Using machine learning techniques for early cost prediction of structural systems of buildings

It is desirable to predict construction costs in the early design stages in order tomake sure that target costs are met and competitive prices are realized. This study investigates the possibility of predicting the cost of construction early in the design phase by using machine learning (ML) techniques. To achieve this objective, artificialneural network (ANN) and case based reasoning (CBR) prediction models were developed in a spreadsheet-based format. An investigation of the impacts of weight generation methods on the ANN and CBR models was conducted. The performance of the ANN model was enhanced by experimenting with the weight generation methods of simplex optimization, back propagation training, and genetic algorithms while the CBR model was augmented by feature counting, gradient descent, genetic algorithms (GA), decision tree methods of binary-dtree, info-top and info-dtree. Cost data belonging to the superstructure of low-rise residential buildings were used to test these models. It was found that both approaches were capable of providing high prediction accuracy, 96% for ANN using simplex optimization for weight determination, and 84% for CBR using GA for attribute weight selection. A comparison of the Excel-based ANN and CBR models was made in terms of prediction accuracy, preprocessing effort, explanatory value, improvement potentials and ease of use. The study demonstrated the practicality of using spreadsheets in developing ANN and CBR models for use in construction management as well as the potential benefits of enhancing ANN and CBR models by using different weight generation methods.