Implementation of Artificial Neural Network on Regression Analysis

Artificial neural network (ANN) works as a very effective tool in both classification and regression problem. The main advantage lies in the fact that it can draw fine distinctions, patterns, or hidden information of data explicitly without complex mathematical considerations. This study aimed to use the neural network on regression problem to achieve a better performance with a lower computational cost. After constructing the initial architecture, we added randomized weights and biases to the model. The performance of the model is evaluated on a well-known benchmark regression dataset named Auto MPG. The datasets were split among train, validation, and testing in a ratio of 50%-30%-20%. After subsequent tuning of hyperparameters, the refined architecture was able to perform predicting regression variable with a great efficiency and less computational time. The metrics of accuracy were taken as mean absolute error (MAE), root mean square error (RMSE) and mean absolute percentage of error (MAPE). Comparative study with the others state-of-the-art metaheuristic algorithms reveals the effectiveness of the proposed ANN. Evaluating the experimental results, it was obvious that, most often ANN did lead the other approaches with all the metrics of accuracy with a least computational time. It is also seen that, in some cases, though ANN revealed a very competitive result with other approaches, however the reduction in computational cost overcame the drawbacks and put neural network in leading.