Metamorphic testing of stochastic optimisation

Abstract

Testing stochastic optimisation algorithms presents an unique challenge because of two reasons. First, these algorithms are non-testable programs, i.e. if the test oracle was known, there wouldn't have been the need for those algorithms in the first place. Second, their performance can vary depending on the problem instances they are used to solve. This paper applies the statistical metamorphic testing approach to stochastic optimisation algorithms and investigates the impact that different problem instances have on testing optimisation algorithms. The paper presents an empirical evaluation of the approach using instances of Next Release Problem (NRP). The effectiveness of the testing method is evaluated using mutation testing. The result shows that, despite the challenges from the stochastic nature of the optimisation algorithm, metamorphic testing can be effective in testing them. © 2010 IEEE.