Allen, Edward B.

Development of reliable, high quality, software requires study and understanding at each step of the development process. A basic assumption in the field of software measurement is that metrics of internal software attributes somehow relate to the intrinsic difficulty in understanding a program. Measuring the information content of a program attempts to indirectly quantify the comprehension task. Information theory based software metrics are attractive because they quantify the amount of information in a well defined framework. However, most information theory based metrics have been proposed with little reference to measurement theory fundamentals, and empirical validation of predictive quality models has been lacking. This dissertation proves that representative information theory based software metrics can be "meaningful" components of software quality models in the context of measurement theory. To this end, members of a major class of metrics are shown to be regular representations of Minimum Description Length or Variety of software attributes, and are interval scale. An empirical validation case study is presented that predicted faults in modules based on Operator Information. This metric is closely related to Harrison's Average Information Content Classification, which is the entropy of the operators. New general methods for calculating synthetic complexity at the system level and module level are presented, quantifying the joint information of an arbitrary set of primitive software measures. Since all kinds of information are not equally relevant to software quality factors, components of synthetic module complexity are also defined. Empirical case studies illustrate the potential usefulness of the proposed synthetic metrics. A metrics data base is often the key to a successful ongoing software metrics program. The contribution of any proposed metric is defined in terms of measured variation using information theory, irrespective of the metric's usefulness in quality models. This is of interest when full validation is not practical. Case studies illustrate the method.