DeepMind Technologies, a company that develops artificial intelligence acquired by Google a little over a year ago, is developing a machine that can play with 49 classic video games of the Atari 2600 console. This system, called Deep Q-network (DQN), achieved a great ability to play, even exceeding the capacity of human beings. So far nothing strange since by now computers can beat humans in complex games but the important fact is that this artificial intelligence learned to on its own.
Demis Hassabis, one of the DeepMind founders, is a known name in the field of video games. In the development of this new artificial intelligence, he decided to make it play with 49 classic Atari arcades such as BreakOut. Normally, the systems used to play are programmed with the rules and strategies they need, instead DQN began without having any knowledge of the Atari games.
DQN was created inspired by the natural learning mechanisms. In the case of video games, it found itself in the same situation as a human player who is for the first time in front of a new arcade with no help and tries to figure out how it works on his own.
The Q in the artificial intelligence system name developed by DeepMind means Q-learning. It’s a reinforcement learning technique algorithm based on the awards that are given for a correct behavior. In the case of video games, those are the right strategies, which lead to scoring points.
The Atari 2600 console video games were classics in the ’80s. They’re very simple by today’s standards but becoming experts is still difficult even for humans. Strategies to get many points in Space Invaders or BreakOut are quite complex and a player takes time to discover them if he isn’t tought them by anyone.
DQN had to learn to play on its own and slowly, attempt after attempt, it succeeded. Its game strategies improved slowly but became more and more sophisticated. For example, in the case of BreakOut, DQN discovered on its own the strategy known by experienced players to dig a hole on the wall side to send the ball to its back side.
The results of this DQN training was published the journal “Nature” in recent days. Google’s idea is to apply this type of learning to other tasks that have a practical use. The interest can be in various technologies developed by the company directly or through its subsidiaries, including some of those acquired in recent years. Robots and self-driving cars are among the candidates for the use of these learning techniques.
[ad name=”Google Adsense 468″]
