Researched
Deep Q-Network (Atari)
DeepMind's DQN (2013; Nature 2015) learned Atari games from raw pixels and the score alone, reaching a level comparable to a professional human tester across 49 games with one network design.
Open in the interactive tree →Mnih and colleagues posted 'Playing Atari with Deep Reinforcement Learning' in December 2013: a convolutional network trained with a form of Q-learning beat earlier methods on six of seven games and a human expert on three. The 2015 Nature paper used one network architecture for all 49 games, with four 84x84 pixel frames as input. Two tricks kept training stable: learning from random samples of past experience, and moving towards target values that are refreshed only periodically.
Prerequisites
- Video Games1972The agent learns by playing Atari video games from their screen pixels.
- Reinforcement Learning (TD-Gammon)1992TD-Gammon showed that reward-driven learning works with a neural network.
- Deep Learning2012A convolutional network reads the raw game pixels.
Unlocks
- AlphaGo2016AlphaGo trains deep networks by reinforcement learning, as DQN first showed on Atari.
- RLHF (Learning from Human Feedback)2017The 2017 preference-learning paper trained deep reinforcement-learning agents on Atari games.