Cooperative Beam Selection for RIS-Aided Terahertz MIMO Networks via Multi-Task Learning

TL;DR

The paper tackles the hard combinatorial problem of cooperative analog beam selection for RIS-aided THz MIMO networks by formulating it as a multi-task classification problem. It introduces the MTL-ABS framework, which employs a shared ResNet-based representation and task-specific heads for beams at the BS, RIS, and users, with a self-attention module for the BS head and a determinant-based sum-rate metric to avoid matrix inversions. An offline IAS-generated dataset of CSI-to-beam labels is used to train the model, and the authors prove blockwise convergence (stationary point) while analyzing online complexity, demonstrating substantial reductions in beam selection overhead. Simulation results show MTL-ABS achieves near-optimal sum-rate comparable to ES/IAS with far lower complexity, validating its practicality for real-time RIS-enabled THz systems and scalability to larger subarray configurations. The work advances real-time, scalable beam management for 6G THz networks by combining multi-task learning with deep feature extraction and attention mechanisms.

Abstract

Reconfigurable intelligent surface (RIS) have been cast as a promising alternative to alleviate blockage vulnerability and enhance coverage capability for terahertz (THz) communications. Owing to large-scale array elements at transceivers and RIS, the codebook based beamforming can be utilized in a computationally efficient manner. However, the codeword selection for analog beamforming is an intractable combinatorial optimization (CO) problem. To this end, by taking the CO problem as a classification problem, a multi-task learning based analog beam selection (MTL-ABS) framework is developed to implement cooperative beam selection concurrently at transceivers and RIS. In addition, residual network and self-attention mechanism are used to combat the network degradation and mine intrinsic THz channel features. Finally, the network convergence is analyzed from a blockwise perspective, and numerical results demonstrate that the MTL-ABS framework greatly decreases the beam selection overhead and achieves near optimal sum-rate compared with heuristic search based counterparts.

Figures (9)

• Figure 1: Proposed MTL-ABS architecture.
• Figure 2: Blockwise structure of proposed MTL-ABS.
• Figure 3: Training loss versus number of epochs: (a) different learning rates; (b) different step sizes.
• Figure 4: Training loss versus the number of batches.
• Figure 5: Training loss versus the number of epochs.