Construction cost prediction based on adaptive boosting and artificial neural networks

Abstract

The artificial bee colony algorithm and multilayer error back-propagation neural networks commonly used in construction project cost forecasting suffer from slow training speeds and high costs. A combination of the beetle antennae search, support vector machines, adaptive boosting and genetic algorithms was proposed to solve these problems. Support vector machine optimisation was accomplished using the beetle antennae search algorithm. The enhanced genetic algorithm was then used directly to swap out the fit solutions for the unfit ones. One hundred projects completed during the last 3 years were chosen from a network integration database to serve as the training data set after developing the prediction model. Using actual cost information and trial and error, appropriate parameters were chosen, and combinations of algorithms were selected for comparison. The maximum relative error of the improved method was 9.01%, which was 34.68% lower than that of the baseline method, while the smallest relative error was 0.59%, which was 1.58% lower than that of the baseline method. The ’innovation of the study lies in the addition of the beetle antennae search algorithm and enhancement of the genetic algorithm. The former significantly increased the search efficiency of the network, while the latter generally increased the population fitness and mitigated the drawback of the genetic algorithm which was prone to local convergence.