Unifying Direct and Indirect Learning for Safe Control of Linear Systems (2504.18331v1)

Abstract: This paper aims to learn safe controllers for uncertain discrete-time linear systems under disturbances while achieving the following two crucial goals: 1) integration of different sources of information (i.e., prior information in terms of physical knowledge and posterior information in terms of streaming data), and 2) unifying direct learning with indirect learning. These goals are achieved by representing a parametrized data-driven constrained matrix zonotope form of closed-loop systems that is conformant to prior knowledge. To this end, we first leverage collected data to characterize closed-loop systems by a matrix zonotope and then show that the explainability of these closed-loop systems by prior knowledge can be formalized by adding an equality conformity constraint, which refines the matrix zonotope obtained by data to a constrained matrix zonotope. The prior knowledge is further refined by conforming it to the set of models obtained from a novel zonotope-based system identifier. The source of data used for zonotope-based system identification can be different than the one used for closed-loop representation, allowing to perform transfer learning and online adaptation to new data. The parametrized closed-loop set of systems is then leveraged to directly learn a controller that robustly imposes safety on the closed-loop system. We consider both polytope and zonotope safe sets and provide set inclusion conditions using linear programming to impose safety through {\lambda}-contractivity. For polytope safe sets, a primal-dual optimization is developed to formalize a linear programming optimization that certifies the set inclusion. For zonotope safe sets, the constrained zonotope set of all next states is formed, and set inclusion is achieved by ensuring the inclusion of this constrained zonotope in a {\lambda}-scaled level set of the safe set.