Supervised learning for robust predictive control: Safe and tunable approach

The paper introduces a safe and real-time tunable approximation of the optimal controller based on a neural network (NN). The approximation enables full integration into industrial applications, easing the computational burden and memory footprint of an optimal control law. Specifically, we propose a neural-network-based approximation of the optimal control law derived from a robust Model Predictive Control (MPC) that preserves adherence to input and state physical constraints, recursive feasibility, and robust closed-loop stability. Additionally, the proposed method is able to tune the approximate control law in real-time, keeping the controller safe and feasible. The real-time tunability improves the performance of the platforms with limited computational and memory resources. Outperforming implicit MPC, the proposed method eliminates the computation of real-time optimisation, and the approximated control law is suitable for certification. Moreover, the computation of control inputs does not rely on external libraries or solvers. In comparison with explicit MPC, the proposed method avoids the necessity to evaluate the explicit solution map or piecewise affine functions, as it removes the real-time point location problem and, therefore, reduces the memory footprint. The advantages of the proposed method are showcased using a double integrator benchmark and quadrotor position control to the origin.