2024 | Ján Drgoňa, Member, IEEE, Aaron Tuor, and Draguna Vrabie, Member, IEEE
The paper introduces Differentiable Predictive Control (DPC), a method for offline learning of constrained neural control policies for nonlinear dynamical systems with performance guarantees. DPC leverages automatic differentiation (AD) to compute the sensitivities of the model predictive control (MPC) objective function and constraints penalties, enabling direct policy gradients. The method guarantees closed-loop stability and constraint satisfaction through probabilistic guarantees based on indicator functions and Hoeffding's inequality. Empirical results demonstrate that DPC can stabilize unstable systems, track time-varying references, and satisfy nonlinear state and input constraints. Compared to alternative approaches, DPC offers faster execution times, reduced memory requirements, and better scalability, making it suitable for real-time applications. The paper also includes a detailed formulation of the DPC problem, a probabilistic verification method, and numerical studies comparing DPC with other control methods.The paper introduces Differentiable Predictive Control (DPC), a method for offline learning of constrained neural control policies for nonlinear dynamical systems with performance guarantees. DPC leverages automatic differentiation (AD) to compute the sensitivities of the model predictive control (MPC) objective function and constraints penalties, enabling direct policy gradients. The method guarantees closed-loop stability and constraint satisfaction through probabilistic guarantees based on indicator functions and Hoeffding's inequality. Empirical results demonstrate that DPC can stabilize unstable systems, track time-varying references, and satisfy nonlinear state and input constraints. Compared to alternative approaches, DPC offers faster execution times, reduced memory requirements, and better scalability, making it suitable for real-time applications. The paper also includes a detailed formulation of the DPC problem, a probabilistic verification method, and numerical studies comparing DPC with other control methods.