Machine Learning for Metamorphic Testing

Abstract: Test oracle is a mechanism used to validate all the functionalities ofsoftware under test. However, the lack of test oracle makes the processof software testing difficult. Metamorphic testing is a state of artapproach for automated software testing without test oracles basedon metamorphic relations. Metamorphic relations are a set of propertiesbetween inputs and outputs that a software could have. However,it is usually difficult to identify metamorphic relations for unknownprograms. This thesis aims for automatic generation of metamorphicrelations by utilizing machine learning algorithms with the methodrandom walk kernel using input from control flow graphs. By applyingKanawala et al. [1] previous work in our targeted system environment,we encountered a series of difficulties, which we also describein this thesis. It is important to introduce an alternative solution forour test suite that is working. The performance of our model is evaluatedby different measures including area under the receiver operatingcharacteristic curve and mean squared error. The results show promisingapplications of automatically predicting metamorphic relations forunknown programs. The study was conducted on software system inthe telecommunicaiton domain.