– An excerpt of a study on AI that teaches itself to succeed at complex tasks (like playing Atari games 2600 games) published in Nature.
AI-minded developers, take note: A team of researchers at Google DeepMind claims to have developed an articial agent capable of learning to play (from scratch) 49 different Atari 2600 games at skill levels comparable to human beings.
Evidence to back up their claims is laid out in a study published in Nature this month, a full copy of which is available for purchase via the Nature website. You can also see footage of the AI agent, known as “DQN”, playing games like Space Invaders and Breakout in a Nature video (embedded above) on the topic.
In brief, the researchers claim to have built an autonomous agent that, when tasked to maximize its score in a title like Breakout with no prior understanding of the rules of the game, learns to play effectively (developing tricks like knocking the ball back behind the brick layer) based on a combination of deep learning and reinforcement learning techniques.
The researchers were interested in developing artificial agents that can successfully learn complex tasks from scratch, not ace old Atari games, but their work is likely to impact the future of game development.
It’s also worth noting that the research group includes Demis Hassabis, the former Bullfrog developer who (at the age of 17) co-designed and programmed Theme Park with Peter Molyneux. He later went on to found AI tech developer DeepMind in 2011, which was then acquired by Google early in 2014.