Speaker
Description
Variational quantum algorithms are widely used in quantum computing but rely heavily on manually designed circuit architectures. Quantum Architecture Search (QAS) addresses this bottleneck by automating the circuit design process using classical machine learning. In this work, we propose novel QAS approaches that advance differentiable QAS by integrating architectural constraints, sequential gate memory, and surrogate neural networks. We evaluate a greedy (G-QAS) and a Differentiable QAS (DQAS) baseline against Architecture-Constrained DQAS (AC-DQAS), Memory DQAS (M-DQAS), and a Combined method (C-DQAS) which includes both architecture constraints and gate memory. Across three distinct benchmarks, we demonstrate that AC-DQAS yields the highest overall accuracy. Additionally, C-DQAS offers a highly effective compromise, having comparable accuracy to AC-DQAS whilst reducing both computational runtime and required circuit evaluations. Furthermore, we incorporate a surrogate neural network in the four differentiable methods, creating NN-DQAS, NN-AC-DQAS, NN-M-DQAS, and NN-C-DQAS. This surrogate is used to predict the best candidates circuit performances during the search, and only those best are evaluated using quantum circuits. We show that including this surrogate slightly decreases performance for all methods, but drastically reduces the needed number of quantum computer evaluations by at least a factor of five, making it more available for deployment on quantum hardware.
| Topic | Hybrid quantum-classical approaches for learning and optimization |
|---|