Tuesday, August 22, 2023, 11:00 - 11:59
In Model Predictive Control, feasible iterate optimization algorithms are a possibility to overcome small computation times due to their possibility of stopping at a suboptimal, but feasible point.
Previous work developed a Feasible Sequential Linear Programming (FSLP) algorithm, that was successfully applied to time-optimal control problems. The algorithm performed well on these problems, but some weaknesses could be identified by exhaustive testing on a broader class of test problems.
In this talk, several enhancements of the original FSLP algorithm are presented. Starting from improving on the feasibility iteration strategy which keeps all iterates feasible, Anderson Accelerated FSLP achieved a speedup of at least 40% in computation time in comparison to the original version. The talk ends with recent work on funnel sequential quadratic programming methods, that can overcome difficulties in initializing the algorithm at a feasible point.