Application of artificial neural networks to cost and duration forecasting for buildings | |
| Author | Sdhabhon Bhokha |
| Call Number | AIT Diss. no. ST-98-02 |
| Subject(s) | Neural networks (Computer science) Building--Estimates |
| Note | A Dissertation submitted in partial fulfillment of requirements for the degree of Doctor of Engineering, School of Civil Engineering |
| Publisher | Asian Institute of Technology |
| Abstract | This dissertation deals with the application of Artificial Neural Networks (ANNs) to pre-design cost and duration forecasting for buildings. The accuracy of forecasting is dependent on the class of estimate, e.g. rough, target, budget, approximate, sketch, screening, conceptual, study, preliminary, and feasibility. Lacking complete information at the pre-design stage, a professional may estimate the construction costs and duration by using knowledge, past experience from completed similar projects, and personal judgement. This process is intuitive and fast, but prone to errors. It is however, acceptable and sufficient for various planning and budgeting purposes. The task is sometimes cost and time consuming and is made difficult by complexity of project and uncertainties in the construction industry. ANN is a new and promising approach to handle tasks of this nature. When usually little known information is input, probable answers of construction costs and duration can be drawn promptly from the embedded knowledge of a trained network. In addition, networks can generalize to give meaningful solutions even when the inputs contain errors or are incomplete. Independent variables found to have major effects on construction costs (in Baht/m2 ) are: 1) building function; 2) structural system; 3) complexity of foundation; 4) height; 5) exterior finishing; 6) decorating quality; and 7) accessibility to the site. They are all binary attributes. Two real variables representing price indices of major construction materials are proposed to incorporate changes in prices. Three dependent variables or outputs of real values are provided for the cost model. Almost all the same inputs are also used to forecast the construction duration (in months) except that the three price indices are superseded by a real-value node representing functional area. One output node of real value is provided as construction technology has hardly changed with the passage of time. The three-layered back-propagation networks developed are fully connected, written in c++ language, and implemented on a Pentium-75 based micro-computer. Generalized Delta Rule (GDR) is used as learning algorithm. One hundred and thirty-six samples of buildings which were built in the Greater Bangkok area during the period 1987-1995 are divided into two equal parts, and fed to train and test the networks. The best networks for cost and duration forecasting consist of { 12, 6, 3}, and { 11, 6, 1} nodes on each layer, respectively. Learning rate at 0.6, and null momentum are used for both networks. Average errors on the test samples are 6.7%, and 18.2% for cost and time respectively. The results show that ANN is a suitable approach for forecasting construction cost and duration at the pre-design phase. It is a distribution free model which uses simple mathematics. At this beginning stage of development, it is recommended that ANN should be used in conjunction with the existing methods of forecasting. The developed networks need periodic maintenance so that they could efficiently solve real world problems in the future. |
| Year | 1998 |
| Type | Dissertation |
| School | School of Civil Engineering |
| Department | Department of Civil and Infrastucture Engineering (DCIE) |
| Academic Program/FoS | Structural Engineering (STE) /Former Name = Structural Engineering and Construction (ST) |
| Chairperson(s) | Ogunlana, Stephen 0. ; |
| Examination Committee(s) | Worsak Kanok-Nukulchai ;Sandananda, Ramakoti ;Nattawuth Udayasen ;Raftery, John J. ; |
| Degree | Thesis (Ph.D) - Asian Institute of Technology, 1998 |