A major hurdle for the application of automata learning to realistic systems is the identification of an adequate alphabet: it must be small enough, in particular finite, for the learning procedure to converge in reasonable time, and it must be expressive enough to describe the system at a level where its behavior is deterministic. In this paper, we combine our automated alphabet abstraction approach, which refines the global alphabet of the system to be learned on the fly during the learning process, with the principle of state-local alphabets: rather than determining a single global alphabet, we infer the optimal alphabet abstraction individually for each state. Our experimental results show that this does not only lead to an increased comprehensibility of the learned models, but also to a better performance of the learning process: indeed, besides the drastic - yet foreseeable - reduction in terms of membership queries, we also observed interesting cases where the number of equivalence queries was reduced.
展开▼
