Testing scientific software: techniques for automatic detection of metamorphic relations

Abstract

Scientific software plays an important role in critical decision making in fields such as the nuclear industry, medicine, and the military. Systematic testing of such software can help to ensure that it works as expected. Comprehensive, automated software testing requires an oracle to check whether the output produced by a test case matches the expected behavior of the program. But the challenges in creating suitable oracles limit the ability to perform automated testing of scientific software. For some programs, creating an oracle may be not possible since the correct output is not known a priori. Further, it may be impractical to implement an oracle for an arbitrary input due to the complexity of a program. The software testing community refers to such programs as non-testable. Many scientific programs fall into this category of non-testable programs, since they are either written to find answers that are previously unknown or they perform complex calculations. In this work, we developed techniques to automatically predict metamorphic relations by analyzing the program structure. These metamorphic relations can serve as automated partial test oracles in scientific software. Metamorphic testing is a method for automating the testing process for programs without test oracles. This technique operates by checking whether a program behaves according to a certain set of properties called metamorphic relations. A metamorphic relation is a relationship between multiple input and output pairs of the program. It specifies how the output should change following a specific change made to the input. A change in the output that differs from what is specified by the metamorphic relation indicates a fault in the program. Metamorphic testing can be effective in testing machine learning applications, bioinformatics programs, health-care simulations, partial differential equations and other programs. Unfortunately, finding appropriate metamorphic relations for use in metamorphic testing remains a labor intensive task that is generally performed by a domain expert or a programmer. In this work we applied novel machine learning based approaches to automatically derive metamorphic relations. We first evaluated the effectiveness of modeling the metamorphic relation prediction problem as a binary classification problem. We found that support vector machines are the most effective binary classifiers for predicting metamorphic relations. We also found that using walk-based graph kernels for feature extraction from graph-based program representations further improves the prediction accuracy. In addition, incorporating mathematical properties of operations in the graph kernel computation improves the prediction accuracy. Further, we found that control flow information of a function are more effective than data dependency information for predicting metamorphic relations. Finally we investigated the possibility of creating multi-label classifiers that can predict multiple metamorphic relations using a single classifier. Our empirical studies show that multi-label classifiers are not effective as binary classifiers for predicting metamorphic relations. Automated testing will make the testing process faster, reduce the testing cost and make it more reliable. Automated testing requires automated test oracles. Automatically discovering metamorphic relations is an important step towards automating oracle creation. Work presented here is the first attempt towards developing automated techniques for deriving metamorphic relations. Our work contributes toward automating the testing process of programs that face oracle problems.