Model learning

This research will be part of the TiCToC (Testing In times of Continuous Change) research project. TiCToC integrates model based testing with automata learning. With model based testing we manually create a model of the system. This can be rather difficult, especially for legacy systems where specifications are missing, incomplete or incorrect.

Model learning is an approach where a model is derived from log files or other observations of the system. It will attempt to construct a model that corresponds to the observed behavior. Model learning is actively researched at the universities of Twente and Nijmegen. One of the leading tools is LearnLib.

The LearnLib approach has some serious shortcomings. Therefore we are also interested in other approaches, for example Test based modeling as done by Tim Willemsen. The types of models that are learned in the aforementioned studies are different than the models that Axini uses. At Axini we use Symbolic Transition Systems that support time and data. These models have pleasant compositional properties to form bigger models by combining smaller models. The models that LearnLib supports, for example Mealy machines and Finite Automata, do not share these compositional properties.

Additionally, current techniques have difficulty with systems that exhibit non-deterministic behavior. We think that LLMs (large language models) are an interesting candidate to apply to model learning, given their capabilities in processing natural language and structured data such as log files.