Dueling Posterior Sampling for Preference-Based Reinforcement Learning
Ellen R. Novoseller, Yibing Wei, Yanan Sui, Yisong Yue, Joel W. Burdick
TL;DR
This work addresses learning under human trajectory-level preferences in reinforcement learning by introducing Dueling Posterior Sampling (DPS), a Bayesian, model-based algorithm that jointly infers environment dynamics and a utility function from trajectory preferences. DPS integrates a credit assignment mechanism to translate sparse trajectory-level feedback into state-action rewards, using Bayesian linear regression and Thompson sampling to balance exploration and exploitation. The authors establish an asymptotic Bayesian no-regret guarantee for DPS under a linear-credit model and demonstrate strong empirical performance across several benchmark domains, robust to the choice of credit assignment model. Overall, the paper extends posterior-sampling RL to the preference-based setting with theoretical guarantees and practical effectiveness, offering a principled framework for learning from human preferences in complex environments.
Abstract
In preference-based reinforcement learning (RL), an agent interacts with the environment while receiving preferences instead of absolute feedback. While there is increasing research activity in preference-based RL, the design of formal frameworks that admit tractable theoretical analysis remains an open challenge. Building upon ideas from preference-based bandit learning and posterior sampling in RL, we present DUELING POSTERIOR SAMPLING (DPS), which employs preference-based posterior sampling to learn both the system dynamics and the underlying utility function that governs the preference feedback. As preference feedback is provided on trajectories rather than individual state-action pairs, we develop a Bayesian approach for the credit assignment problem, translating preferences to a posterior distribution over state-action reward models. We prove an asymptotic Bayesian no-regret rate for DPS with a Bayesian linear regression credit assignment model. This is the first regret guarantee for preference-based RL to our knowledge. We also discuss possible avenues for extending the proof methodology to other credit assignment models. Finally, we evaluate the approach empirically, showing competitive performance against existing baselines.