Data-driven MPC applied to non-linear systems for real-time applications

Model Predictive Control (MPC) is a traditional technique widely employed on the control of constrained non-linear systems. In light of the increasing popularity of neural networks, data-driven MPC has emerged as an alternative to alleviate the computation burden of traditional control strategies. This paper aims to replace the constrained optimization problem by training a feed-forward neural network with data collected from an offline MPC simulation. The network's performance is then compared to that of traditional MPC using the Van der Pol oscillator as toy example. Finally, the convergence of the network training error is analytically proven by extending the analysis of neural tangent kernels (NTK) to underparameterized networks.