Why Generalization in RL is Difficult: Epistemic POMDPs and Implicit Partial Observability @rail7462
Why Generalization in RL is Difficult: Epistemic POMDPs and Implicit Partial Observability  @rail7462
Uploaded October 2021 | Updated November 2026, 4 days ago
A presentation of "Why Generalization in RL is Difficult: Epistemic POMDPs and Implicit Partial Observability", originally recorded for NeurIPS 2021.

This paper shows that when learning to generalize in RL actually requires solving an implicit partially observable problem (a POMDP), even though the original problem itself may be fully-observable (an MDP).

Ghosh*, Rahme*, Kumar, Zhang, Adams, Levine. Why Generalization in RL is Difficult: Epistemic POMDPs and Implicit Partial Observability, NeurIPS 2021

Paper: arxiv.org/abs/2107.06277
Why Generalization in RL is Difficult: Epistemic POMDPs and Implicit Partial ObservabilityCS 285: Lecture 22, Meta-Learning, Part 1CS 285: Lecture 23, Part 1: Challenges & Open ProblemsCS 185/285 (Spring 2026): Guest Lecture, Seohong ParkCS 285: Lecture 17, Part 1: RL TheoryHybrid Imitative Planning with Predictive and Geometric Costs in Offroad EnvironmentsCS 182: Lecture 2, Part 2: Machine Learning BasicsCS 285: Lecture 21, Transfer Learning, Part 5LangRob Workshop @ CoRL 2023 (Part 3)CS 182: Lecture 16: Part 3: Actor-Critic & Q-LearningCS 182: Lecture 20: Part 2: Adversarial ExamplesCS 285: Lecture 2, Imitation Learning. Part 3
RAIL |

Why Generalization in RL is Difficult: Epistemic POMDPs and Implicit Partial Observability

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER