Test-Suite Reduction Does Not Necessarily Require Executing the Program under Test

Removing redundancies from test-suites is an important task of software testing in order to keep test-suites as small as possible, but not to harm the test-suite's fault detection capabilities. A straightforward algorithm for test-suite reduction would select elements of the test-suite randomly and remove them if and only if the reduced test-suite fulfills the same or similar coverage or mutation score. Such algorithms rely on the execution of the program and the repeated computation of coverage or mutation score. In this paper, we present an alternative approach that purely relies on a model learned from the original test-suite without requiring the execution of the program under test. The idea is to remove those tests that do not change the learned model. In order to evaluate the approach we carried out an experimental study showing that reductions of 60-99% are possible while still keeping coverage and mutation score almost the same.