Abstract
The application of neural networks in solving quantum many-body systems has been a hot point in the research of theoretical quantum physics recently. Despite its success in dealing with some simple quantum systems, the neural network methods encounter difficulties in finding correct sign structures in frustrated systems. In this thesis, we use a numerical optimisation method named evolution strategy (ES) to train the network, which is a practice not attempted in previous studies. We show that ES can optimise a sign network as an independent or auxiliary sign structure of the whole wave function. The performance of ES is tested on the J1-J2 model on square and pyrochlore lattices. Our method based on ES gives the best neural network result in these frustrated systems compared with previous studies.