MinedDeep

DQN (Deep Q-Network)

2 min read

DQN (Deep Q-Network)

Type: project Slug: project—DQN Sources: human-level-control-through-deep-reinforcement-learning—hassabis, overcoming-catastrophic-forgetting-in-neural-networks—hassabis Last updated: 2026-05-13

Summary

DQN is the first DeepMind project to demonstrate that deep neural networks combined with reinforcement learning can achieve human-level performance across diverse tasks. Published in Nature (2015), DQN played 29 Atari 2600 games from raw pixel input, achieving human-level or superhuman performance on 23. Its two key innovations — experience replay and target networks — became standard in deep RL. The project established DeepMind as a serious research lab and contributed to Google’s 2014 acquisition.

Core content

Architecture (paper—human-level-control-through-deep-reinforcement-learning): A convolutional neural network takes raw 210×160 pixel frames as input, outputs Q-values for each of 18 possible actions. Trained with Q-learning using two stabilisation innovations: (1) experience replay — storing transitions in a replay buffer and sampling random mini-batches to break temporal correlations; (2) target network — a slowly-updated copy of the Q-network that provides stable targets for TD learning.

Results: Human-level or superhuman on 23/29 games. Strongest on games requiring precise control (Breakout, Enduro); weakest on games requiring long-horizon planning (Montezuma’s Revenge).

Continual learning extension (paper—overcoming-catastrophic-forgetting-in-neural-networks): Elastic weight consolidation (EWC) addressed DQN’s vulnerability to catastrophic forgetting when learning sequentially across games, drawing on synaptic consolidation in neuroscience.

Key personnel: Volodymyr Mnih (first author), Koray Kavukcuoglu, David Silver, Ioannis Antonoglou, Daan Wierstra, Demis Hassabis.

Connections

  • Theme: theme—deep-RL, theme—reinforcement-learning, theme—game-playing-AI
  • Period: period—early-deepmind
  • Precedes: project—AlphaGo (DQN’s deep RL paradigm extended to Go)

Honest Gaps

  • The 2010–2014 development history is entirely absent from the corpus.
  • No source documents whether the neuroscience background of the lab influenced DQN’s design (e.g., experience replay as analogous to hippocampal replay).
  • The actual Nature paper has more authors than listed in metadata.
  • No comparison with contemporaneous deep RL approaches from other labs.

Graph View

Backlinks