| Summary: | 碩士 === 國立高雄第一科技大學 === 營建工程所 === 97 === Construction projects are often quite complex with a lot of uncertainty about the required duration. For a project planner, it is important to ensure project completion within the deadline. The critical path method (CPM) uses single-time estimation and does not consider time variability. On the other hand, the program evaluation and review technique (PERT) produces estimates of project duration based on the path with the longest mean time but ignores the influence of other paths. For a project network with two or more paths with similar lengths, PERT will not be able to give accurate estimates of the mean duration and the duration for a given confidence level. Where the probability distributions for time of two paths are close or have large variances, the deviation of the PERT estimate becomes evident.
In the construction environment, the activity time can be seen as a random variable, so the project duration can also be seen as a random variable. It is rather complicated to obtain the probability distribution for project duration using analytical or numerical methods, while simulation through a large number of samplings can only produce an approximate solution. Therefore, this research proposes the similar paths network evaluation technique (SPNET) as an improved method and uses it to provide training and testing data for developing neural networks. The neural networks developed are intended for estimating the mean duration and the duration for a given confidence level for a project network with two as well as three similar paths, achieving the mean absolute percentage errors of 0.314%, 0.554%, 0.813%, and 1.52%, respectively. Comparisons of accuracy in estimating the mean and confidence level durations are made between the developed neural networks, JPDM, PNET, and JPNET, as the alternatives to the traditional PERT.
The neural network model can reduce the amount of calculation in estimating project duration and obtain a speedier solution than PERT, JPDM, PNET, and JPNET. The results of applying the neural networks developed from this research to estimating the mean duration and the duration for a given confidence level for two project cases show that they can achieve a mean absolute percentage errors within 1.5%.
|
|---|
