Deep Q Network (DQN) uses deep learning to do Q-learning

DQN is a variant of Q-learning done by DeepMind to train an AI to play Atari games.

An agent in the DQN Architecture:

• makes observations of what is happening in the game
• knows about states and rewards
• acts in the game by taking actions

The DQN agent in training has:

• a replay buffer
• a Q-function that is trained.

During training, Epsilon-greedy dictates the action of the agent during training.

Resources:

• Paper: https://arxiv.org/abs/1312.5602

mlaics