Abstract

Abstraction is the key when learning behavioral models of realistic systems, but also the cause of a major problem: the introduction of non-determinism. In this paper, we introduce a method for refining a given abstraction to automatically regain a deterministic behavior on-the-fly during the learning process. Thus the control over abstraction becomes part of the learning process, with the effect that detected non-determinism does not lead to failure, but to a dynamic alphabet abstraction refinement. Like automata learning itself, this method in general is neither sound nor complete, but it also enjoys similar convergence properties even for infinite systems as long as the concrete system itself behaves deterministically, as illustrated along a concrete example.