Adaptive Tuning of Online Feedback Optimization for Process Control Applications

Online Feedback Optimization creates controllers by embedding an optimization algorithm inside a feedback loop. This is useful for processes where you lack a complete mathematical model. The difficulty is choosing the right parameters for the optimizer because they do not map directly to physical quantities and their effects are hard to predict. The authors propose making the tuning automatic by measuring how sensitive the performance objective is to changes in those parameters. They focus on scaled projected gradient descent and reduce the operator's choices to two simple numbers: one that sets the maximum allowed change in the control inputs from one step to the next, and one that sets the maximum allowed change in the objective value. These scalars replace the usual collection of hard-to-set algorithm parameters. No extra system information or repeated experiments are required. The method is tested in simulation on a gas-lift oil-production system and on a continuously stirred tank reactor. In both cases the adaptive version produces better closed-loop behavior than the same controller with parameters chosen by hand. The core idea is to let the controller observe its own sensitivity and adjust its internal step sizes on the fly.