Exploring Symbolic Regression and Genetic Algorithms for Astronomical Object Classification

Fabio R. Llorella, José A. Cebrián · 2025

Read Exploring Symbolic Regression and Genetic Algorithms for Astronomical Object Classification by Fabio R. Llorella, José A. Cebrián (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-09220.txt.

Introduction

Fabio Llorella and Jose Cebrian wrote this astronomy paper, published in two thousand twenty-five, on sorting stars, galaxies, and quasars with formulas a human can read. Black-box models sort well but cannot say why, so the team aimed for an explicit expression instead.

They used a public data set of one hundred thousand sky objects from a sky survey. Symbolic regression searched for a math expression that splits the three classes, trained on a tenth of the data. A genetic algorithm then tuned the expression's settings. The final formula reached a Cohen's kappa of 0.81, a strong match with the true labels. The authors say the pairing of symbolic regression with genetic tuning offers accurate, open models. Such models let astronomers argue with the rule, not just trust it.

Catalog record

CREATORFabio R. Llorella, José A. Cebrián
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE45 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-09220

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Exploring Symbolic Regression and Genetic Algorithms for Astronomical Object Classification Fabio R. Llorella, José A. Cebrián Abstract This study explores the use of symbolic regression (SR) combined with genetic algorithms (GA) to classify astronomical objects. Using the SDSS17 dataset from Kaggle, which includes 100,000 observations of stars, galaxies, and quasars, we applied SR to 10% of the data to derive a mathematical expression capable of distinguishing these classes. A genetic algorithm was then employed to optimize the hyperparameters of the expression, refining the model’s performance. The final model achieved a Cohen’s kappa value of 0.81, indicating a strong agreement with true classifications. Our results demonstrate that the SR+GA approach can produce interpretable and accurate models for the classification of astronomical objects, offering a promising alternative to traditional black-box machine learning methods. keywords Symbolic Regression, Galaxy, Star, Quasar, Genetic Algorithm, Machine Learning 1 Introduction The classification of astronomical objects, such as galaxies, stars, and quasars, is a fundamental task in astrophysics and cosmology. Accurate classification allows researchers to analyze the structure and evolution of the universe, study the properties of different celestial bodies, and identify potential candidates for further investigation. The advent of large-scale sky surveys, particularly the Sloan Digital Sky Survey (SDSS), has revolutionized this field by providing vast datasets that catalog millions of celestial objects with high precision. The SDSS [ Kollmeier et al.(2019)Kollmeier, Anderson, Blanc, Blanton, Covey, Crane, Drory, Frinchaboy, Froning, Johnson, Kneib, Kreckel, Merloni, Pellegrini, Pogge, Ramirez, Rix, Sayres, Sánchez-Gallego, Shen, Tkachenko, Trump, Tuttle, Weijmans, Zasowski, Barbuy, Beaton, Bergemann, Bochanski, Brandt, Casey, Cherinka, Eracleous, Fan, García, Green, Hekker, Lane, Longa-Peña, Mathur, Meza, Minchev, Myers, Nidever, Nitschelm, O’Connell, Price-Whelan, Raddick, Rossi, Sankrit, Simon, Stutz, Ting, Trakhtenbrot, Weaver, Willmer, and Weinberg ] , for example, offers an extensive database that includes photometric and spectroscopic observations, which are invaluable for various astrophysical studies. However, the sheer volume of data generated by these surveys presents significant challenges. Traditional manual classification methods are no longer feasible, leading to the widespread adoption of automated techniques. Machine learning (ML) has emerged as a powerful tool for this purpose, enabling the rapid and efficient classification of astronomical objects based on their observed features [ Cheng et al.(2020)Cheng, Conselice, Aragón-Salamanca, Li, Bluck, Hartley, Annis, Brooks, Doel, García-Bellido, James, Kuehn, Kuropatkin, Smith, Sobreira, and Tarle , Zammit and Zarb Adami(2023) , Hinners et al.(2018)Hinners, Tat, and Thorp , Abraham et al.(2024)Abraham, Deville, and Kinariwala ] . Methods such as neural networks, support vector machines, and random forests have been extensively used in recent years, showing remarkable success in achieving high classification accuracy. Despite their effectiveness, these traditional machine learning approaches often suffer from a critical limitation: a lack of interpretability [ Räz(2024) ] . Many of these methods, particularly deep learning models, operate as ”black boxes,” making it difficult for researchers to understand how the model arrives at a particular decision [ Hassija et al.