Efficient boosting-based algorithms for shear strength prediction of squat RC walls

Reinforced concrete shear walls have been considered as an effective structural system due to their optimal cost and great behavior in resisting lateral loads. For the slender type of these walls, failure modes are mainly related to flexure, while for the squat type with height-to-length ratios less than two, shear is the dominant factor. Thus, accurate estimation of shear strength for squat shear walls is necessary for design applications and can also be complex due to the various effective parameters. In order to address this issue, first a comprehensive dataset with 558 samples of squat shear walls is conducted, and three hybrid models consisting of genetic algorithms and boosting-based ensemble learning methods, i.e., XGBoost, CatBoost, and LightGBM, are used for estimation of shear strength. The results showed high prediction accuracy, with a coefficient of determination of at least 98.6% for all three models. Genetic algorithm has been proven to be an effective method for tuning boosting-based algorithms compared to manual testing. In addition, the results of the algorithms are compared to their default hyperparameters and other conventional regression Models. Also, multicollinearity and principal component analysis (PCA) were studied. Furthermore, the performance of three tuned models is compared with that of a mechanics-based semi-empirical model and other genetic programming (GP)-based models. Finally, parametric and sensitivity analyses were performed, to demonstrate the ability of the models to identify the most critical parameters with significant influence on shear strength.