New networked operator parametrization allows unconstrained learning!
Our paper "Unconstrained learning of networked nonlinear systems via free parametrization of stable interconnected operators" got accepted for the European Control Conference 2024 (ECC24)!
This paper introduces a novel parametrization technique for nonlinear networked incrementally L2-bounded operators in discrete time. The key innovation is a "free" parametrization that yields a sparse large-scale operator with bounded incremental L2 gain regardless of parameter values. This enables efficient parameter optimization using unconstrained gradient descent, making it applicable in large-scale optimal control and system identification. Moreover, it allows embedding prior knowledge about system interconnection topology and stability directly into the designed distributed operator. The method can incorporate state-of-the-art Neural Network (NN) parametrizations seamlessly and is demonstrated through a simulation example, highlighting its superiority over standard NN-based identification methods without enforced priors on system topology and stability.
NCCR Automation