Recognizing good variational quantum circuits with Monte Carlo Tree Search

Many investigators have recently turned to the study of quantum architecture search since it is laborious to manually design a high-performing quantum model and corresponding training strategies. For some tasks, it is more realistic in practice to search for a model architecture. In this paper, we introduce the Monte Carlo Tree Search algorithm which has achieved great success in classical neural architecture search to find good variational quantum circuits for two real-world tasks of ground state energy estimations and multimodal fusion. We adapt the Monte Carlo Tree Search to the quantum scenario by considering more sophisticated classifiers within the tree nodes to partition the search space into several subregions based on the model performance. The experimental results indicate that our proposed method has the ability to recognize good models from the vast search space in both tasks. More importantly, the discovered variational quantum circuits demonstrate their advantages in fusing multimodal features under the comprehensive consideration of parameter number and performance.