An Empirical Study of Optimizers for Quantum Machine Learning
Yiming Huang, Hang Lei, Xiaoyu Li · 2020
The hybrid quantum-classical framework has been attracted more researchers, no matter in classical machine learning or quantum physics and quantum chemistry. Although the existing models are implemented on small and intermediate scale system or datasets, they also provide the foresight of the large scale applications in the future. Because the hybrid framework consists of the parametrized quantum circuit and well-developed classical optimizers, the selection of appropriate optimizer for different task matters the final performance of the model. In this paper, we take nine widely employed optimizers, including gradient based and gradient-free based, into account, and present the empirical comparison between their performance on a typical scenario, i.e. supervised learning. In our experiments, we found the gradient based optimizer provides a relatively better solution in each case. However, the gradient-free based optimizers have less running time than that of gradient-based optimizers. For small scale problems, the gradient-free optimizers such as COBYLA are more practical. We believe what we found in this work might provide some help for researchers interested in the related problems.