Solution of a non-convex optimization arising in PI/PID control design (Q1858859)

This is a well-written paper dealing with robust proportional-integral (PI) controller design for linear systems. Designing a stabilizing PI controller subject to closed-loop robustness constraints is in general a non-convex optimization problem in the space of controller parameters. In order to handle this non-convexity, the authors reformulate the design problem in the frequency domain (Nyquist plot). Closed-loop robustness constraints can be formulated as Euclidean distance constraints in the Nyquist plot. It is shown that at a given frequency, the two-dimensional set of stabilizing PI controller parameters satisfying robustness constraints is an ellipse. When sweeping over the frequency, the union of these ellipses forms a connected region in the parameter plane. Obviously, determining the boundary of this region amounts to solving the PI controller design problem. Once the boundary is computed, one can even optimize over controller parameters to ensure some additional performance such as the minimization of the integral of error subject to a bounded disturbance input. The authors propose an algorithm to construct the boundary of the parameter region using the notion of principal points associated with a differentiable complex-valued mapping and a path-following algorithm. The original non-convex problem can thus be solved globally without a gradient-based approach. Numerical examples are described to illustrate the algorithm.