Software Fault Prediction and Test Data Generation Using Articial Intelligent Techniques

Abstract

The complexity in requirements of the present-day software, which are often very large in nature has lead to increase in more number of lines of code, resulting in more number of modules. There is every possibility that some of the modules may give rise to varieties of defects, if testing is not done meticulously. In practice, it is not possible to carry out white box testing of every module of any software. Thus, software testing needs to be done selectively for the modules, which are prone to faults. Identifying the probable fault-prone modules is a critical task, carried out for any software. This dissertation, emphasizes on design of prediction and classication models to detect fault
prone classes for object-oriented programs. Then, test data are generated for a particular task to check the functionality of the software product. In the eld of object-oriented software engineering, it is observed that Chidamber and Kemerer (CK) software metrics suite is more frequently used for fault prediction analysis, as it covers the unique aspects of object - oriented programming such as the complexity, data abstraction, and inheritance. It is observed that one of the most important goals of fault prediction is to detect
fault prone modules as early as possible in the software development life cycle (SDLC). Numerous authors have used design and code metrics for predicting fault-prone modules. In this work, design metrics are used for fault prediction. In order to carry out fault prediction analysis, prediction models are designed using machine learning methods. Machine learning methods such as Statistical methods, Articial neural network, Radial basis function network, Functional link articial neural network, and Probabilistic neural network are deployed
for fault prediction analysis. In the rst phase, fault prediction is performed using the CK metrics suite. In the next phase, the reduced feature sets of CK
metrics suite obtained by applying principal component analysis and rough set theory are used to perform fault prediction. A comparative approach is drawn to nd a suitable prediction model among the set of designed models for fault prediction. Prediction models designed for fault proneness, need to be validated for their eciency. To achieve this, a cost-based evaluation framework is designed to evaluate the eectiveness of the designed fault prediction models. This framework, is based on the classication of classes as faulty or not-faulty. In this cost-based analysis, it is observed that fault prediction is found to be suitable where normalized estimated fault removal cost (NEcost) is less than
certain threshold value. Also this indicated that any prediction model having NEcost greater than the threshold value are not suitable for fault prediction, and then further these classes are unit tested. All the prediction and classier models used in the fault prediction analysis are applied on a case study viz., Apache Integration Framework (AIF). The metric data values are obtained
from PROMISE repository and are mined using Chidamber and Kemerer Java Metrics (CKJM) tool. Test data are generated for object-oriented program for withdrawal task in Bank ATM using three meta-heuristic search algorithms such as Clonal selection algorithm, Binary particle swarm optimization, and Articial bee colony algorithm. It is observed that Articial bee colony algorithm is able
to obtain near optimal test data when compared to the other two algorithms. The test data are generated for withdrawal task based on the tness function derived by using the branch distance proposed by Bogdan Korel. The generated test data ensure the proper functionality or the correctness of the programmed module in a software.