Skip to content
Open access

NCP-DRL: Geometry-Aware Deep Reinforcement Learning for Real-Time Transmission in Multi-HAP-Assisted Satellite Networks

2026 · IEEE Open Journal of the Communications Society · Vol 7, pp. 13324-13338 · 0 citations · 33 references

Abstract

Non-terrestrial networks (NTNs) integrating low Earth orbit (LEO) satellites and high-altitude platforms (HAPs) are expected to play an important role in beyond-fifth-generation (B5G) and sixth-generation (6G) wireless systems by providing wide-area coverage and resilient connectivity. However, the coexistence of multiple aerial and spaceborne platforms introduces highly dynamic network topologies, time-varying interference, and mobility-dependent channel variations, which significantly complicate resource management. Although deep reinforcement learning (DRL) has recently emerged as a promising approach for adaptive NTN optimization, existing methods often rely on generic state representations and do not explicitly exploit the geometry-dependent spatial structure of NTN propagation environments. To address this limitation, this paper proposes a Neural Channel Primitives-assisted DRL (NCP-DRL) framework for joint user association and beamforming in multi-HAP-assisted LEO satellite networks. The proposed NCP module, inspired by Instant Neural Graphics Primitives (Instant-NGP), performs multi-resolution spatial encoding to extract compact geometry-aware propagation features from user and transmitter locations. These features are incorporated into the DRL state representation to improve learning efficiency and policy adaptation under dynamic NTN conditions. The joint optimization problem is formulated as a Markov decision process (MDP) with hybrid discrete–continuous actions and QoS-aware reward design. Simulation results based on dynamic LEO satellite motion and HAP mobility demonstrate that the proposed framework achieves improved convergence behavior and enhanced spectral efficiency versus power-consumption trade-offs compared with conventional Soft Actor-Critic (SAC), successive convex approximation-matching algorithm (SCA-MA) benchmark schemes and the state-of-the-art generative artificial intelligence-based DRL framework, named Gen-DRL. The results highlight the potential of geometry-aware representation learning for DRL-based resource management in dynamic multi-layer NTN environments.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.