Source-linked AI summary
Accelerate Monte Carlo Simulations with Restricted Boltzmann Machines
Li Huang, Lei Wang
TL;DR
Monte Carlo sampling can mix slowly near phase transitions, while designing generally effective updates is difficult. The paper trains a restricted Boltzmann machine on the physical distribution and uses its learned structure for Monte Carlo updates, keeping acceptance high across the Falicov-Kimball phase transition and reducing autocorrelation time by at least a factor of two.
Problem
Monte Carlo methods can mix slowly near phase transitions, and devising generally effective update strategies remains difficult.
Method
The approach trains an RBM as a proxy for the physical distribution and uses its latent structure to propose nonlocal Monte Carlo updates satisfying detailed balance.
Results
RBM updates maintain high acceptance across the Falicov-Kimball phase transition and reduce autocorrelation time by at least a factor of two.
Takeaways & Limitations
The RBM provides an adaptive way to guide more efficient Monte Carlo exploration of the sampled physical system.
Takeaways & Limitations
The approach’s scalability to larger and more complicated physical systems remains to be checked.
Abstract
from arXiv · showhide
Despite their exceptional flexibility and popularity, the Monte Carlo methods often suffer from slow mixing times for challenging statistical physics problems. We present a general strategy to overcome this difficulty by adopting ideas and techniques from the machine learning community. We fit the unnormalized probability of the physical model to a feedforward neural network and reinterpret the architecture as a restricted Boltzmann machine. Then, exploiting its feature detection ability, we utilize the restricted Boltzmann machine for efficient Monte Carlo updates and to speed up the simulation of the original physical system. We implement these ideas for the Falicov-Kimball model and demonstrate improved acceptance ratio and autocorrelation time near the phase transition point.
Learned weights at T/t = 0.13
At T/t = 0.13, the RBM’s learned weights change drastically near the critical temperature. Relative to T/t = 0.15, more hidden neurons control extended visible-variable regions, while the low-temperature checkerboard pattern is more visible.
- The RBM’s learned weights change drastically near the critical temperature.
- Compared with T/t = 0.15, more hidden neurons control extended regions of the visible variables.This indicates an enlarged correlation length at the lower temperature.
- The checkerboard pattern of the low-temperature phase is more visible.
Proof of the detailed balance conditions
The section proves that the RBM simulation shown in Fig. 4 satisfies detailed balance, covering both block Gibbs sampling and Gibbs sampling with additional Metropolis updates for hidden variables.
- Proof of the detailed balance conditions: The simulation of the RBM shown in Fig. 4 satisfies the detailed balance condition.The proof addresses the simulation procedure directly.
- Proof of the detailed balance conditions: The proof treats block Gibbs sampling as shown in Fig. 4(a).This is one of the two update schemes considered.
- Proof of the detailed balance conditions: It also treats a Gibbs sampler with additional Metropolis steps for hidden variables as shown in Fig. 4(b).The hidden-variable Metropolis update obeys T(h →h′)/T(h′ → h) = p(h′)/p(h), and the proof generalizes to compositions of several such updates.