The global IT industry has now matured. As more and more systems grow old and enter into the maintenance stage, software maintenance (SM) is becoming one of the most carried out and challenging tasks. Besides, the industry is also facing a shift in traditional technical environment by way of use of newer tools and approaches of software development, migration from legacy software to current software and dynamic changes in the SM environment. The challenge then lies in accurately modeling and predicting the SM effort, schedule and risk involved, under the above circumstances. This work employs a neural network (NN) approach to model and predict the software maintenance effort based on an available real life dataset of outsourced maintenance projects (Rao and Sarda, 36 projects of 14 drivers). A comparison between results obtained by NN and regression modeling is also presented. It is concluded that NN is able to successfully model the complex, non-linear relationship between a large number of effort drivers and the software maintenance effort, with results closely matching the effort estimated by experts.

Abstract — Software development effort prediction is one of the most key activities in software industry. Many models have been proposed to build a relationship between software size and effort; however we still have problems for effort prediction. This is because project data, available in the primary stages of project is often inadequate, unpredictable, uncertain and unclear. The need for accurate effort estimation in software industry is an ongoing challenge. Artificial Neural Network models are more apt in such situations. The present paper is concerned with developing software effort prediction models based on artificial neural networks. The models are designed to improve the performance of the network that suits to the COCOMO Model. Artificial Neural Network models are created using Radial Basis and Generalized Regression. A case study based on the NASA 93 database compares the proposed neural network models with the Intermediate COCOMO. The results were analyzed using