> ML_LITERATURE // MNIH-2015-HUMAN-LEVEL-CONTROL-DEEP-REINFORCEMENT-LEARNING-NATURE_v1.0
Human-level control through deep reinforcement learning (Nature DQN)
Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Sadik Sadik, Anton Antonoglou, Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, Demis Hassabis · Nature (2015)
foundational2015foundationalthirdPartyReproduced
Principal Contribution
Achieved human-level performance across 49 Atari games directly from raw pixel inputs using deep Q-learning with experience replay and target networks.
Operational Relevance
Landmark Nature cover paper proving that deep neural networks can master complex behavioral policy control from sensory inputs.
Assumptions
- Experience replay decorrelates consecutive transition tuples; periodic target network freezes stabilize non-stationary Bellman targets
Limitations
- Sample inefficient (required tens of millions of frames); overestimates Q-values (resolved by Double DQN)