Deep Q-Network
DQN approximates Q(s, a) with a deep net, using experience replay and a frozen target network so the TD target does not chase itself every step.
Definition
Mnih et al. train a CNN on stacked Atari frames to predict Q-values. Replay samples past transitions; a target network is copied every N steps to stabilize the bootstrap. Nature (2015) is the journal version of the 2013 workshop paper.
Why it matters
DQN is the proof that deep function approximation can do RL from pixels. Every later value-based method (Rainbow, C51, IQN) is a footnote to this stack.
Case
The same agent that plays Breakout from pixels will look brilliant in a market simulator whose latency and fees were forgotten. That is not a DQN failure — it is an environment specification failure.
How to read it
Ask what the observation, action repeat, and reward clip were. Atari results are not transferable recipes for a trading MDP.
Key papers
Landmark references with real publisher or arXiv links — not token-overlap noise.
- Human-level control through deep reinforcement learningMnih et al. · 2015 · NatureDQN journal paper.