Hybrid Quantum-Classical Neural Networks for Downlink Beamforming Optimization

TL;DR

This work investigates hybrid quantum-classical neural networks to optimize downlink beamforming in a multiuser MISO system. It introduces two architectures—one inserting a parameterized quantum circuit after a classical CNN (QNN) and another placing a quantum convolutional layer before a classical CNN (QCNN)—and demonstrates that these hybrids can achieve comparable or superior sum-rate performance with far fewer trainable parameters, even under NISQ-like noise. The study provides detailed parameter-count analyses, training via the parameter-shift rule, and robustness evaluations against quantum noise using software simulators and hardware emulators. The results suggest practical potential for quantum-assisted resource optimization in wireless networks, with transfer learning and QCNN variants offering scalability advantages as the number of users and antennas grows.

Abstract

This paper investigates quantum machine learning to optimize the beamforming in a multiuser multiple-input single-output downlink system. We aim to combine the power of quantum neural networks and the success of classical deep neural networks to enhance the learning performance. Specifically, we propose two hybrid quantum-classical neural networks to maximize the sum rate of a downlink system. The first one proposes a quantum neural network employing parameterized quantum circuits that follows a classical convolutional neural network. The classical neural network can be jointly trained with the quantum neural network or pre-trained leading to a fine-tuning transfer learning method. The second one designs a quantum convolutional neural network to better extract features followed by a classical deep neural network. Our results demonstrate the feasibility of the proposed hybrid neural networks, and reveal that the first method can achieve similar sum rate performance compared to a benchmark classical neural network with significantly less training parameters; while the second method can achieve higher sum rate especially in presence of many users still with less training parameters. The robustness of the proposed methods is verified using both software simulators and hardware emulators considering noisy intermediate-scale quantum devices.

Figures (8)

• Figure 1: The classical neural network structure to solve \ref{['generalproblem']}.
• Figure 2: The proposed hybrid QNN structure to solve \ref{['generalproblem']}.
• Figure 3: The parameterized quantum circuit used in QNN, $Q=4$. $c_i$ is the $i$-th classical input data. There are $L$ entanglement layers in the middle part.
• Figure 4: The hybrid QCNN structure to solve \ref{['generalproblem']}.
• Figure 5: Detailed circuit of the quantum convolutional layer. Here we assume it processes a block of $2\times2$ classical data, and the number of qubits is $Q=2$. $c_i$ is classical channel data.