COMBINING BAYESIAN METHODS WITH HIERARCHICAL REINFORCEMENT LEARNING FOR EFFICIENT POLICY TRANSFER
Abstract
Policy transfer remains a central challenge in reinforcement learning (RL), particularly when agents must adapt to related tasks in complex and uncertain environments. This paper presents a Bayesian Hierarchical Reinforcement Learning (BHRL) framework that combines hierarchical policy structure with Bayesian posterior inference. High-level policies select goals, low-level policies select primitive actions, and posterior updates represent uncertainty during target-task adaptation. The framework also introduces a policy-transfer-efficiency measure and a structured algorithm for mapping, regularizing, and refining source policies. A benchmark-oriented analytical comparison suggests that Bayesian uncertainty modeling may improve adaptability and sample efficiency relative to conventional RL and standard HRL, although confirmation requires reproducible experiments with raw benchmark data. The proposed formulation provides a consistent basis for future evaluation and application of uncertainty-aware hierarchical policy transfer.
Keywords : Hierarchical reinforcement learning, Bayesian methods, policy transfer, reinforcement learning, adaptive algorithms, uncertainty modeling, machine learning.
