An Age of Service and Transformer-Driven Transfer Learning Framework for Holographic MIMO-Enabled 6G Networks

Download PDF
Open Webpage

Apurba Adhikary, Yu Qiao, Avi Deb Raha, Luyao Zou, Mrityunjoy Gain, Zhu Han, Choong Seon Hong

Published: 01 Jan 2025, Last Modified: 06 Nov 2025IEEE Transactions on Network Science and EngineeringEveryoneRevisionsCC BY-SA 4.0
Abstract: The forthcoming 6G wireless communication networks are expected to provide widespread mobile connectivity, ultra-fast data services with minimal power consumption, enhanced performance, and seamless integration of diverse technologies for efficient beamforming. To achieve these objectives, an AI framework empowered by age of service (AoS) and transfer learning is proposed to optimize power allocation for users within the coverage areas of holographic MIMO-enabled base stations (HM BSs). The AI framework activates the required number of grids from the respective HM BSs, utilizing AoS of user service information from the HM BSs to identify the teacher HM BS (THM BS) among the serving HM BSs. The HM BS serving the highest number of users is designated as the THM BS, while the others are classified as student HM BSs (SHM BSs). An optimization problem is formulated to maximize the utility function for the achievable rate (UFA), incorporating user information freshness, achievable user rates, and the learning cost of the HM BSs, thereby leading to enhanced signal-to-interference-plus-noise ratio (SINR), improved achievable rates (AR), and significant power savings. A transformer-based AI framework, assisted by AoS and transfer learning, is employed to efficiently allocate power to users within the coverage areas of the THM and SHM BSs. The THM BS is initially trained using a large dataset, after which the trained model's knowledge is transferred to the SHM BSs, which require smaller datasets, enabling efficient power allocation for users in both THM and SHM BSs with reduced learning costs. Finally, simulation results demonstrate the effectiveness of the proposal, which outperforms baseline methods such as gated recurrent unit and long short-term memory, achieving significant improvements in AoS prediction accuracy, power savings, SINR, and AR.