Convex Neural Network-Based Cost Modifications for Learning Model Predictive Control

Developing model predictive control (MPC) schemes can be challenging for systems where an accurate is not available, or too costly to develop. With the increasing availability of data and tools treat them, learning-based MPC has late attracted wide attention. It recently been shown that adapting only model, but also its cost function conducive achieving optimal closed-loop performance when cannot provided. In learning context, this modification performed via parametrizing adjusting parameters via, e.g., reinforcement (RL). framework, simple parametrizations effective, underlying theory suggests rich in principle useful. paper, we propose such a parametrization using class neural networks (NNs) preserves convexity. This choice avoids creating difficulties solving problem sensitivity-based solvers. addition, ensures nominal stability resulting scheme. Moreover, detail how applied economic problems generic therefore does necessarily fulfill any specific property.