Construction duration prediction using neural network methodology

The work presented in the thesis is concerned with the investigation and development of a neural network model for predicting construction duration. Previous works in construction duration prediction were reviewed and the limitations of the existing prediction models were highlighted. As a result, a newly emerging Al technique namely, neural networks is proposed as an alternative modelling environment.The fundamentals of neural networks are introduced and the selection of the most appropriate neural network paradigm is justified. The basics of the neural network paradigm selected i.e. backpropagation, are conveyed using a systems perspective, and a mathematical exposition, at the depth suitable for understanding the system Implementation and simulation, is given. Discussions to address issues of convergence focus on several aspects, including practical properties, of the algorithm. A synthesized methodology for developing neural network models is presented, using principled heuristic approaches.Guided by this methodology, a construction duration prediction model was developed and tested. The effects of network configuration and learning parameters, namely initial weight, learning rate coefficient and momentum, on network performance were investigated. The optimal network configuration for this construction duration problem was found to be 22:19:1 with initial weight = 0.1, = 1.0 and a = 0.9, giving predictive accuracy, stated in term of R2, of 96.6%. The work also interpreted the connection weights of the trained network, to demystify the black box image of the technique. Results of the neural network model were compared to a form of stepwise multiple regression and also with human experts predictions. The comparative study results indicated that the overall performance of the neural network exceeds that of both of this methods.From a neural network point of view, this research has established that there is a significant predictive relationship between a number of projects characteristic factors and its construction duration. It Is concluded that neural networks provide an effective alternative approach to organising available information on construction projects for use in early stage prediction of construction duration, and it is suggested that its application to cost prediction could be equally effective.