DQN

← Back to the article

DQN is Q-Learning with a neural network that can look at a game screen.

What it does

DQN is Q-Learning with a neural network that can look at a game screen.

Deep Network: Complex AI brain • State: Current game screen

How it works

  • Look at the current screen or state.
  • Use a neural network to estimate good actions.
  • Choose an action that should lead to more reward.

Uses and applications

  • A game AI can learn from pixels in an Atari-style game.
  • A robot simulator can learn actions from screen-like views.
  • A training system can test choices in a safe virtual world.

In short

DQN combines Q-Learning with deep learning.

Related methods

  • Q-Learning: The table-based method DQN extends.
  • Neural Network: Provides the function that estimates action values.
  • AlphaZero: Combines learned values with tree search.

Back to Reinforcement learning, agents and planning · All topics

The interactive demonstration and the Python walkthrough for this topic need JavaScript.