We employ Difference of Log-Sum-Exp neural networks to generate a data-driven feedback controller based on Model Predictive Control (MPC) to track a given reference trajectory. By using this class of networks to approximate the MPC-related cost function subject to the given system dynamics and input constraint, we avoid two of the main bottlenecks of classical MPC: the availability of an accurate model for the system being controlled, and the computational cost of solving the MPC-induced optimization problem. The former is tackled by exploiting the universal approximation capabilities of this class of networks. The latter is alleviated by making use of the difference-of-convex-functions structure of these networks. Furthermore, we show that the system driven by the MPC-neural structure is practically stable.
On the Use of Difference of Log-Sum-Exp Neural Networks to Solve Data-Driven Model Predictive Control Tracking Problems
Possieri Corrado
2021
Abstract
We employ Difference of Log-Sum-Exp neural networks to generate a data-driven feedback controller based on Model Predictive Control (MPC) to track a given reference trajectory. By using this class of networks to approximate the MPC-related cost function subject to the given system dynamics and input constraint, we avoid two of the main bottlenecks of classical MPC: the availability of an accurate model for the system being controlled, and the computational cost of solving the MPC-induced optimization problem. The former is tackled by exploiting the universal approximation capabilities of this class of networks. The latter is alleviated by making use of the difference-of-convex-functions structure of these networks. Furthermore, we show that the system driven by the MPC-neural structure is practically stable.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.