SOFTWARE EFFORT ESTIMATION USING HYBRIDIZED SEARCH BASED TECHNIQUES

Abstract

Prediction of resource requirements of a software project is crucial for the timely delivery of quality-assured software within a reasonable timeframe. Software effort estimation is the process of prognosticating the amount of effort required to build a software project. Most cost estimation models attempts to generate an effort estimation, which can then be mapped into project duration and cost. Many conventional (model-based) and Artificial Intelligence (AI) oriented (model-free) resource estimators have been proposed in the recent past. In this thesis two search based Effort Estimation techniques are discussed .Firstly we evaluates a genetically trained neural network (NN) predictor trained on historical data. Secondly, Particle Swarm Optimization (PSO) technique which operates on data sets clustered using the K-means clustering algorithm. Hence PSO and Genetic Algorithm (GA) based search techniques are employed to perform optimized search in solution space. The comparison of this new predictor is established using n-fold cross validation and Student’s t-test. The data is obtained on various partitions of merged COCOMO data set and Kemerer data sets incorporating data from 78 real-life software projects. PSO is employed to generate parameters of the COCOMO (Constructive Cost Model) model for each cluster of data values. The clusters and effort parameters are then trained to a Neural Network by using Back propagation technique, for classification of data. Here we have tested the model on the COCOMO dataset and also compared the obtained values with standard COCOMO model. By making use of the experience from Neural Networks and the efficient tuning of parameters by PSO operating on clusters, the proposed model is able to generate comparable results and it can be applied efficiently to larger data sets. It goes without saying that a predictor trained on historical data can only be as accurate as the data set itself. Hence, there is a need to continue collection of data on diverse projects with wide range of attributes to construct a sizable historical database for training neural predictors. Using search based techniques to train NN; we are looking to overcome this limitation to possible extent.