From Passive to Active: Learning Timed Automata Efficiently

Abstract

Model-based testing is a promising technique for quality assurance. In practice, however, a model is not always present. Hence, model learning techniques attain increasing interest. Still, many learning approaches can only learn relatively simple types of models and advanced properties like time are ignored in many cases. In this paper we present an active model learning technique for timed automata. For this, we build upon an existing passive learning technique for real-timed systems. Our goal is to efficiently learn a timed system while simultaneously minimizing the set of training data. For evaluation we compared our active to the passive learning technique based on 43 timed systems with up to 20 locations and multiple clock variables. The results of 18060 experiments show that we require only 100 timed traces to adequately learn a timed system. The new approach is up to 755 times faster.