Data-driven control for SISO feedback linearizable systems with unknown control gain

Paulo Tabuada, Lucas Fraile

Abstract: In this paper we consider the problem of controlling an unknown system without making use of prior data or training. By relying on a feedback linearizability assumption we show how, based on prior ideas by Fliess and co-workers on model-free control, it is possible to accomplish such objective. The key idea is to learn a model that is only valid at the current state and re-learn this model as time progresses. Since this requires learning two real numbers rather than functions, it results in an approach quite different from: 1) deep learning since it requires no prior data neither large amounts of data; 2) reinforcement learning since it converges much faster and does not suffer from the curse of dimensionality.