Meta-Learning for Resource Allocation in Uplink Multi-Active STAR-RIS-aided NOMA System

TL;DR

This work addresses uplink communications with multiple active STAR-RISs supporting NOMA, aiming to maximize sum-rate by jointly optimizing transmit powers, active beamforming, STAR-RIS reflection/transmission coefficients, and user-RIS associations. A Meta-DDPG framework is introduced to solve the non-convex problem by modeling it as an MDP and integrating meta-learning with DDPG for rapid adaptation in dynamic wireless environments. Empirical results show Meta-DDPG outperforms standard DDPG by about 19%, and second-order reflections among RISs provide up to 74.1% higher data rates, with multi-active STAR-RISs yielding superior performance over single-active RIS configurations. The approach demonstrates a practical path to enhanced spectral efficiency in RIS-assisted uplink networks, balancing performance gains with manageable computational overhead.

Abstract

Simultaneously transmitting and reflecting reconfigurable intelligent surface (STAR-RIS) is a novel technology which enables the full-space coverage. In this letter, a multi-active STAR-RIS-aided system using non-orthogonal multiple access in an uplink transmission is considered, where the second-order reflections among multiple active STAR-RISs assist the transmission from the single-antenna users to the multi-antenna base station. Specifically, the total sum rate maximization problem is solved by jointly optimizing the active beamforming, power allocation, transmission and reflection beamforming at the active STAR-RISs, and user-active STAR-RIS assignment. To solve the non-convex optimization problem, a novel deep reinforcement learning algorithm is proposed which integrates Meta-learning and deep deterministic policy gradient (DDPG), denoted by Meta-DDPG. Numerical results reveal that our proposed Meta-DDPG algorithm outperforms the DDPG algorithm with $19\%$ improvement, while second-order reflections among multi-active STAR-RISs provide $74.1\%$ enhancement in the total data rate.

Figures (5)

• Figure 1: System model.
• Figure 2: The reward in different episodes.
• Figure 3: The system's total rate in different episodes.
• Figure 4: The system's total rate versus $P_{\text{max}}$.
• Figure 5: The system's total rate versus the number of elements.