Dissipativity Learning Control through Estimation from Online Trajectories*

Dissipativity, as a behavioral characterization of nonlinear dynamic systems, contains sufficient information for designing controllers. Such dissipativity information, if learned from data instead of being derived from the model, can be used for model-free data-driven control. The learning of dissipativity for general multi-input-multi-output nonlinear systems, however, is difficult and typically demands large offline trajectory datasets. In this work, a method for inferring the dissipativity information from a single trajectory collected online is proposed, which yields a conservative estimation of the set of supply rate functions, thus guaranteeing the stabilizing model-free control performance. The efficacy of the proposed approach is demonstrated with a two-phase reactor case study.