The authors propose IDS-based RLHF algorithms with a surrogate environment and a new ℓ_g distance, proving Bayesian regret bounds of order O(H^(3/2) sqrt(T log K)).
However, Lemma C.1 can only be applied to handle the difference between two value functions with the same policy and different environments, while inV eE ∗ t 1,π∗ E (st
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Sample-Efficient Reinforcement Learning from Human Feedback via Information-Directed Sampling
The authors propose IDS-based RLHF algorithms with a surrogate environment and a new ℓ_g distance, proving Bayesian regret bounds of order O(H^(3/2) sqrt(T log K)).