Room 4 · The Awakening · 2015

DeepMind Learns to Play Atari

DeepMind Learns to Play Atari (DeepMind team, 2015): Learning to play by looking only at the screen.

A network called Deep Q-Network learned to play dozens of Atari video games receiving only the screen's pixels and the score. In 29 of the 49 games tested, it reached more than 75% of a professional human tester's score. It was published in Nature in 2015.

In February 2015 the journal Nature published "Human-level control through deep reinforcement learning," by Volodymyr Mnih and colleagues at DeepMind, a London company that Google had acquired the year before. The work built on an earlier version the team had shared in 2013.

The system, called a Deep Q-Network or DQN, learned to play classic Atari 2600 video games. It received only what a human player sees, the pixels on the screen, along with the score, and it could press the controller's buttons. Nobody told it the rules, what the objects on screen were or what a good strategy looked like.

It learned by reinforcement: try actions, see whether the score goes up, and gradually prefer the actions that lead to more points in the long run. A convolutional network, like the ones elsewhere in this room, turned the screen into useful features. Two practical tricks kept training stable. One, called experience replay, stored past moments of play and reused them in random order. The other was a slowly updated copy of the network used to compute learning targets.

The same network design and settings were used for all 49 games tested. DQN reached a level comparable to a professional human tester in many of them and clearly beat earlier methods. In Breakout, it discovered that digging a tunnel through one side of the wall sends the ball bouncing behind it, a strategy its designers had not programmed.

It also had clear limits. In games with sparse rewards, such as Montezuma's Revenge, where you need long sequences of actions before any points appear, it barely made progress.

Ask your AI: why is a game with rare rewards so much harder to learn than one with frequent rewards?

Did you know? Google had bought DeepMind in 2014.

Source: doi.org/10.1038/nature14236

See it in the museum →

← Generative Adversarial Networks · ResNet: 152-Layer Networks →

All of Room 4: The Awakening

Cite this: Nazareno, I. (2026). DeepMind Learns to Play Atari (2015). New Mind Museum. Retrieved November 17, 2026, from https://newmindmuseum.com/museum/deepmind-learns-to-play-atari-2015/