Deep Q-network
Also: DQN, deep Q-learning, deep Q-networks
A network Q(s, a; θ) replaces the Q-table so the method scales to huge state spaces like raw pixels. Training samples random minibatches from a replay buffer and regresses toward r + γ·maxₐ′ Q(s′, a′; θ⁻) from a frozen target network.
L(θ) = E[ (r + γ maxₐ′ Q(s′, a′; θ⁻) − Q(s, a; θ))² ]
Example
For Atari, the input is a stack of 4 grayscale 84×84 frames and the output is one Q-value per joystick action; the agent plays the action with the highest output.