Predicting strengths of concrete-type specimens using hybrid multilayer perceptrons with center-unified particle swarm optimization

This study explores the effectiveness of results obtained by using proposed hybrid multilayer perceptron (HMLP) networks to predict strength in concrete cylinders, reinforced-concrete deep beams, and reinforced-concrete squat walls. Such HMLP networks were designed to incorporate one linear and three high-order layer connections. Of the latter, one, employed only in the first layer connection, was derived from drawings referenced in the literature and two were developed by the author for this study. To calculate appropriate network coefficients, this study designed a center-unified particle swarm optimization (CUPSO) approach, composed of a center particle and global and local variants, which is quite effective for optimization tasks. This study gathered 103, 62, and 62 datasets, respectively, from drawings in three cases reported in the literature. Results, which showed that certain high order HMLP models perform better than their traditional counterpart, evidence the efficacy of proposed HMLP families. Each family, comprising high-order models and a linear counterpart, achieved results that were superior to those attained using traditional MLP networks only.