Prueba de mutación evolutiva en entornos orientados a objetos
- Autores: Pedro Delgado Pérez
- Directores de la Tesis: Inmaculada Medina Bulo (dir. tes.) , Juan José Domínguez Jiménez (codir. tes.)
- Lectura: En la Universidad de Cádiz ( España ) en 2017
- Idioma: español
- Tribunal Calificador de la Tesis: Sergio Segura Rueda (presid.) , Juan Boubeta Puig (secret.) , Antonio García Domínguez (voc.)
- Enlaces
- Resumen
Mutation testing is acknowledged as a powerful method to evaluate the strength of test suites in detecting possible faults in the code. However, its application is expensive, which has traditionally been an obstacle for a broader use in the industry. While it is true that several techniques have shown to greatly reduce the cost without losing much effectiveness, it is also true that those techniques have been evaluated in limited contexts, especially in the scope of traditional operators for procedural programs. To illustrate this fact, Evolutionary Mutation Testing has only been applied to WS-BPEL compositions, despite the positive outcome when selecting a subset of mutants through an evolutionary algorithm with the aim of improving a test suite. As a result, it is unknown whether the same benefits can be extrapolated to other levels and domains.
In particular, we wonder in this thesis to what extent Evolutionary Mutation Testing is also useful to reduce the number of mutants generated by class mutation operators in object-oriented systems. More specifically, we focus on the C++ programming language, since the development of mutation testing with regard to this widely-used language is clearly immature judging from the lack of papers in the literature tackling this language. Given that C++ has been hardly addressed in research and practice, we deal with all the phases of mutation testing: from the definition and implementation of mutation operators in a mutation system to the evaluation of those operators and the application of Evolutionary Mutation Testing among other cost reduction techniques.
We define a set of class mutation operators for C++ and implement them in MuCPP, which allows us to perform experiments with real programs thanks to the facilities incorporated into this mutation tool. These mutation operators are automated following a set of guidelines so that they produce the expected mutations. In general, class-level operators generate far fewer mutants than traditional operators, a higher equivalence percentage and they are applied with varying frequency depending on the features of the tested program. Developing improvement rules in the implementation of several mutation operators help further reduce the number of mutants, avoiding the creation of uninteresting mutants. Another interesting finding is that the set of class mutants and the set of traditional mutants complement each other to help the tester design more effective test suites.
We also develop GiGAn, a new system to connect the mutation tool MuCPP and a genetic algorithm to apply Evolutionary Mutation Testing to C++ object-oriented systems. The genetic algorithm allows reducing the number of mutants that would be generated by MuCPP as it guides to the selection of those mutants that can induce the generation of new test cases (strong mutants). The performance of this technique shows to be better than the application of a random algorithm, both when trying to find different percentages of strong mutants and also when simulating the refinement of the test suite through the mutants selected by each of these techniques. The stability of EMT among different case studies and the good results of the simulation in the programs that lead to the largest set of mutants are additional observations.
Finally, we conduct an experiment to assess individually these mutation operators from a double perspective: how useful they are for the evaluation of the test suite (TSE) and its refinement (TSR). To that end, we rank the operators using two different metrics: degree of redundancy (TSE) and quality to guide on the generation of high-quality test cases (TSR). Based on these rankings, we perform a selective study taking into account that the less valuable operators are at the bottom of the classification. This selective approach reveals that an operator is not necessarily as useful for TSE as for TSR, and that these rankings are appropriate for a selective strategy when compared to other rankings or the selection of mutants randomly. However, favouring the generation of individual mutants from the best-valued operators is much better than discarding operators completely because each of the operators targets a particular object-oriented feature. Altogether, these evaluations about class operators suggest that their nature can limit the benefits of any cost reduction technique.
