An Empirical investigation into metrics for object-oriented software.

Abstract

Object-Oriented methods have increased in popularity over the last decade, and are now the norm for software development in many application areas. Many claims were made for the superiority of object-oriented methods over more traditional methods, and these claims have largely been accepted, or at least not questioned by the software community. Such was the motivation for this thesis. One way of capturing information about software is the use of software metrics. However, if we are to have faith in the information, we must be satisfied that these metrics do indeed tell us what we need to know. This is not easy when the software characteristics we are interested in are intangible and unable to be precisely defined. This thesis considers the attempts to measure software and to make predictions regarding maintainabilty and effort over the last three decades. It examines traditional software metrics and considers their failings in the light of the calls for better standards of validation in terms of measurement theory and empirical study. From this five lessons were derived. The relatively new area of metrics for object-oriented systems is examined to determine whether suggestions for improvement have been widely heeded. The thesis uses an industrial case study and an experiment to examine one feature of objectorientation, inheritance, and its effect on aspects of maintainability, namely number of defects and time to implement a change. The case study is also used to demonstrate that it is possible to obtain early, simple and useful local prediction systems for important attributes such as system size and defects, using readily available measures rather than attempting predefined and possibly time consuming metrics which may suffer from poor definition, invalidity or inability to predict or capture anything of real use. The thesis concludes that there is empirical evidence to suggest a hypothesis linking inheritance and increased incidence of defects and increased maintenance effort and that more empirical studies are needed in order to test the hypothesis. This suggests that we should treat claims regarding the benefits of object-orientation for maintenance with some caution. This thesis also concludes that with the ability to produce, with little effort, accurate local metrics, we have an acceptable substitute for the large predefined metrics suites with their attendant problems.