Intelligently Predict Project Effort by Reduced Models Based on Multiple Regressions and Genetic Algorithms with Neural Networks

Estimating the amount of effort required for developing a software system is one of the most important project management concerns. This study successfully produces an optimal reduced linear model for software cost estimation by employing a series of methods of multiple regressions to identify the most significant explanatory variables of the fifteen COCOMO cost drivers. The results yielded by the linear models are then compared with their counterparts obtained from the simulation using genetic algorithms with feed-forward neural networks (NN) with back-propagation learning algorithms. The performance of the resulted optimal reduced linear model is very close to that of the full regression and neural network models, and is also comparable to that of the COCOMO'81 intermediate in terms of MMRE and Pred (25). As both linear and nonlinear reduction methods described in this paper are applied and the most significant nine explanatory variables selected among the fifteen COCOMO cost drivers in these reduced models are the identical and their effort estimation accuracy is highly acceptable, the reduced models can be concluded as accurate and robust.