Combining Machine Learning and Formal Methods for Complex Systems Design

During the last 20 years, model-based design has become a standard practice in many fields such as automotive, aerospace engineering, systems and synthetic biology. This approach allows a considerable improvement of the final product quality and reduces the overall prototyping costs. In these contexts, formal methods, such as temporal logics, and model checking approaches have been successfully applied. They allow a precise description and automatic verification of the prototype's requirements. In the recent past, the increasing market requests for performing and safer devices shows an unstoppable growth which inevitably brings to the creation of more and more complicated devices. The rise of cyber-physical systems, which are on their way to become massively pervasive, brings the complexity level to the next step and open many new challenges. First, the descriptive power of standard temporal logics is no more sufficient to handle all kind of requirements the designers need (consider, for example, non-functional requirements). Second, the standard model checking techniques are unable to manage such level of complexity (consider the well-known curse of state space explosion). In this thesis, we leverage machine learning techniques, active learning, and optimization approaches to face the challenges mentioned above. In particular, we define signal measure logic, a novel temporal logic suited to describe non-functional requirements. We also use evolutionary algorithms and signal temporal logic to tackle a supervised classification problem and a system design problem which involves multiple conflicting requirements (i.e., multi-objective optimization problems). Finally, we use an active learning approach, based on Gaussian processes, to deal with falsification problems in the automotive field and to solve a so-called threshold synthesis problem, discussing an epidemics case study.

During the last 20 years, model-based design has become a standard practice in many fields such as automotive, aerospace engineering, systems and synthetic biology. This approach allows a considerable improvement of the final product quality and reduces the overall prototyping costs. In these contexts, formal methods, such as temporal logics, and model checking approaches have been successfully applied. They allow a precise description and automatic verification of the prototype's requirements. In the recent past, the increasing market requests for performing and safer devices shows an unstoppable growth which inevitably brings to the creation of more and more complicated devices. The rise of cyber-physical systems, which are on their way to become massively pervasive, brings the complexity level to the next step and open many new challenges. First, the descriptive power of standard temporal logics is no more sufficient to handle all kind of requirements the designers need (consider, for example, non-functional requirements). Second, the standard model checking techniques are unable to manage such level of complexity (consider the well-known curse of state space explosion). In this thesis, we leverage machine learning techniques, active learning, and optimization approaches to face the challenges mentioned above. In particular, we define signal measure logic, a novel temporal logic suited to describe non-functional requirements. We also use evolutionary algorithms and signal temporal logic to tackle a supervised classification problem and a system design problem which involves multiple conflicting requirements (i.e., multi-objective optimization problems). Finally, we use an active learning approach, based on Gaussian processes, to deal with falsification problems in the automotive field and to solve a so-called threshold synthesis problem, discussing an epidemics case study.