Active Automata Learning as Black-Box Search and Lazy Partition Refinement

• Published in: A Journey from Process Algebra Via Timed Automata to Model Learning Vol. 13560; pp. 321 - 338
• Main Authors: Howar, Falk, Steffen, Bernhard
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer 2022; Springer Nature Switzerland
• Series: Lecture Notes in Computer Science
• Subjects: Active Automata Learning; Minimally Adequate Teacher; Model Learning
• ISBN: 9783031156281; 3031156285
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-031-15629-8_17

We present a unifying formalization of active automata learning algorithms in the MAT model, including a new, efficient, and simple technique for the analysis of counterexamples during learning: Lλ\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L^{\!\lambda }$$\end{document} is the first active automata learning algorithm that does not add sub-strings of counterexamples to the underlying data structure for observations but instead performs black-box search and partition refinement. We analyze the worst case complexity in terms of membership queries and equivalence queries and evaluate the presented learning algorithm on benchmark instances from the Automata Wiki, comparing its performance against efficient implementations of some learning algorithms from LearnLib.