The paper addresses data-driven control based on input–output data in the absence of an underlying dynamic model. It proposes a dissipativity learning control (DLC) framework which involves the data-based learning of the dissipativity property of the control system, followed by a dissipativity-based controller design procedure. Specifically, independent component analysis and parametric distribution inference are adopted to estimate a polyhedral region of input–output trajectory samples, whose dual cone characterizes the dissipativity property; subsequently, an optimal controller in the L2 sense is designed by solving a nonlinear semidefinite programming problem. The applicability of the proposed method is demonstrated by case studies on regulating control of a polymerization reactor and tracking control of an oscillatory chemical reactor.
Bibliographical noteFunding Information:
This work was supported by National Science Foundation (NSF-CBET) and a Doctoral Dissertation Fellowship (DDF) of University of Minnesota for Wentao Tang.
- Data-driven control
- Dissipative systems
- Machine learning