Deep Q-Network (DQN)

A Deep Q-Network, or DQN, is the agent that first showed deep RL could work at scale: in 2015 it learned to play dozens of Atari games straight from the screen, reaching human level on many using one architecture and one set of hyperparameters. You hand it raw pixels and a score, and it discovers on its own which actions earn points. It was the moment RL stopped being a toy-problem discipline and started beating people at real games.

Under the hood DQN is Q-learning with a convolutional network estimating Q(s,a) — the expected return of each action — for every game. It trains by minimising the gap between its current estimate and a bootstrapped target r plus gamma times the best next value. Two tricks make this stable: an experience replay buffer that breaks the correlation between consecutive frames, and a separate target network that holds the target fixed for a while. Without both, the network chases its own shifting predictions and diverges.