With a single algorithm, a computer can learn dozens of classic video games, researchers from Google DeepMind in London report in the Feb. 26 Nature. With no prior knowledge of video games, the algorithm considers the visual elements of Atari 2600 games including Pong, Q*bert and Ms. Pac-Man. It then tries out various game moves and takes the resulting score into account to build a strategy.
For 29 of the 49 games tested, the algorithm scored at least 75 percent as high as a professional human games tester.
“This is the first artificial agent capable of learning to excel across a diverse array of challenging tasks,” team member Demis Hassabis said at a press briefing.
After playing the classic Atari game Breakout hundreds of times, a new computer algorithm learns the best strategies for high scores.
Google DeepMind (with permission from Atari Interactive Inc.)