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.
