Efficient Automaton Learning Algorithms for Software Engineering Applications

Abstract

Automaton learning is a domain in which the target system aka System Un der Learning (SUL) is inferred by the automaton learning algorithm in the form of an automaton, by synthesizing a finite number of inputs and their corresponding outputs. Automaton learning makes use of a Minimally Ad equate Teacher (MAT). The learner learns the SUL by posing membership queries to the MAT. In this thesis, we try to infer behavioral models of the system to be tested and verified. For this, we propose efficient automaton learning algorithms. The proposed algorithms can be used to infer a behavioral representation of the underlying system automatically in the form of an automaton or a Kripke structure which can be used for software testing, formal verification, or CE GAR (counterexample guided abstraction refinement) which are all applica tions of software engineering and need some sort of model (built manually, semi-automatically or automatically). The proposed algorithms can do this job more efficiently as they have reduced the complexity compared to other algorithms available in the literature. The existing automaton learning algorithms have at least polynomial (cubic i.e. n3) time complexity therefore, for learning and testing complex soft ware systems the existing automaton learning algorithms take a lot of time. Sometimes, even these algorithms can fail to learn large complex software xiii systems. All the existing automaton learning algorithms learn the SUL from the initial state to the final state(s). In this research work, we introduce a novel concept of automaton learning in the form of inverse transitions (δ−1). In this thesis, we inquire about the feasibility of learning an automaton in the inverse direction (from the final state(s) to the initial state) and prove that with the help of Inverse Query (IQ) and δ−1 transitions, we can de sign more efficient automaton learning algorithms than existing algorithms. For this, we design and implement complete learning algorithms (DLIQ and BDLIQ), an incremental learning algorithm (IDLIQ), and a multi-bit incre mental Kripke learning algorithm (BIKL) based on the concepts of inverse transition (δ−1) and Inverse Query (IQ). An inverse query is a question that is posed to the teacher (MAT) about the predecessor state(s) of a state. To answer against the IQ, we enhance the capability of existing MAT. Besides this, for performance verification of our proposed algorithms, we implement an evaluation framework to compare their performance with existing algo rithms. The results depict that our proposed automaton learning algorithms are more efficient than existing algorithms in terms of time complexity. In the early stages, theoretical automaton learning has successful real-world applications. It goes beyond formal verification and allows to infer behav ioral model of black-box systems. The improvements in tool support and raising competition focuses on the practicality of automaton learning and en couraging Learning-based Testing (LBT) techniques, which shows the grow ing interest in the field of research. Therefore, in this thesis, we have se lected LBT for real-life benchmarks. For this, we have selected three differ ent real-world systems as our case studies. Results of all these case studies illustrate that our proposed multi-bit Kripke learning algorithm; the BIKL outperforms existing Kripke structure learning algorithms. As the notion of automaton learning through inverse query improves the complete and incre mental learning process of DFA and Kripke structure therefore, we believe that this concept can be helpful for the software engineering community to innovate new methods for learning and testing domains