Open Conference Systems, StatPhys 27 Main Conference

Font Size: 
Improving Rare Event Sampling with Functional Estimators for Optimal Dynamics from Deep Learning
Tom Oakes, Adam Moss, Juan P Garrahan

##manager.scheduler.building##: Edificio San Jose
##manager.scheduler.room##: Auditorio 1
Date: 2019-07-11 11:30 AM – 11:45 AM
Last modified: 2019-06-10

Abstract


There has been a large amount of interest in improving the sampling of rare events when looking at the large deviations of stochastic many body systems. The focus has been on utilizing the so-called generalized Doob transformation, which provides the optimal dynamics to sample rare behavior in the longtime limit. However obtaining the Doob transformation is difficult as it requires exact diagonalization of the stochastic generator, and thus many approximation schemes have been developed, for example [1,2,3,4,5]. Here, we provide a means of arriving at an approximation for the transformation using the functional estimation provided by deep learning. This approach offers a clear advantage over current methods since it requires no prior knowledge of the system. Using a paradigmatic model, the fully packed classical dimer model Fig.1a, we show how the sampling can be improved via a neural network as a functional estimation Fig.1b.

[1] T.Oakes, S.Powell, C.Castelnovo, A.Lamacraft, and J.P.Garrahan, Phys. Rev. B. 98, 6, 064302, 2018

[2] T.Nemoto, R.L.Jack, and V.Lecomte, Phys. Rev. Lett. 118, 11, 115702, 2017

[3] K.Klymko, P.Geissler, J.P.Garrahan, and S.Whitelam, Phys. Rev. E 97, 6, 064302, 2018

[4] U.Ray, G.K.Chan, and D.T.Limmer, Phys. Rev. Lett. 120, 21, 210602

[5] G. Ferré, H. Touchette, J. Stat. Phys. 172, 6, 1525, 2018