Boris Houska, Xuhui Feng (presented by Boris)

ShanghaiTech University, China

Friday, September 02, 2016, 10:00

Room 01-210, Georges-Köhler Allee 102, Freiburg 79110, Germany

This talk is about a novel control scheme, named self-reflective model predictive control, which takes its own limitations in the presence of process noise and measurement errors into account. In contrast to existing output-feedback MPC and persistently exciting MPC controllers, self-reflective MPC controllers do not only propagate a matrix-valued state forward in time in order to predict the variance of future state-estimates, but they also propagate a matrix-valued adjoint state backward in time. This adjoint state is used by the controller to compute and minimize a second order approximation of its own expected loss of control performance in the presence of random process noise and inexact state estimates. A second part of the talk introduces a real-time algorithm, which can exploit the particular structure of the self-reflective MPC problems in order to speed-up the online computation time. It is shown that, in contrast to generic state-of-the-art optimal control problem solvers, the proposed algorithm can solve the self-reflective optimization problems with reasonable additional computational effort compared to standard MPC. The advantages of the proposed real-time scheme are illustrated by applying it to a benchmark predator-prey-feeding control problem.