DeepMind's DQN learns Atari games from raw pixels

What happened

DeepMind researchers led by Volodymyr Mnih, Koray Kavukcuoglu, and David Silver published 'Human-level control through deep reinforcement learning' in Nature. Their deep Q-network, or DQN, combined a convolutional neural network with reinforcement learning to play Atari 2600 games, receiving only the raw pixels on screen and the game's score as input, with no hand-coded knowledge of any game's rules. DQN used a technique called experience replay, storing past game transitions and randomly sampling from that memory during training rather than learning strictly in the order actions occurred, which stabilized the learning process. Using the same algorithm, network architecture, and hyperparameters across all games, DQN matched or exceeded the level of a professional human game tester on a majority of the 49 Atari games tested.

How we know

The Nature paper is confirmed genuine via its DOI record in Crossref, matching title, authors, journal, and page numbers; DeepMind's own blog post on the research independently describes the same experience-replay method and Atari results.

Sources

• Volodymyr Mnih, Koray Kavukcuoglu, David Silver, et al., Nature. Human-level control through deep reinforcement learning · Primary source (author-declared)web.stanford.edu · Cited as a "primary" source (no stronger domain match).
• Google DeepMind. Deep Reinforcement Learning · General sourcedeepmind.google · Cited as a "reference" source (no stronger domain match). · Link is live and its text matches the event's key terms (Jul 2026)