Applicability of Neural Network Based Models for Software Effort Estimation

Effort Estimation is a very challenging task in the software development life cycle. Inaccurate estimations may cause the client dissatisfaction and thereby, decrease the quality of the product. Considering the problem of software cost and effort prediction, it is conceivable to call attention to that the estimation procedure considers the qualities present in the data set, as well as the aspects of the environment in which the model is embedded. Existing literatures have the instances where machine learning techniques such as Linear Regression (LR), Support Vector Machine (SVM), K-Nearest Neighbor (KNN) have been used to estimate the effort required to develop any software. Yet it is quite uncertain for any particular model to perform well with all the data sets. Most of the research is based on the dataset of any single organization. Consequently, the results obtained through these models cannot be generalized. So, the main objectives of this research are: i) to use different data preparation techniques such as selection, cleaning, and transformation to improve the quality of data set given to the model ii) to use other machine learning models such as Multi-Layer Perceptron Neural Network (MLPNN), Probabilistic Neural Network (PNN), and Recurrent Neural Network (RNN) to increase the performance of software effort estimation process iii) to use different optimization techniques to tune the parameters of machine learning models iv) to use ensemble methods to improve the accuracy of software effort estimation process. In this study, first, we found out the most influential attributes in the Desharnais data set, then, MLPNN has been applied on reduced data set with to improve the accuracy of software effort estimation. Then, the performance of the MLPNN model is compared with LR, SVM and KNN models in the literature to find the best model fitting this dataset. Results obtained from the study demonstrate that some of the variables are more important in comparison to others for effort estimation. Also among the various models used in this study, the best-obtained R2 value is 79 % for the MLPNN model.