LearnLib: a framework for extrapolating behavioral models

Abstract

The original publication is available at www.springerlink.comInternational audienceIn this paper, we present the LearnLib, a library of tools for automata learning, which is explicitly designed for the systematic experimental analysis of the profile of available learning algorithms and corresponding optimizations. Its modular structure allows users to configure their own tailored learning scenarios, which exploit specific properties of their envisioned applications. As has been shown earlier, exploiting application-specific structural features enables optimizations that may lead to performance gains of several orders of magnitude, a necessary precondition to make automata learning applicable to realistic scenarios

Formal Models of Bank Cards for Free

“…Access to the smartcards was realised via a standard smartcard reader and a testing harness discussed in Section IV-A. We connected the SUT to the LearnLib library [4], which served as Learner, see Fig. 2.…”

“…The regular inference algorithm [3], [4] provides sequences of inputs to a System Under Test (SUT) and observes the responses to infer a Mealy machine, a special form of finite state machine, as explained in Section III. In addition to standard learning methods, abstraction techniques for data parameters [5], [6] can be used to learn a more detailed model of the system.…”

Improving active Mealy machine learning for protocol conformance testing

“…In query learning, access to an oracle is needed that can answer specific types of questions such as: whether a specific string is part of the language (membership queries), and whether a given model is a model for the language (equivalence queries). In software model synthesis, the actual software system can be used for this purpose, see, e.g., [49]. When such an oracle is available that can be queried often and quickly, it is possible to identify very large realistic models.…”

“…LearnLib LearnLib [49] is a tool that supports active learning of Mealy machines or DFAs based on the L * algorithm. It contains many optimizations that reduce the number of queries asked by L * , and a means of selecting them in a graphical user interface.…”

“…This is (usually) cheap in the case of software systems and our methods can consequently ask millions of queries within minutes. The Learnlib toolkit [49], winner of the 2010 Zulu active state machine learning competition [21], can therefore learn models consisting of up to tens of thousands of states. Active state machine learning tools have been used successfully for many different applications in software engineering such as regression testing of software components [40], fuzz testing of protocol implementations [22], and inference of botnet protocols [18].…”