(2023)Hassija, Chamola, Mahapatra, Singal, Goel, Huang, Scardapane, Spinelli, Mahmud, and Hussain ] . In scientific disciplines like astronomy, where understanding the underlying mechanisms and relationships between variables is crucial, this opacity can be a significant drawback. The interpretability of a model not only enhances its transparency, but also allows scientists to gain new insights into the physical processes governing the behavior of celestial objects. To address these challenges, alternative approaches that balance accuracy with interpretability have been explored. Symbolic regression (SR) is one such technique that holds considerable promise. Unlike traditional machine learning methods, SR does not produce a single opaque model but rather derives explicit mathematical expressions that describe the relationships between variables. These expressions can then be analyzed and interpreted, offering insights into the underlying physical phenomena. SR has been successfully applied in various scientific fields, from physics to biology, where understanding the functional forms of relationships is as important as predicting outcomes [ Makke and Chawla(2024) ] . However, the effectiveness of symbolic regression can be limited by its computational complexity and the challenge of finding the best possible expression from an almost infinite space of potential equations. This is where genetic algorithms (GA), a type of evolutionary algorithm, come into play. GAs are heuristic search methods inspired by the principles of natural selection and genetics. They are particularly well-suited for optimizing complex, multidimensional problems, such as those encountered in symbolic regression. By simulating the process of evolution, GAs can efficiently explore the space of possible mathematical expressions, selecting and refining those that best fit the data. In this study, we propose a hybrid approach that combines symbolic regression with genetic algorithms (SR+GA) to classify galaxies, stars, and quasars. We used the SDSS17 dataset, a comprehensive collection of 100,000 observations of celestial objects, extracted from the Kaggle machine learning platform [ kag() ] . Our methodology involves using symbolic regression on a subset of the data (10%) to generate a candidate mathematical expression capable of distinguishing between these object classes. We then employ a genetic algorithm to optimize the free hyperparameters of this expression, fine-tuning the model to enhance its predictive accuracy. The use of SR+GA offers several advantages. First, the resulting models are highly interpretable, providing explicit expressions that can be directly analyzed and understood. This contrasts with traditional machine learning models, which often sacrifice interpretability for performance. Secondly, the optimization of hyperparameters via genetic algorithms ensures that the derived expressions are not only interpretable, but also competitive in terms of accuracy. This approach has the potential to generate models that are both scientifically meaningful and practically useful for large-scale astronomical classifications. Our results demonstrate the efficacy of this hybrid approach. The final model achieved a Cohen’s kappa value of 0.81, indicating a strong level of agreement with true classifications. This performance underscores the viability of using symbolic regression combined with genetic algorithms as a tool for astronomical object classification. Moreover, the interpretability of the resulting expressions provides valuable insight into the relationships between the observed features of galaxies, stars, and quasars. Before proceeding, it is important to review the key features of galaxies, stars, and quasars. • Galaxies are vast systems of stars, gas, dust, and dark matter, bound together by gravity. They come in various shapes and sizes, ranging from spiral galaxies like the Milky Way, characterized by their rotating disks and spiral arms, to elliptical galaxies, which are more three-dimensional and lack the well-defined structure seen in spirals. Galaxies are often categorized based on their morphology, star formation activity, and the presence of central supermassive black holes [ Sparke and Gallagher(2000) ] . • Stars are luminous spheres of plasma held together by their own gravity. They are the primary building blocks of galaxies and can vary widely in size, mass, temperature, and brightness. Stellar classification is typically based on spectral characteristics, which reflect a star’s temperature and composition. Main-sequence stars, giants, and dwarfs represent different stages in stellar evolution, each with distinct observational signatures [ Carroll and Ostlie(2017) ] . • Quasars (quasi-stellar objects) are among the most energetic and distant members of the universe. Powered by supermassive black holes at the centers of young galaxies, quasars emit enormous amounts of energy as the black holes accrete matter. They are often characterized by their bright, point-like appearance and extreme redshifts, which provide insights into the

Read the complete text