Reinforcement Learning
Playing Atari with Deep Reinforcement Learning
L4 · ExpertMnih, Kavukcuoglu, Silver, Graves, Antonoglou, Wierstra, Riedmiller · 2013 · NeurIPS Deep Learning Workshop 2013
TL;DR
Learns to play Atari games straight from raw pixels by combining Q-learning with a deep convolutional network, the Deep Q-Network (DQN).
Why it matters
DQN showed that deep networks could learn control policies end to end from high-dimensional input, launching modern deep reinforcement learning.
Key ideas
- Approximate the action-value function Q with a convolutional network over pixels.
- Experience replay stores past transitions and samples them to stabilize learning.
- A separate, slowly-updated target network reduces feedback instability.
- One architecture learned many games with the same hyperparameters.
Related concepts
reinforcement-learningdeep-learning