SOFTWARE ENGINEERING PROJECT EFFORT ESTIMATION USING FUZZY NEURAL NETWORK

Abstract

“Estimation” is an important task in software engineering. Estimation has application in many areas e.g. cost estimation, effort estimation. The most crucial thing is the “effort estimation of the software project”. Most of the software projects fail due to improper effort estimation. Effort estimation directly affects the budget of the software project. Software Project Mangers consider effort estimation of software engineering project a very difficult and challenging task because of its inherent imprecision. Therefore it is important and most crucial to have right effort estimate at the right time. Accurate effort estimation in software engineering projects is a challenging task, and it is one of the most crucial project management activities. In this dissertation, we propose a Fuzzy Neural Network (FNN) model for effort estimation of software engineering project. This approach provides dual benefits of incorporating qualitative knowledge of experts and learning from historical data obtained from previous projects. The main focus of this research is to minimize the error by combining neural network and fuzzy logic and train the FNN with evolutionary algorithm. Three different datasets are used in this research work for training and testing purposes; each dataset is divided into three equal parts and each part is in turn used for testing purpose. For performance measurement, a total of 216 experiments were performed for the three datasets including the three combinations of each dataset, out of which 108 are with crossover operation and 108 are without crossover operations. The results show that the FNN is well trained by giving small values of Root Mean Square Error (RMSE) Moreover, the results show that greater the population size, lower the RMSE. The accuracy of the FNN is also based on the number of samples which are provided for training; the more the number of samples, the more the accuracy. Once trained, the FNN is supposed to predict the effort (in person months) of a software engineering project. The results show that network is trained at an acceptable level of accuracy. Some of the examples are compared with the COCOMO model and the result shows that the proposed model behaves comparably well in these examples.