Multi-Branch Attention Convolutional Neural Network for Online RIS Configuration with Discrete Responses: A Neuroevolution Approach

TL;DR

The paper addresses online joint configuration of RIS unit elements with discrete phase states $\phi_i(t)\in\{-1,1\}$ and a codebook-based TX precoder in RIS-empowered MISO links operating under time-varying channels. It introduces a Multi-Branch Attention Convolutional Neural Network (MBACNN) whose weights are trained with NeuroEvolution (CoSyNE) to handle non-differentiable RIS states, and extends the architecture to distributed multi-RIS systems. Across stochastic and geometrical channels, the proposed method (NE-MBACNN) outperforms DRL-based policies, lightweight discrete optimizers, and larger FF networks, with robustness to hyper-parameter settings and good generalization. The distributed variant reduces control signaling and scales to multiple RISs while maintaining superior performance, making online RIS configuration more practical for 6G-era networks.

Abstract

In this paper, we consider the problem of jointly controlling the configuration of a Reconfigurable Intelligent Surface (RIS) with unit elements of discrete responses and a codebook-based transmit precoder in RIS-empowered Multiple-Input Single-Output (MISO) communication systems. The adjustable elements of the RIS and the precoding vector need to be jointly modified in real time to account for rapid changes in the wireless channels, making the application of complicated discrete optimization algorithms impractical. We present a novel Multi-Branch Attention Convolutional Neural Network (MBACNN) architecture for this design objective which is optimized using NeuroEvolution (NE), leveraging its capability to effectively tackle the non-differentiable problem arising from the discrete phase states of the RIS elements. The channel matrices of all involved links are first passed to separate self-attention layers to obtain initial embeddings, which are then concatenated and passed to a convolutional network for spatial feature extraction, before being fed to a per-element multi-layered perceptron for the final RIS phase configuration calculation. Our MBACNN architecture is then extended to multi-RIS-empowered MISO communication systems, and a novel NE-based optimization approach for the online distributed configuration of multiple RISs is presented. The superiority of the proposed single-RIS approach over both learning-based and classical discrete optimization benchmarks is showcased via extensive numerical evaluations over both stochastic and geometrical channel models. It is also demonstrated that the proposed distributed multi-RIS approach outperforms both distributed controllers with feedforward neural networks and fully centralized ones.

Figures (7)

• Figure 1: The real and imaginary parts of a $10\times50$ channel matrix $\mathbf{H}_1(t)$ constituting a realization of the Ricean distribution AlexandroPervasive with $7$ dB $\kappa$-factor. It can be observed that adjacent matrix elements have similar values (spatial correlation), a fact that motivates the investigation of attention mechanisms as a means to extract important channel features.
• Figure 2: The proposed MBACNN architecture comprising a multi-branch attention module, followed by a Convolutional Neural Network (CNN) module, a Multi-Layered Perceptron (MLP) module for the RIS phase configuration selection, and an additional MLP module for the TX precoding vector selection.
• Figure 3: Achievable rate with the proposed MBACNN framework versus (vs.) different system and channel parameters.
• Figure 4: Received SNR with the proposed MBACNN framework vs. TX power level for the considered geometric channels.
• Figure 5: Sensitivity analysis of the proposed NE-MBACNN scheme for the considered geometric channels.

Theorems & Definitions (1)

• Remark 1: Extension to Arbitrary Numbers of $\phi_i(t)$ Values