Control by Interconnection using Reinforcement Learning

Abstract

The dynamics of many physical processes can be described by port-Hamiltonian (PH) models where the importance of the energy function can be seen. In Control by Interconnection (CbI), the controller is another PH system that is connected to the plant through a power preserving interconnection to add up the energy functions. However, a major issue in this is that the choice of Casimir function and controller Hamiltonian is left to the discretion of the designer and requires experience to make a good choice. In this thesis, an attempt is made to eliminate this problem by using machine learning algorithms (in particular, reinforcement learning) to let the computer "learn" the best controller design. Moreover, the assumption that both the plant and the controller must be passive leads to what is known as the dissipation obstacle, which means that dissipation is allowed only on those states/coordinates of the energy function which do not require shaping. This imposes restrictions on the applications. Here, it is attempted to try to go beyond this dissipation obstacle and achieve a dynamic feedback controller.