Learning optimal switching feedback controllers from data

In this paper we present a data-driven approach for synthesizing optimal switching controllers directly from experimental data, without the need of a global model of the dynamics of the process. The set of controllers and the switching law are learned by using a coordinate descent strategy: for a fixed switching law, the controllers are sequentially optimized by using stochastic gradient descent iterations, while for fixed controllers the switching law is iteratively refined by unsupervised learning. We report examples showing that the approach performs well when applied to control processes characterized by hybrid or nonlinear dynamics, outperforming control laws that are single-mode (no switching) or multi-mode but with the switching law defined a priori.