Toward improvement on Monte Carlo method by machine learning

2022 Fiscal Year Final Research Report

• Project/Area Number: 18K13548
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 15010:Theoretical studies related to particle-, nuclear-, cosmic ray and astro-physics
• Research Institution: Institute of Physical and Chemical Research
• Principal Investigator: Tanaka Akinori 国立研究開発法人理化学研究所, 革新知能統合研究センター, 上級研究員 (20791924)
• Project Period (FY): 2018-04-01 – 2023-03-31
• Keywords: 機械学習 / 場の量子論

Outline of Final Research Achievements

The purpose of this project is developing Markov Chain Monte-Carlo (MCMC) method accelerated by machine learning techniques. To achieve it, I focus on so-called Self Learning Monte-Carlo (SLMC) method that enables training during execution of MCMC and modifying the bias of the trained model in the generation step, and I apply this method to some physical theories. In the end of this project, we have achieved to make SLMC supported simulation of lattice gauge theory with dynamical fermions and non-commutative gauge group (SU(2)). We have checked its results and shown reduction of the autocorrelation.

Free Research Field

機械学習、数理物理

Academic Significance and Societal Importance of the Research Achievements

機械学習の手法を物理学に、より広くは科学分野に応用しようという動きが広がっているが、その場合におそらく最も重要な問題は、機械学習モデルの「間違い」を、精密さが要求される科学分野でどのように取り扱うかだと思われる。その点で、SLMCを用いたシミュレーションは（応用先の理論がわかっている場合には）、Metropolis-Hastingsテストを挟むことで正確さを担保するという意味で一つの解法を与えている。