Optimization of physical quantities in the autoencoder latent space

Abstract

We propose a strategy for optimizing physical quantities based on exploring in the latent space of a variational autoencoder (VAE). We train a VAE model using various spin configurations formed on a two-dimensional chiral magnetic system. Three optimization algorithms are used to explore the latent space of the trained VAE. The first algorithm, the single-code modification algorithm, is designed for improving the local energetic stability of spin configurations to generate physically plausible spin states. The other two algorithms, the genetic algorithm and the stochastic algorithm, aim to optimize the global physical quantities, such as topological index, magnetization, energy, and directional correlation. The advantage of our method is that various optimization algorithms can be applied in the latent space containing the abstracted representation constructed by the trained VAE model. Our method based on latent space exploration is utilized for efficient physical quantity optimization.

A Stage-Specific Gene Expression Framework for Promoting Genetic Algorithm

“…Wu et al proposed a new genetic algorithm which can divide the initial largescale problem into several small-scale subproblems, then use a hybrid search strategy to solve the subproblem respectively [12]. Minh et al proposed a strategy for optimizing physical quantities based on exploring the latent space of a variational autoencoder (VAE) [13]. This algorithm trains a VAE using an existing dataset whose elements have good performance for fitness functions.…”

Optimizing training trajectories in variational autoencoders via latent Bayesian optimization approach *

“…Generally, a hyperparameter is a parameter which controls the learning process of the ML models, and the hyperparameter tuning (or optimization) is the problem of choosing a set of optimal values for those hyperparameters to optimize the learning process. Extensive effort has been dedicated towards optimal tuning of ML models, with different optimization techniques such as gradient based method, genetic algorithm (GA), Bayesian optimization (BO) etc [13][14][15][16][17][18][19]. In the process of tuning ML models where the training cost is computationally high, any exhaustive or manual parameter space search is a highly non-desirable approach.…”