<![CDATA[Abstract: Technological developments have driven the application of Deep Learning, particularly Artificial Neural Networks (ANN), in solving regression and classification tasks. ANNs consist of interconnected artificial neurons that are effective in classification, prediction, and pattern recognition. This study aims to build and analyze the impact of variations in Artificial Neural Network (ANN) architecture on prediction accuracy, using a quantitative experimental method with a controlled randomized design. The BostonHousing.csv dataset was used for regression, and the Iris.csv dataset for classification. Regression evaluation uses MSE, R², and MAE; while classification uses accuracy and confusion matrix. The best regression results were obtained from the 4-hidden-layer architecture (512, 256, 128, 64), with ReLU, sigmoid, tanh, and ReLU activation functions, a learning rate of 0.001, achieving an R² of 89.2% and an MAE of 1.94. For classification, the best architecture (128, 64, 32, 16) with a softmax output yielded an accuracy of 99.8% and a model accuracy of 100%. Keywords – deep learning, artificial neural network, regression, classification, ANN architecture]]>
Copyrights © 2025
