An artificial neural network model for estimating road construction cost in the Philippines

• Main Authors: Cristobal, Jerald B., Hao, Sara Eunice M., Rabino, Rachelle Marie G., Revalde, Fulgencio M., Jr.
• Format: text
• Published: Animo Repository 2015
• Subjects: Roads > Design and construction > Estimates > Automation; Roads > Design and construction > Estimates > Philippines > Automation
• Online Access: https://animorepository.dlsu.edu.ph/etd_honors/385
• Institution: De La Salle University

Estimating the total project cost early on the project conceptualization phase is very important to the client and contractor however, not all factors are given that early on the project as several external factors are still unforeseen. When the stakeholders need an immediate estimate of the budget for the project, in-depth analysis of the costing of the project may take a lot of time, sacrificing resources for feasibility studies. In the Philippine context, there are developments yet to come when it comes to fast and efficient early cost estimation. This paper adapted artificial neural network, a modern technology that adapts the ability of the brain to learn through pattern recognition. The objective of this paper is to determine which road factors are the most significant to the total cost and how many hidden layers would one need to arrive at a near estimate to the project's total cost. The initial assumed factors were fifteen road factors, through when inspected of their coefficient of correlation and analyzed through learned road knowledge, they were narrowed down to seven, namely: project length, project duration, average site clearing and grubbing (ASCG), earthworks volume (EWV), surface class, water body, and consumer price index (CPI). Forty-one (41) road projects were obtained from the Department of Public Works and Highways were simulted using the following ANN programs: MATLAB, Microsoft Excel Solver, NeuroShell 2 and NeuroSolutions. From the results of the simulation with four, five, six and fifteen hidden neurons, the automated progam NeuroShell 2 produced the lowest mean squared error and the neaest coefficient of linear regression to one, deeming it the program to validate the architectures with. Through two simulation using forty-two (42) sample projects and six (6) test projects, the results were consistent. The network architecture that produced the least MSE and closest r-value to one has six hidden nodes, having the 7- 6-1 ANN architecture as the best ANN architecture to estimate