Automatic Generation of Quantum Neural Networks with Reinforcement Learning*

Quantum Machine Learning (QML) is at the intersection of quantum computing and machine learning. QML models are often of a hybrid quantum-classical nature, where only a subroutine of the algorithm is executed on a quantum computer. In the area of practical QML, numerous questions arise regarding the trainability and generalization of the models. The essential building block of any data-driven QML model is the encoding strategy used to embed the data into the quantum state space. Depending on the specific model, this encoding is often referred to as a parameterized quantum circuit or equivalently as a Quantum Neural Network (QNN).

Based on the principles of classical machine learning, it is well understood that the optimal performance of a data-driven algorithm depends on its tailored design for the problem at hand. However, orchestrating this architecture search can be a complex and labor-intensive process, typically requiring a deep understanding of the specific problem domain. While contemporary QML research has produced advanced QNN architectures, these designs are rarely tailored to the problem at hand. Consequently, this lack of problem-specific tailoring can lead to reduced model performance and trainability.

In this work, we discuss the automatic generation of QNNs with a model-based reinforcement learning approach. This allows us to build problem specific circuit architectures, which we subsequently benchmark against contemporary circuits used in the literature for a variety of problems. We apply our approach to problems based on quantum data and on classical data. Our results show that tailoring the QNNs to the problem at hand can significantly improve the performance of the algorithm.