Resource allocation in RIS-aided wireless communications : from optimization to continual learning

Abstract

The quest for physical-layer technologies pivotal to beyond-fifth-generation systems has begun, with Reconfigurable Intelligent Surfaces (RIS) being identified as a promising candidate. RIS can collect wireless signals from a transmitter and passively beamform them toward the receiver, enhancing performance and energy efficiency by reconfiguring the wireless propagation environment. To fully harness the potential of RIS, the development of corresponding resource allocation algorithms for RIS-aided wireless systems is essential.

Due to the unique properties of discrete amplitude coefficients and the passive nature of RIS, existing resource allocation algorithms are not well-suited for RIS-aided wireless systems, impacting both system performance and algorithm efficiency. Additionally, the computational complexity of RIS resource allocation makes it difficult to ensure real-time implementation. To address these issues, this thesis proposes advanced resource allocation algorithms for RIS-aided wireless communications.

Firstly, optimization-based resource allocation algorithms are explored, focusing on the wireless mobile edge computing scenario. Conventional methods struggle with feasibility issues due to RIS’s passive nature and face difficulties with discrete variables arising from transmission and reflection coefficients. To overcome these challenges, the hidden relationship between RIS phase shifts and quality-of-service constraints is exploited and converted into an explicit form for optimization, improving both the feasibility of the algorithms and system performance. For the discrete variables, an efficient smoothing-based method is proposed to decrease convergence error, in contrast to the conventional penalty-based method, which often results in undesired stationary points and local optima. These solutions provide robust frameworks for addressing the specific feasibility and discrete variable issues inherent in RISs, ensuring effective and efficient resource allocation.

Next, to better support the real-time implementation of resource allocation algorithms for RIS-aided wireless systems, the focus is shifted to the deep learning-based method. The limitations of traditional deep learning-based methods are critically analyzed, as they typically perform well only when the training data distribution matches the actual one. However, due to the nonstationary nature of wireless channels, this assumption is often not realizable in real-world scenarios. To address this issue, a novel continual learning paradigm is proposed. This approach enables deep neural networks to continuously learn from new data while preventing catastrophic forgetting, adapting to changes in channel conditions and ensuring robust performance in nonstationary environments. This new paradigm effectively addresses the challenge of resource allocation in RIS-aided wireless systems using traditional deep learning-based methods.