improvement on its previous record of two.
Google isn't the only company tapping the power of video games to teach computers to be smart. In March, Microsoft announced it would let researchers use Minecraft as a testing ground for artificial intelligence in the hopes that the game's open world will teach AIs to better think for themselves.
Popular Science reports that DeepMind is planning to task its AI with more more complex games in future, including sci-fi strategy title StarCraft. When AIs get that clever, it might be time to start getting worried. Saying that, Google's other AI has already managed to beat a world-champion Go player, meaning that time may already be here.