SPA-QNAS: Improving Search Efficiency and Stability in Evolutionary Quantum Neural Architecture Search

Area of Science:

• Quantum Computing
• Quantum Machine Learning
• Artificial Intelligence

Background:

• Quantum Neural Architecture Search (QNAS) designs parameterized quantum circuits (PQCs) for quantum machine learning.
• Existing Quantum Evolutionary Algorithm (QEA)-based QNAS methods face challenges with slow convergence and poor adaptivity in large search spaces.

Purpose of the Study:

• To introduce Structural Probability Adaptive Quantum Neural Architecture Search (SPA-QNAS) to enhance QNAS efficiency and performance.
• To address limitations in search dynamics and update adaptivity in QEA-based QNAS.

Main Methods:

• SPA-QNAS enhances the EQNAS framework with Structural Probability Enhancement (SPE) and Adaptive Evolutionary Control (AEC).
• SPE focuses sampling distribution on high-fitness regions.
• AEC stabilizes updates for non-elite individuals using fitness feedback.

Main Results:

• SPA-QNAS demonstrated superior classification accuracy and stability compared to EQNAS on MNIST and Warship datasets.
• Achieved 99.42% accuracy on MNIST and 85.33% on Warship under identical resource constraints.
• SPA-QNAS showed improved performance within the same search space and circuit template.

Conclusions:

• Enhancing Quantum Probability Vector (QPV)-based evolutionary update dynamics is crucial for robust QNAS.
• SPA-QNAS offers a more stable and effective approach for designing PQCs under fixed quantum resource limitations.