BROWSE
Detail
Conference Paper
Reinforcement Learning-Based Traffic Steering for TN-NTN Integrated DualSteer Systems
Cited - time in 
Share 
Share
- Authors
- Jimin Jeon, Sung Hyuk Byun, HyunKyung Yoo, Seokwon Jang, Namseok Ko
- Issue Date
- 2026-06
- Citation
- European Conference on Networks and Communications (EuCNC) 2026, pp.1158-1164
- Publisher
- IEEE
- Language
- English
- Type
- Conference Paper
- Abstract
- Integrating terrestrial networks (TN) with non-terrestrial networks (NTN) is a key enabler of ubiquitous 6G connectivity, yet exploiting both paths simultaneously remains challenging due to the disparity in latency and channel dynamics between the two access types. Leveraging the 3GPP Release 19 DualSteer architecture, in which user equipment (UE) maintains concurrent sessions on both TN and NTN, this paper proposes a reinforcement learning (RL)-based traffic steering method that dynamically selects the optimal downlink path. We formulate the multi-user steering problem as a Markov decision process and design a Deep Q-Network (DQN) agent whose reward function maximizes reliability-to-latency efficiency, jointly capturing packet loss rate (PLR) and round-trip time (RTT). Simulation results show that the proposed method reduces PLR by 79% relative to the 3GPP ATSSS Smallest Delay policy (Min-RTT) while achieving comparable performance to the computationally infeasible exhaustive search with up to 838 \times inference speedup, demonstrating that near-optimal reliability and latency are achievable in real time.
- Keyword
- TN-NTN integration, DualSteer, traffic steering, reinforcement learning, deep Q-network
- KSP Keywords
- Deep Q-Network, Learning-based, Markov Decision process, Reward function, Round Trip Time(RTT), Steering method, Traffic steering, User equipment(UE), exhaustive search(ES), multi-user, packet loss rate