This website requires JavaScript.

Learning a convex cost-to-go for single step model predictive control

E. M. TuranZ. MdoeJ. J\"aschke
Feb 2024
For large systems that consider uncertainty the online solution of model predictive control problems can be computationally taxing, and even infeasible. This can be offset by using a shorter horizon, however this can in turn result in poor controller performance. In this work we consider the task of learning convex objective terms to allow the use of a short, potentially single step, control horizon to reduce the online computational cost. We consider two surrogates: (1) a convex interpolating function and (2) an input-convex neural network. We highlight that irrespective of the choice of surrogate the behaviour of the surrogate near the origin and ability of the surrogate to describe the feasible region are key concerns for the closed loop performance of the new MPC problem. We address these concerns by tailoring the design of the surrogate to ensure good performance in both aspects.The paper concludes with a numerical example, showing the clear and significant reduction in computational complexity through the use of these convex surrogates.
发布时间 · 被引用数 · 默认排序
发布时间 · 被引用数 · 默认排序