Hybrid RSM–GA Hyperparameter Tuning of Artificial Neural Networks for Academic Performance Prediction

Abstract

The development of artificial intelligence has encouraged the use of Artificial Neural Networks (ANNs) foracademic performance prediction in competency-based education. However, in automotive vocational education, ANNhyperparameters are commonly determined through trial-and-error procedures, which may produce unstable and suboptimalmodels. This study proposes an integrated Central Composite Design (CCD), Response Surface Methodology (RSM),and Genetic Algorithm (GA) framework to optimize ANN hyperparameters for predicting the academic performance ofAutomotive Vocational Education (AVE) students at Vocational High School (VHS) NU 1 Karanggeneng, Lamongan.The quadratic RSM model explained 82.09% of the response variation and identified the learning algorithm as the mostinfluential optimization factor. At the original optimization-response scale, the RSM–GA procedure produced an optimal ANN configuration with three hidden layers, 20 neurons, a tansig–purelin transfer-function combination, the trainlm learning algorithm, a learning rate of 0.005, and 200 epochs, achieving an MSE of 0.07 and an S/N Ratio of 22.94. In the normalized response-level benchmark, the Opt RSM–GA ANN obtained the most favorable normalized repeated MSE response, with a Mean MSE of 0.012, SD MSE of 0.0019, and S/N Ratio of 38.65. These findings indicate that the CCD–RSM–GA workflow provides a structured and reproducible approach for ANN hyperparameter optimization. Broader validation using larger, multi-school, and multi-cohort datasets is still required before practical implementation.