Local model learning for asynchronous services

Software services are often composed into more complex systems. Existing methods ensure the correctness of service compositions by automatically generating a mediator/adaptor service: a service in the middle to properly coordinate the interactions in the system towards satisfying a desired temporal property. This is accomplished using formal behavioural models for the participating services. However, such models are not always provided, which makes it difficult to compose systems containing incompletely specified services. We developed a black-box model learning method specifically adapted for stateful asynchronous services. Often, such services exhibit uncontrollable behaviour, which is not addressed by current learning techniques. Our technique interleaves runtime exploration with model refinement in order to learn an approximation of the real behaviour that allows for a safe system composition. Furthermore, the service model is learned locally, thus allowing parallelism in the inference process when more than one black-box service model has to be learned. Experiments performed show that obtained models are precise enough to be used for adaptor synthesis.