Dissertation

Hybrid Quantum-Classical Metaheuristics for Automated Machine Learning Applications

This thesis investigates how quantum, quantum-inspired, and hybrid quantum-classical computation can enhance key points of the automated machine learning (AutoML) pipeline under the constraints of noisy intermediate-scale quantum (NISQ) devices.

Author

D. Von Dollen

Date

18 November 2025

Links

Thesis in Leiden Repository

We explore whether quantum effects such as superposition, entanglement, and interference can provide practical advantages in feature selection, kernel approximation, and optimization metaheuristics. The research introduces QUBO-based formulations for selection while benchmarking classical, quantum annealing, and hybrid solvers across real-world and synthetic domains. We extend these approaches to evolution strategies, hyperparameter optimization, and neuroevolution. Additionally, we propose hybrid quantum-classical surrogate models for sparse Gaussian process regression and analyze their impact on Bayesian optimization, including experiments with both classical and quantum-generated data. Across these studies, we identify qualitative improvements in search behavior, robustness, and diversity even within NISQ limitations, and highlight scenarios where hybrid approaches outperform purely classical baselines. The findings point to promising strategies for integrating quantum resources into AutoML, offering machine learning practitioners new tools for exploring complex search spaces.

• automated machine learning
• automl
• bayesian optimization
• evolutionary algorithms
• gaussian process regression
• quantum machine learning