MODIFIED HIGH CONVERGENCE GENETIC ALGORITHM FOR NUMERICAL OPTIMIZATION TASKS SOLVING IN THE FIELD OF THZ-IR SPECTROSCOPY

Abstract

The Modified High Convergence Genetic Algorithm (HCGA) for solving highly parametrized, unregularized, and non-convex mathematical tasks is presented. HCGA is the superposition of the genetic algorithm (GA) without modification and Levenberg-Marquardt iteration algorithm. The key feature of HCGA is the partial optimization of an individual's genome using LM algorithm. Each individual is presented with its genome. An individual's genome is a set of model parameters to optimize. HCGA operates with a model — a parametrized mathematical function, and a target function as a quality metric. The architecture of HCGA combines the advantages of GA and LM algorithm. In relation to GA, the key advantage is the stochastic feature, which makes it possible to not stop in a local extremum region of a target function. The main advantage of LM algorithm is the high speed of convergence. As the result, HCGA effectively scans a surface of a target function, finds a set of extremum regions, and defines a global extremum.
The characteristic for the tasks of optics and spectroscopy model function is provided. Among these tasks are the following: molecular potential energy surface approximation, estimation of rotational constants with THz-IR spectra, evaluation of optical features of an isotropic medium using THz-TDS, and etc. Using the model function and RMSE as the target function, the comparative analysis between HCGA, GA, LM, and Adam algorithms was conducted.
This research was funded by the Ministry of Science and Higher Education of the Russian Federation grant number 075-15-2024-557 dated 04/25/2024.

Speaker

Tretyakov Akim Konstantinovich
Tomsk State University
Russia

Discussion

Ask question