Educational Evolutionary Neural Architecture Search for Time Series Prediction

Martha Escalona-Llaguno, Sergio Sarmiento-Rosales

2024

Abstract

This paper presents a sophisticated tool designed to teach Evolutionary Neural Architecture Search (ENAS) in time series analysis. The goal is to create a flexible and modular algorithm that helps to understand evolutionary algorithms in the context of neural architecture optimization. The tool allows parameter tuning and search space exploration. Its initial setup can include a population size of 20 individuals, spanning 5 generations, with an elitism rate of 20% and crossover and mutation probabilities set to 90% and 10%, respectively. However, these hyperparameters are completely modular, so that the effect on the algorithm can be studied. The parameter ranges from 1 to 20 for neurons and delays. The neural networks are extensively trained using the MATLAB narnet function and the PJM hourly energy consumption dataset, which is split into 70% for training, 15% for validation, and 15% for testing. The goal is to maximize the correlation coefficient r obtained from the test dataset. This approach offers an interactive platform for experimentation and learning about the evolutionary process of neural architectures, thus improving the understanding of evolutionary algorithms applied to Neural Architecture Search (NAS). Our experiments show efficiency due to the limited search space and the absence of specialized hardware requirements such as GPU, making it accessible and practical for educational and research environments. Using only an AMD Ryzen 7 7800X3D CPU, all architectures within the search space were trained in less than 3 hours, demonstrating the agility of ENAS in various configurations and its effectiveness in facilitating practical understanding of the evolutionary process in NAS. All datasets, tutorials and essential codes to apply this work are publicly accessible at the following link: https://github.com/SergioSarmientoRosales/ENAS-Time-Series.

Paper Citation

in Harvard Style

Escalona-Llaguno M. and Sarmiento-Rosales S. (2024). Educational Evolutionary Neural Architecture Search for Time Series Prediction. In Proceedings of the 16th International Joint Conference on Computational Intelligence - Volume 1: ECTA; ISBN 978-989-758-721-4, SciTePress, pages 234-241. DOI: 10.5220/0012948900003837

in Bibtex Style

@conference{ecta24,
author={Martha Escalona-Llaguno and Sergio Sarmiento-Rosales},
title={Educational Evolutionary Neural Architecture Search for Time Series Prediction},
booktitle={Proceedings of the 16th International Joint Conference on Computational Intelligence - Volume 1: ECTA},
year={2024},
pages={234-241},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0012948900003837},
isbn={978-989-758-721-4},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 16th International Joint Conference on Computational Intelligence - Volume 1: ECTA
TI - Educational Evolutionary Neural Architecture Search for Time Series Prediction
SN - 978-989-758-721-4
AU - Escalona-Llaguno M.
AU - Sarmiento-Rosales S.
PY - 2024
SP - 234
EP - 241
DO - 10.5220/0012948900003837
PB - SciTePress