PSI - Issue 70

Ch. Vihas et al. / Procedia Structural Integrity 70 (2025) 461–468

464

2.2 Model Training

The machine learning models that predict concrete compressive strength are typically regression-based and designed to predict continuous values. The ML models selected for the study are RF, XG Boost, ANN, and SVM. Initially, the ML models are trained on the training dataset in the forward propagation. The backward propagation of the weights is adjusted using the optimization algorithm. The hyperparameters of each model were adjusted to enhance the efficiency of the machine learning model. The ML models were evaluated using various metrics coefficient of determinations (R²), MSE, MAE, and RMSE (Kaveh, A., & Khavaninzadeh, N. (2023)).

2.2.1 Random Forest

An RF is a machine learning model that uses the results of many decision trees to train the model. The final predictions are the mean of predictions made by the trees for regression tasks. It uses bagging to generate multiple datasets by random sampling with replacements from the original training dataset. Each tree is trained on a different subset of the data, which helps reduce overfitting. When constructing the decision tree, RF selects a subset of features and splits at its each node, thereby further reducing the correlation between the trees.

2.2.2 XG Boost

XG Boost is an implementation of gradient boosting that incorporates enhancements, making it faster and more accurate in predicting outcomes. The base learner in XGBoost is a decision. Each subsequent tree attempts to minimize the error of the previous tree by focusing on the residuals. Essentially, the model constructs trees to correct the mistakes made by earlier trees in the ensemble. XGBoost employs a custom objective function that combines the loss function with a regularization that penalizes model, such as the depth of the trees and the number of leaves. XGBoost introduces two regularization terms to reduce overfitting: L1 regularization (Lasso), which regularizes the absolute values of the weights, and L2 regularization (Ridge), which regularizes the squared values of the weights. This helps to prevent overfitting and enhances model stability.

2.2.3 Artificial Neural Networking (ANN)

ANN is a machine learning model influenced by the human nervous system. It consists of an input, hidden, and output layer. The input layer consists of neurons made of features in the dataset. Hidden layers form a deep learning model from one or more layers to utilize activation functions. The output layer consists of the predicted results from the hidden layer. It typically consists of one neuron with either no activation function or a linear activation function.

2.2.4 Support Vector Machine (SVM)

The Support Vector Machine (SVM) is a machine learning model that identifies the hyperplane which seperates the data into distinct classes. A hyperplane is a decision boundary that will separate various classes in the space of features. The Support Vectors are points that are closer to the hyperplane. These points are crucial because they affect the position and orientation of the hyperplane. The Kernel Function will be applied to transform data into a higher dimension, making it separable linearly if it is not in its original form. The SVM algorithm aims to maximize the margin, the displacement between the hyperplane and the nearest vectors. The Decision Function is a mathematical function that classifies new data points based on their position relative to the hyperplane .