The Effectiveness of t-Way Test Data Generation

Ellims, Michael (2009). The Effectiveness of t-Way Test Data Generation. PhD thesis The Open University.



Modern society is increasingly dependent on the correct functioning of software and increasingly so in areas that are considered safety related or safety critical. Therefore, there is an increasing need to be able to verify and validate that the software is in fact correct and will perform its intended function. Many approaches to this problem have been proposed; however, none seems likely to supplant the role of testing in the near future.

If we accept that there is, and will be, a continuing need to be able to test software then the question becomes one of how can this be done effectively, both in terms of ability to detect errors and in terms of cost. One avenue of research that offers prospects of improving both of these aspects is the automatic generation of test data.

There has recently been a large amount of work conducted in this area. One particularly promising direction has been the application of ideas from the field of experimental design and in particular, the field of t-way adequate factorial designs.

The area however, is not without issues; there is evidence that the technique is capable of detecting errors but that evidence is not unequivocal. Moreover, as with almost all work in the area of automatic test generation, there has been very little comparative work comparing the technique with other test data generation techniques. Worse, there has been effectively no work done that compares any automatic test data generation technique with the effectiveness of tests generated by humans. Another major issue with the technique is the number of tests that applying the technique can result in. This implies that there is a need for an automated oracle if the technique is to be successfully applied. The flaw with this is of course that in most situations the oracle is the human that is conducting the tests, a point often ignored in testing research.

The work presented here addresses both of these points. To do this I have used a code base taken from an industrial engine control system that has an existing set of high quality unit tests developed by hand. To complement this, several other techniques for automatically generating test data have been applied, namely random testing, random experimental designs and a technique for generating single factor experiments. To address the issue of being able to compare the error detection ability of all of the sets of test vectors, rather than the usual effectiveness surrogates of code coverage I have used mutation analysis on the code base to directly measure the ability of each set of test vectors to discover common coding errors. The results presented here show that test data generation techniques based on t-way factorial designs are at least as effective as handgenerated tests and superior to random testing and the factor experimental technique.

The oracle problem associated with the factorial design techniques was addressed using a test set minimisation approach. The mutation tool monitored which vectors could “kill” which code mutants. After a subset of the test vectors had been run, the most effective vectors were retained and the rest discarded. Likewise, mutants that were killed were removed from further consideration and the process repeated. Experimental results show that this minimisation procedure is effective at reducing computational overhead and is capable of producing final sets of test vectors that are comparable in size with the sets of hand-generated tests and so amenable to final hand checking.

Viewing alternatives

Download history


Public Attention

Altmetrics from Altmetric

Number of Citations

Citations from Dimensions

Item Actions