Domain-Specific Optimization in Automata Learning Hardi Hungar, Oliver Niese, and Bernhard Steffen University of Dortmund {Hardi.Hungar, Oliver.Niese, Bernhard.Steffen}@udo.edu Abstract. Automatically generated models may provide the key to- wards controlling the evolution of complex systems, form the basis for test generation and may be applied as monitors for running applica- tions. However, the practicality of automata learning is currently largely preempted by its extremely high complexity and unrealistic frame con- ditions. By optimizing a standard learning method according to domain- specific structural properties, we are able to generate abstract models for complex reactive systems. The experiments conducted using an industry- level test environment on a recent version of a telephone switch illustrate the drastic effect of our optimizations on the learning efficiency. From a conceptual point of view, the developments can be seen as an instance of optimizing general learning procedures by capitalizing on specific ap- plication profiles. 1 Introduction 1.1 Motivation The aim of our work is improving quality control for reactive systems as can be found e.g. in complex telecommunication solutions. A key factor for effective quality control is the availability of a specification of the intended behavior of a system or system component. In current practice, however, only rarely pre- cise and reliable documentation of a system’s behavior is produced during its development. Revisions and last minute changes invalidate design sketches, and while systems are updated in the maintenance cycle, often their implementa- tion documentation is not. It is our experience that in the telecommunication area, revision cycle times are extremely short, making the maintenance of spec- ifications unrealistic, and at the same time the short revision cycles necessitate extensive testing effort. All this could be dramatically improved if it were pos- sible to generate and then maintain appropriate reference models steering the testing effort and helping to evaluate the test results. In [6] it has been pro- posed to generate the models from previous system versions, by using learning techniques, and incorporating further knowledge in various ways. We call this general approach moderated regular extrapolation, which is tailored for a poste- riori model construction and model updating during the system’s life cycle. The general method includes many different theories and techniques [14]. W.A. Hunt, Jr. and F. Somenzi (Eds.): CAV 2003, LNCS 2725, pp. 315–327, 2003. c  Springer-Verlag Berlin Heidelberg 2003