Section: Research Program

Transversal concern: missing models

We are concerned with one important lesson derived from our involvement in several application domains. Most of our background gets in force as soon as a perfect model of the system under study is available. Then verification, control, diagnosis, test, etc. can mobilize a solid background, or suggest new algorithmic problems to address. In numerous situations, however, assuming that a model is available is simply unrealistic. This is a major bottleneck for the impact of our research. We therefore intend to address this difficulty, in particular for the following domains.

• Model building for diagnosis. As a matter of fact, diagnosis theory hardly touches the ground to the extent that complete models of normal behavior are rarely available, and the identification of the appropriate abstraction level is unclear. Knowledge of faults and their effects is even less accessible. Also, the actual implemented systems may differ significantly from behaviors described in the norms. One therefore needs a theory for incomplete and erroneous models. Besides, one is often less bothered by partial observations than drowned by avalanches of alerts when malfunctions occur. Learning may come to the rescue, all the more that software systems may be deployed in sandpits and damaged for experimentation, thus allowing the collection of masses of labeled data. Competition on that theme clearly comes from Machine Learning techniques.

• Verification of large scale software. For some verification problems like the one we address in the IPL HAC-Specis, one does not have access to a formal model of the distributed program under study, but only to executions in a simulator. Formal verification poses new problems due to the difficulties to capture global states, to master state space explosion by gathering and exploiting concurrency information.

• Learning of stochastic models. Applications in bioinformatics often lead to large scale models, involving numerous chains of interactions between chemical species and/or cells. Fine grain models can be very precise, but very inefficient for inference or verification. Defining the appropriate levels of description/abstraction, given the available data and the verification goals, remains an open problem. This cannot be considered as a simple data fitting problem, as elements of biological knowledge must be combined with the data in order to preserve explainability of the phenomena.

• Testing and learning timed models: during conformance testing of a black-box implementation against its formal specification, one wants to detect non-conformances but may also want to learn the implementation model. Even though mixing testing and learning is not new, this is more recent and challenging for continuous-time models.

• Process mining. We intend to extend our work on process discovery using Petri net synthesis [35] by using negative information (e.g. execution traces identified as outliers) and quantitative information (probabilistic or fuzzy sets of execution traces) in order to infer more robust and precise models.