N -to-1 Knowledge Transfer in Reinforcement Learning via Adaptive Q -Function Selection

Read full text Summarise with VIRAT
Authors: Marko Ruman, Tatiana V. Guy
Cited by: 0
References: 20

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This paper tackles multi-source (N-to-1) knowledge transfer in reinforcement learning (RL), where an agent must adaptively solve a new task by leveraging a library of pre-learned skills. We consider a setting where these skills are represented as multipleQ-functions and corresponding environment models, without any explicit labels to guide the selection. To address this scenario, we introduce a theoretically grounded method that dynamically selects the most suitableQ-function at each learning stage. Instead of relying on noisy, short-term signals, our approach makes a farsighted choice by simulating the long-term performance of each skill while simultaneously evaluating the trustworthiness of its underlyingworld model. This allows the agent to intelligently transition from relying on transferred knowledge to using its own newly acquired policy. The proposed method is evaluated in three scenarios: (1) selection among multipleQ-functions to solve a fixed RL task, (2) adaptation in a dynamically changing environment, and (3) switching from a partially learnedQ-function to a newly learned one. In all cases, our method accelerates learning and demonstrates robust adaptation, confirming its effectiveness for scalable multi-source transfer in RL.

Fields of Study & Subjects

Reinforcement Learning in Robotics Domain Adaptation and Few-Shot Learning AI-based Problem Solving and Planning

Publication details

Journal / source
IEEE Access
Volume
14
Pages
45964–45976
Publication date
1 January 2026
Language
EN