Applying Automata Learning to Complex Industrial Software

Learning models of machines comes very naturally to humans. They construct a mental model of the behaviour of machines while trying out different options. For machines this is much harder. Algorithms have been developed to learn the behaviour of systems and describe them in state machine models. Once such a model has been learned it can be used by software engineers to improve their software. They can simulate and analyse the behaviour of the system, test newer versions of the system and get insight in legacy systems of which no documentation exists. This master thesis describes a case study done at Océ for the ITALIA project at the Radboud University Nijmegen. The goal of this research was to find out whether state-of-the-art techniques and tools for automata learning are powerful enough to learn models of industrial control software. Specifically whether the LearnLib tool developed at the University of Dortmund was able to learn a model of the Engine Status Manager (ESM) developed at Océ. The ESM controls the transition from one status to another in a printer. It delegates status requests to the connected hardware components and coordinates their responses. Software like the ESM can be found in many embedded systems in one form or another. Although such the ESM may seem simple, the many details and exceptions involved make it hard to learn. A system is learned in two alternating phases: A learning phase in which a hypothesis model is created, and a testing phase in which the algorithm searches for a counterexample. This is implemented in the LearnLib. In order to find counterexamples during the test phase using less queries novel techniques had to be developed. Although a correct model of the ESM could not be learnt during this research novel techniques were developed that greatly reduce the number of test queries needed.