[Paper Review] Selecting Initial States from Genetic Tempering for Efficient Monte Carlo Sampling

Thomas E. Baker|arXiv (Cornell University)|Jan 29, 2018

Neural Networks and Applications3 citations

TL;DR

This paper proposes Genetic Tempering on a GPU (GT-GPU), an efficient Monte Carlo sampling method that uses independent Metropolis-Hastings threads on a GPU, selecting initial states via a genetic algorithm to minimize autocorrelation and accelerate convergence. It achieves high-precision results in seconds—matching or exceeding traditional methods—by sampling multiple minima simultaneously and selecting threads from within hysteresis loops of partially relaxed states.

ABSTRACT

An alternative to Monte Carlo techniques requiring large sampling times is presented here. Ideas from a genetic algorithm are used to select the best initial states from many independent, parallel Metropolis-Hastings iterations that are run on a single graphics processing unit. This algorithm represents the idealized limit of the parallel tempering method and, if the threads are selected perfectly, this algorithm converges without any Monte Carlo iterations--although some are required in practice. Models tested here (Ising, anti-ferromagnetic Kagome, and random-bond Ising) are sampled on a time scale of seconds and with a small uncertainty that is free from auto-correlation.

Motivation & Objective

• To address the inefficiency and long sampling times of traditional Monte Carlo methods due to slow equilibration and autocorrelation.
• To leverage the massive parallelism of GPUs to run independent Metropolis-Hastings threads simultaneously, improving sampling efficiency.
• To reduce systematic errors and improve statistical precision by selecting threads based on their position within hysteresis loops from partially relaxed samples.
• To achieve accurate ensemble statistics with minimal Monte Carlo iterations by focusing on initial state selection rather than long thermalization.
• To demonstrate the method's effectiveness on frustrated and disordered spin models, including Ising, anti-ferromagnetic Kagome, and random-bond Ising models.

Proposed method

• Run multiple independent Metropolis-Hastings simulations in parallel on GPU threads, each with a unique initial spin configuration.
• Use a genetic algorithm-inspired selection process to choose the best-performing threads from a hysteresis loop generated during partial relaxation.
• Select threads based on their energy and magnetization values within the hysteresis loop to ensure diverse sampling of minima and reduced autocorrelation.
• Apply the selection criterion to propagate only the most representative threads forward, effectively mimicking the idealized limit of parallel tempering.
• Use single-precision arithmetic on the GPU to maximize performance while minimizing memory overhead and overflow risks.
• Buffer random numbers over η cycles to reduce on-the-fly generation costs, maintaining high throughput during simulations.

Experimental results

Research questions

• RQ1 Can independent Metropolis-Hastings threads on a GPU, when intelligently selected, outperform traditional Monte Carlo methods in sampling efficiency and accuracy?
• RQ2 Does selecting initial states from within a hysteresis loop of partially relaxed samples significantly reduce autocorrelation and improve statistical precision?
• RQ3 To what extent can GT-GPU achieve high-precision results with minimal Monte Carlo iterations, even in the presence of energy barriers and frustration?
• RQ4 How does GT-GPU perform on disordered and frustrated spin systems such as the random-bond Ising model compared to established benchmarks?
• RQ5 Can the method achieve accurate phase transition detection (e.g., critical temperature) with minimal computational cost and without long thermalization?

Key findings

• GT-GPU achieved a critical temperature estimate of $ T_c^{ ext{GT}} = 2.267 \pm 0.005 $ for the 2D Ising model, closely matching the exact value of $ T_c \approx 2.269 $.
• For the random-bond Ising model, GT-GPU estimated $ p_c^{ ext{GT}} = 0.1097 \pm 5 \times 10^{-4} $, consistent with the reference value of $ p_c \approx 0.1094 \pm 2 \times 10^{-4} $.
• The method achieved uncertainty levels comparable to $ \mathcal{N}_{\text{MC}} = 10^6 $ iterations using only $ \mathcal{J} = 10^3 $ threads and $ \mathcal{N}_{\text{GPU}} = 10^4 $ GPU iterations.
• GT-GPU reduced statistical uncertainty by equal contributions from the number of threads $ \mathcal{J} $ and GPU iterations $ \mathcal{N}_{\text{GPU}} $, as per the uncertainty formula $ \sigma \propto 1/\sqrt{\mathcal{J} \mathcal{N}_{\text{GPU}}} $.
• The method achieved high-precision results in under 20 minutes per system size, with negligible autocorrelation effects due to thread selection from hysteresis loops.
• The approach demonstrated robustness across models with frustration (Kagome) and disorder (RBIM), showing consistent accuracy and fast convergence.