DeepMind has trained programs to play classic Atari video games by watching YouTube videos, The Register reported.
Researchers said they chose three classic Atari games that are difficult for typical reinforcement learning algorithms: Montezuma’s Revenge, Pitfall, and Private Eye.
A self-learning program, or agent, typically learns how to beat a game through a rewards system.
As the agent progresses through the game it earns rewards or is penalised for making a wrong move, reinforcing whether its approach was good or bad.
Exploration games like the three chosen for the research are difficult for reinforcement learning algorithms, however, as there is no indication of where to go next.
The researchers said they used two techniques to train their algorithms to imitate a human player: temporal distance classification (TDC), and cross-modal temporal distance classification (CDC).
With TDC, they effectively enabled their algorithm to detect changes in the game’s visual environment. CDC allowed the agent to learn from audio queues, with certain sounds corresponding to certain actions.
The video below shows the agent in action.