ABSTRACT Model predictive control (MPC) is a powerful control strategy that delivers optimal performance while ensuring the satisfaction of safety and operational constraints at all times. However, its reliance on online optimization significantly increases computational demand, limiting its deployment in safety‐critical systems with limited onboard computational resources. A promising solution is to replace the MPC optimizer with a neural network trained to approximate the MPC policy. However, the effectiveness of this neural MPC framework relies heavily on both the quality and efficiency of the training process. Supervised learning requires extensive data collection, whereas unsupervised methods often suffer from convergence issues and instability. To address these challenges, this paper introduces a hybrid training method that leverages the strengths of both supervised and unsupervised learning approaches to accelerate training and improve the performance of the resulting neural MPC framework. Extensive numerical experiments show that the proposed training method is up to 147 times faster than conventional approaches, while yielding neural MPC frameworks that improve safety and tracking performance by up to 99.37% and 84.7%, respectively. The practicality and applicability of the hybrid training approach are further validated through its deployment in neural MPC designs for drone hovering and thermal regulation tasks.
Li et al. (Thu,) studied this question.