Hybridized artificial neural network with metaheuristic algorithms for bearing capacity prediction

• Published in: Ain Shams Engineering Journal Vol. 14; no. 5; p. 101980
• Main Author: Abdullahi Mu'azu, Mohammed
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.05.2023; Elsevier
• Subjects: Artificial neural network; Bearing capacity analysis; Metaheuristic algorithms; Settlement measurement; Settlement measurement; Artificial neural network; Bearing capacity analysis; Metaheuristic algorithms
• ISSN: 2090-4479
• DOI: 10.1016/j.asej.2022.101980

Choosing a suitable training technique out of so many is a critical step, and its importance cannot be overemphasized, especially for a problem such as bearing capacity (BC) analysis. The idea is to optimize the configuration of the artificial neural network (ANN) hybridized with the cuttlefish optimization algorithm (CFOA), electrostatic discharge algorithm (ESDA), and Henry gas solubility optimization algorithm (HGSOA), and sine cosine algorithm (SCA) algorithms for soil BC analysis. Applied stress (X7) is discovered as the most important input factor through an unbiased predictor. The training errors are 0.0119, 0.0094, 0.0135, and 0.0139, and the training RMSE values are 0.01560, 0137, 0.0178, and 0.0175 for CFOA–ANN, ESDA–ANN, HGSO–ANN and SCA–ANN, respectively. Also, the testing errors are 00117, 0.0106, 0.0155, and 00151, and the testing RMSE is 0.0159, 0.0153, 0.0202, and 0.0192 for CFOA–ANN, ESDA–ANN, HGSO–ANN and SCA–ANN, respectively. According to RMSE values, the HGSO-ANN has the lowest value and ESDA-ANN has the highest value of RMSE. In compairing the MAE values, the best method was SCA-ANN and the worst one was ESDA-ANN. The second and third methods in presenting the MAE was HGSO-ANN and CFOA-ANN, respectively. While Pearson correlation (Rp) factors in the testing phase is 0.8809, 0.8904, 0.8136, and 0.8252, and in the training phase is 0.8757, 0.9048, 0.8370, and 0.8406 for CFOA–ANN, ESDA–ANN, HGSO–ANN and SCA–ANN respectively. In evaluating the Rp value, HGSO-ANN presented the best value and ESDA-ANN has highest value of Rp. A potentially applicable equation is developed for possible utilization in a compatible practical scenario for ESDA–ANN model as the most accurate. It could be employed as a less time-consuming yet, precise replacement for the conventional approach for BC analysis. As the results indicate, due to the lowest value of RMSE and MAE, and the highest value of Rp, ESDA–ANN model approximated the training and testing datasets slightly better than the others in terms of RMSE, MAE and Rp. After that, CFOA-ANN, SCA-ANN, and HGO-ANN models was the second, third and forth methods in predicting the bearing capacity, respectively.