Learning-based test programming for programmers

Abstract

While a diverse array of approaches to applying machine learning to testing has appeared in recent years, many efforts share three central challenges, two of which are not always obvious. First, learning-based testing relies on adapting the tests generated to the program being tested, based on the results of observed executions. This is the heart of a machine learning approach to test generation. A less obvious challenge in many approaches is that the learning techniques used may have been devised for problems that do not share all the assumptions and goals of software testing. Finally, the usability of approaches by programmers is a challenge that has often been neglected. Programmers may wish to maintain more control of test generation than a "push button" tool generally provides, without becoming experts in software testing theory or machine learning algorithms, and with access to the full power of the language in which the tested system is written. In this paper we consider these issues, in light of our experience with adaptation-based programming as a method for automated test generation. © 2012 Springer-Verlag.