DR3: Value-Based Deep RL Requires Explicit Regularization @rail7462
DR3: Value-Based Deep RL Requires Explicit Regularization  @rail7462
Uploaded July 2021 | Updated November 2026, 3 days ago
Paper at: arxiv.org/abs/2112.04716. Presented at NeurIPS 2021 Deep Reinforcement Workshop (Contributed talk), ICML 2021 workshop on Overparameterization: Pitfalls and Opportunities (Spotlight), and ICML 2021 workshop on RL for Real Life Workshop (Spotlight).

If you are interested in how implicit bias affects representations in deep Q-learning algorithms, consider checking out our previous work:

Implicit Under-Parameterization Inhibits Data-Efficient Deep Reinforcement Learning
openreview.net/forum?id=O9bnihsFfXU (Aviral Kumar*, Rishabh Agarwal*, Dibya Ghosh, Sergey Levine)
DR3: Value-Based Deep RL Requires Explicit RegularizationCS 182 Lecture 3: Part 1: Error AnalysisCS 185/285 (Spring 2026): Lecture 7, Value-Based RLCS 182: Lecture 18: Part 3: Latent Variable ModelsCS 285: Lecture 3, Part 5CS 182: Lecture 10: Part 1: Recurrent Neural NetworksCS 285: Chelsea Finn (Stanford)Ensuring Safety in Online Reinforcement Learning by Leveraging Offline DataCS 285: Guest Lecture: Bo DaiLangRob Workshop @ CoRL 2023 (Part 2)CS 182: Lecture 19: Part 3: GANsCS 182: Lecture 4: Part 1: Optimization
RAIL |

DR3: Value-Based Deep RL Requires Explicit Regularization

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER