Hyperparameter Tuning for Quantum Machine Learning

Abstract

In recent years, the computational requirements of modern Machine Learning (ML)applications have increased significantly. The upcoming post-Moore era therefore forces scientists to search for alternative forms of computing that can meet computational demands beyond the capabilities of classical von Neumann architectures. Quantum computing emerged as a very promising paradigm for providing the necessary computational resources, as several quantum algorithms have proven to be more efficient for certain problems. The great interest in exploiting the capabilities of quantum hardware to speed up machine learning applications contributed to the rise of Quantum Machine Learning (QML).The most promising approach for QML are Variational Quantum Algorithms (VQAs), that combine classical hardware to overcome the limitations of current quantum hardware. Variational Quantum Algorithms (VQAs) use an optimizer on classical hardware to train a Parameterized Quantum Circuit (PQC), that is used to find the quantum state containing the solution to the problem. However, the optimal choice of the optimizer, the structure of the PQC and other hyperparameters is problem-specific and has a major impact on the performance of VQAs. The large number of available options makes manual testing extremely time-consuming and therefore requires automated solutions. In classical ML, automated hyperparameter tuning is widely used, but there are only few studies on its application to QML. In this thesis, we therefore investigate the applicability and performance of different automated hyperparameter tuning algorithms for QML classification tasks.Our results show that choosing the right hyperparameter tuning algorithm is essential and allows to reliably find near optimal configurations. Nevertheless, we also see that the barren plateau phenomenon significantly impacts the runtime of these algorithms and must be considered in future QML projects. Overall, our results highlight the complexity of hyperparameter tuning for QML applications and provide valuable insights for future projects.