The Teachers' Crowd: The Impact of Distributed Oracles on Active Automata Learning

In this paper we address the major bottleneck of active automata learning, the typically huge number of required tests, by investigating the impact of using a distributed testing environment (a crowd of teachers) to execute test cases (membership queries) in parallel. This kind of parallelization of automata learning has the best potential when the time for test case execution is dominant, an assumption valid for most practical applications. Our investigation explicitly focuses on the impact of the structure of the system under learning (number of states, size of alphabet) and the degree of supported parallelism. It comprises three variants of active learning algorithms with different test case generation profiles. These differences can be observed directly at the level of the run-times, which all show a linear speedup for moderate degrees of parallelization, but with different saturation points beyond which further parallelization does not pay off.