January 9, 2023Adnane Saoud (CentraleSupelec)
In the last few years, learning-based techniques have shown a great success to control complex cyber-physical systems (CPS), where learning-based tools are used either to provide mathematical models of a system, based on which one can synthesize a controller, or to learn directly the controller. However, the use of learning-based techniques in the context of safety critical CPS is particularly problematic, since learning-based components are typically viewed as black box-type systems, lacking formal guarantees. In this talk, we first recall how symbolic control techniques can be used in the context of the control of CPS. In the second part of the talk, we focus on how to provide guarantees when learning is used at the model’s level, namely we show how to learn a symbolic model from data, while providing formal guarantees. This symbolic model will be then used to construct the controller for complex specifications, such as linear temporal logics, based on existing tools in the formal methods community. In the last part of the talk, we focus on the use of learning at the controller’s level. We first use tools from viability theory to develop a general framework for the computation of interval controlled invariants for nonlinear systems. We then show how this framework makes it possible to provide safety guarantees when using neural networks to the control of dynamical systems.