Reducing Autocorrelation Times in Lattice Simulations with Generative Adversarial Networks
Authors:
Preprint
Abstract
Short autocorrelation times are essential for a reliable error assessment in Monte Carlo simulations of lattice systems. A generative adversarial network (GAN) can provide independent samples, thereby eliminating autocorrelations in the Markov chain. We address the question of statistical accuracy by implementing GANs as an overrelaxation step, incorporated into a traditional hybrid Monte Carlo algorithm. This allows for a sensible numerical assessment of ergodicity and consistency. Results for scalar $\phi^4$-theory in two dimensions are presented. We achieve a significant reduction of autocorrelations while accurately reproducing the correct statistics. We discuss possible improvements as well as solutions to persisting issues and outline strategies towards the application to gauge theory and critical slowing down.Keywords
All Related Versions
This publication has 26 references indexed in Scilit:
- Self-learning Monte Carlo method with Behler-Parrinello neural networksPhysical Review B, 2020
- Neural-network quantum state tomographyNature Physics, 2018
- Sampling algorithms for validation of supervised learning models for Ising-like systemsJournal of Computational Physics, 2017
- Efficient representation of quantum many-body states with deep neural networksNature Communications, 2017
- Machine learning quantum phases of matter beyond the fermion sign problemScientific Reports, 2017
- Learning phase transitions by confusionNature Physics, 2017
- Machine learning phases of matterNature Physics, 2017
- Learning thermodynamics with Boltzmann machinesPhysical Review B, 2016
- Gauge cooling in complex Langevin for lattice QCD with heavy quarksPhysics Letters B, 2013
- Equation of State Calculations by Fast Computing MachinesThe Journal of Chemical Physics, 1953