Speaker
Description
Gaussian process regression is a widely used kernel-based supervised machine learning method. The kernel choice highly affects the regression quality and usually requires extensive hyperparameter tuning. Quantum fidelity kernels have emerged as expressive alternatives for regression tasks. However, designing a suitable circuit requires domain-specific knowledge and is often done heuristically. We propose a fully automatic quantum kernel architecture search routine for Gaussian process regression, based on proximal policy optimization. Unlike previously proposed approaches that only optimize the gate sequence and number of layers, our method additionally optimizes the number of qubits, enabling a significantly larger search space. We define a custom reward function that directly minimizes the negative log marginal likelihood. Furthermore, we demonstrate that our method generates quantum kernels that surpass established benchmark kernels found in the literature on a synthetic regression problem.
| Topic | Quantum kernel methods and their applications |
|---|