Modeling Dredging Project Cost Variations

Abstract

The U.S. Army Corps of Engineers maintains a database of cost information for competitively bid dredging projects. These data were used to construct linear regression models and radial-basis-function neural networks to predict the completed cost of the dredging projects. The stepwise linear regression procedure was used to construct equations to predict the completed cost based on input of the low bid, government cost estimates, and estimated project-dredging quantities. A data transformation using the natural logarithm enhanced the linear relationships between the variables. An exponential relationship between the low bid and completed cost indicated that large dredging projects could be completed for less than the bid amount. The variables used as inputs to the neural networks were the low bid, the government estimate, estimated quantity, the type of dredge, the method of dredged material disposal, the number of bidders, and the class of work. The addition of categorical variables like the type of dredging and disposal method did not improve the predictive performance of the neural network. The best neural network model was able to predict 40.4% of the test set projects within 10% of the actual cost. The best regression model predicted 51.4% of the projects within 10% of the actual cost.