LearnLib Tutorial: From Finite Automata to Register Interface Programs

Abstract

In the past decade, active automata learning, an originally merely theoretical enterprise, got attention as a method for dealing with black-box or third party systems. Applications ranged from the support of formal verification, e.g. for assume guarantee reasoning [4], to usage of learned models as the basis for regression testing. In the meantime, a number of approaches exploiting active learning for validation [17,20,6,7,2,1] emerged.Today, active automata learning is on the verge of becoming a valuable asset in bringing formal methods to systems lacking formal descriptions (e.g., the huge class of legacy systems): This edition of ISoLA alone features a track on active learning in formal verification [16], one on model-based testing and model inference [12], this tutorial, and is co-located with the STRESS summer school, 3 where active automata learning is part of the curriculum.In particular when dealing with black-box systems, i.e., systems that can be observed, but for which no or little knowledge about the internal structure or even their intent is available, active automata learning can be considered as a key technology due to its test-based approach to model inference. However, the testbased interaction introduces a number of challenges when using active automata learning to infer models of real word systems, which have been summarized in [21]:A: Interacting with real systemsThe interaction with a realistic target system comes with two problems. The technical problem of establishing an adequate interface that allows one to apply test cases for realizing so-called membership queries, and a conceptual problem of bridging the gap between the abstract learned model and the concrete runtime scenario. B: Membership Queries 3 http://info.santoslab.org/event/stress2012

Extracting automata from neural networks using active learning

“…We have implemented our approach in a prototype in Java, wherein we use the library LearnLib (Howar et al, 2012) to implement the MAT learning framework and the RPNI algorithm. Moreover, to find the true counterexamples faster, we use the Wpmethod test (Fujiwara et al, 1991) 4 in the equivalence query between the hypothesis and the abstract models.…”

“…Many approaches in grammatical inference can be described as passive learning. In the paper, we consider the polynomial-time RPNI algorithms provided in the library LearnLib (Howar et al, 2012). Oncina & Garca (1992) proposed the Regular Positive and Negative Inference (RPNI) algorithm for DFA learning.…”

“…Finally, we have implemented our approach in Java, wherein we use the library LearnLib (Howar et al, 2012) to implement the active learning framework. To evaluate our approach, we conducted a series of experiments on a classifier for the MNIST dataset, a large database of handwritten digits that is commonly used for training various image processing systems.…”

Integration testing of communicating systems with unknown components

“…There are variations of Angluin's L* algorithm such as [16], where essentially a 1-quotient of the SUT is inferred. Many algorithms and heuristic improvements on L* have been developed for software testing, and most of them have been implemented in the LearnLib framework [17], which can be considered currently as the reference for this kind of inference. LearnLib specific algorithms actually won several inference competitions such as Zulu and RERS.…”

Refactoring of Legacy Software Using Model Learning and Equivalence Checking: An Industrial Experience Report

“…Instead we decided to use a different combination of tools, similar to the approach of [13,2]. The model learning tool LearnLib [15] was used to learn Mealy machine models of the legacy and the refactored implementation. These models were then compared to check if the two implementations are equivalent.…”