Learning Finite State Machine Specifications from Test Cases

Abstract

Up-to-date software systems are often modular and need to be changeable. While modularity is believed to reduce software production costs, it also leads to increased di culty in testing, as the future uses of a module are getting more varied and therefore harder to guess. Also, reused modules are often represented as black boxes, whose interior structure is hidden from the system. In consequence, on reusing a module, the testing focuses on integration testing and monitoring the interfaces of the module.In order to monitor a system, a model of the system is needed to compare the observed traces to. However, with the advent of agile development methods and decreasing time to market, formal models of a system are seldom available. While formal modeling has not been widely adopted in industry, the importance of testing has increased, in part thanks to the same agile development methods that obsolete explicit modeling. An example is the test rst paradigm of eXtreme Programming, which requires that the tests for any software have to be written before the software itself. Therefore, test cases are available even for systems without a formal model.In consequence, we propose to generate a system model suitable for monitoring from a test suite for the system. The approach is based on automata learning. Angluin's learning algorithm is used to generate an appropriate model, while state-merging methods are applied to represent the test cases in a format that can be processed by the learning algorithm.Both Angluin's algorithm and state-merging are tailored to the characteristics of testing. For Angluin's algorithm, this comprises a mapping of the query mechanisms onto a test suite. The state-merging is used to construct a generic representation of arbitrary test suites by exploiting the properties of a given test speci cation language for a better coverage. The approach is implemented in a prototypical tool and validated by a case study.Like Alice in Wonderland, the research for this thesis has taken me to various unknown shores and topics. Many friends and colleagues have accompanied me along the path of my scienti c journey, and to all of them I o er my thanks for their company.First and foremost, I want to give my thanks to my supervisor, Prof. Dr. Jens Grabowski. Without his patience and persistence, this thesis would never have been completed.I am also very grateful to all colleagues who o ered their advice on this thesis. Franz Schenk was the rst one to read my initial e orts, and his comments on structure have greatly eased my task. Ste en Herbold proved to be the person with the most insight on the mathematical foundations of my research topic; therefore he was the one to spot any theoretical inconsistencies. Philip Makedonski has helped to polish my English phrasing. Special thanks are due to Prof. Dr. Wolfgang May for his valuable comments on correctly formulating mathematical de nitions.Many grateful thoughts also go to all my friends and family, whose constant support and never ending inquiries as to the progress o...

Application of machine learning methods in the classification of satellite images

“…To znači da model treba da pokazuje dobre performanse i u slučajevima ili situacijama koje nisu direktno obuhvaćene podacima korišćenim za njegovo treniranje. Suštinski, cilj je razviti model koji će efikasno donositi pouzdane zaključke i pružiti rešenja za buduće instance problema, time demonstrirajući pravu vrednost i primenljivost u praksi [8][9][10]. Greške u modelovanju podataka su sastavni deo MU, bilo tokom treniranja ili kasnije primene modela.…”