[9p-Z09-1~18] 23.1 Joint Session N "Informatics"

▲ [9p-Z09-11] Q-learning for square lattice Ising model

〇ChihChieh Chen¹, Kodai Shiba^1,2, Masaru Sogabe¹, Katsuyoshi Sakamoto², Tomah Sogabe^2,1 (1.Grid Inc., 2.UEC)

Keywords:Machine learning

The exponentially large degrees of freedom of quantum many-body systems make them difficult to simulate using computational approaches. Machine learning provides new opportunities to tackle the curse of dimensionality in quantum many-body problems. We use the Q-learning algorithm to study periodic Ising model on the 2-dimensional square lattice. In experiments on 2 by 2 and 4 by 4 lattices, the agent predicts correct antiferromagnetic ground states using Q-table. The learned Q-table has higher Q values for the degenerated ground states. We also implemented and tested the algorithm using Deep Q Network. We found some hyperparameters leading to stable learning curves and correct ground states. The results are compared and discussed.