The ANALYSIS OF THE EFFECT OF ARTIFICIAL NEURAL NETWORK ARCHITECTURE AND TRAINING PARAMETERS ON REGRESSION AND CLASSIFICATION PERFORMANCE
ANALYSIS OF THE EFFECT OF ARTIFICIAL NEURAL NETWORK ARCHITECTURE AND TRAINING PARAMETERS ON REGRESSION AND CLASSIFICATION PERFORMANCE
DOI:
https://doi.org/10.54317/elka.v7i2.1013Abstract
Artificial Neural Network (ANN) is one of the computational models used in machine learning to recognize patterns and solve prediction and classification problems. This study aims to analyze the effect of hidden layer configuration, number of neurons, activation functions, and learning rate on ANN performance for regression and classification tasks. The practical experiments were conducted using the Boston Housing dataset for regression and the Iris dataset for classification through Google Colab. The experimental results show that the number of hidden layers affects model performance, but increasing the number of layers does not always produce better results. In the regression experiment, the best neuron configuration achieved an R² value of 0.9930 with an MAE of 0.63, while the best activation function configuration achieved an R² value of 0.9958 with an MAE of 0.50. The best learning rate in the regression experiment was 0.005, producing an R² of 0.9961 and MAE of 0.44. In the classification experiment, the best neuron configuration and several learning rate configurations achieved an accuracy of 1.0000. The final activation function experiment showed that Softmax achieved the highest accuracy of 1.0000. These results indicate that appropriate selection of ANN architecture and training parameters is important for obtaining optimal model performance.





