The integration of data-driven algorithms into feedback loops is transforming modern engineering. Neural networks are now used to model, estimate and control dynamical systems, and learning-based components increasingly take decisions inside and among controlled systems. Yet, while these tools offer remarkable flexibility, they seldom come with the guarantees required for deployment in safety-critical or societally sensitive applications: generalization guarantees from a learning perspective and, in control-theoretic terms, certificates of stability, robustness and safety.
For the control systems community, the challenge is therefore not only to adopt learning tools, but to analyze, guide and certify them. This workshop aims to open discussions on this challenge at three nested scales[, from the learning model to the closed loop and to interacting systems]. At the level of the learning model itself, neural networks can be viewed as dynamical systems, which naturally calls for control-theoretic analysis of their behavior, robustness and training. At the level of the closed loop, learned models, controllers and estimators should be embedded in feedback architectures in ways that preserve stability and safety guarantees. At the level of interacting systems, multiple learning-enabled agents[, whether engineered or human,] pursue their own objectives while sharing resources and influencing one another, raising questions of strategic interaction, equilibrium and fairness. Across these scales, the common thread is the same: using the tools of control theory to explain, enhance and certify learning-based systems.
The program combines a mini-course on neural networks from a control perspective, invited talks by established and early-career researchers, and junior sessions for PhD students. Our goal is a shared picture of available tools and open problems, including scalability of certificates, conservatism of guarantees, and the gap between theory and implementation.