Fusing Rewards and Preferences in Reinforcement Learning

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Khorasani, Sadegh, Salehkaleybar, Saber, Kiyavash, Negar, Grossglauser, Matthias
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866911106191065088
author Khorasani, Sadegh
Salehkaleybar, Saber
Kiyavash, Negar
Grossglauser, Matthias
author_facet Khorasani, Sadegh
Salehkaleybar, Saber
Kiyavash, Negar
Grossglauser, Matthias
contents We present Dual-Feedback Actor (DFA), a reinforcement learning algorithm that fuses both individual rewards and pairwise preferences (if available) into a single update rule. DFA uses the policy's log-probabilities directly to model the preference probability, avoiding a separate reward-modeling step. Preferences can be provided by human-annotators (at state-level or trajectory-level) or be synthesized online from Q-values stored in an off-policy replay buffer. Under a Bradley-Terry model, we prove that minimizing DFA's preference loss recovers the entropy-regularized Soft Actor-Critic (SAC) policy. Our simulation results show that DFA trained on generated preferences matches or exceeds SAC on six control environments and demonstrates a more stable training process. With only a semi-synthetic preference dataset under Bradley-Terry model, our algorithm outperforms reward-modeling reinforcement learning from human feedback (RLHF) baselines in a stochastic GridWorld and approaches the performance of an oracle with true rewards.
format Preprint
id arxiv_https___arxiv_org_abs_2508_11363
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Fusing Rewards and Preferences in Reinforcement Learning
Khorasani, Sadegh
Salehkaleybar, Saber
Kiyavash, Negar
Grossglauser, Matthias
Machine Learning
I.2.6
I.2.6
We present Dual-Feedback Actor (DFA), a reinforcement learning algorithm that fuses both individual rewards and pairwise preferences (if available) into a single update rule. DFA uses the policy's log-probabilities directly to model the preference probability, avoiding a separate reward-modeling step. Preferences can be provided by human-annotators (at state-level or trajectory-level) or be synthesized online from Q-values stored in an off-policy replay buffer. Under a Bradley-Terry model, we prove that minimizing DFA's preference loss recovers the entropy-regularized Soft Actor-Critic (SAC) policy. Our simulation results show that DFA trained on generated preferences matches or exceeds SAC on six control environments and demonstrates a more stable training process. With only a semi-synthetic preference dataset under Bradley-Terry model, our algorithm outperforms reward-modeling reinforcement learning from human feedback (RLHF) baselines in a stochastic GridWorld and approaches the performance of an oracle with true rewards.
title Fusing Rewards and Preferences in Reinforcement Learning
topic Machine Learning
I.2.6
I.2.6
url https://arxiv.org/abs/2508.11363