Основні поняття
Reinforcement learning of Koopman models enhances performance in economic nonlinear model predictive control.
Анотація
The content discusses the application of end-to-end reinforcement learning of Koopman models for optimal performance in economic nonlinear model predictive control (NMPC). It introduces a method to train dynamic surrogate models using reinforcement learning algorithms, focusing on task-optimal performance. The study compares the effectiveness of models trained through system identification and reinforcement learning techniques. Results demonstrate that end-to-end trained models outperform those trained using traditional methods, showcasing adaptability to changes in control settings without retraining.
The article is structured as follows:
- Introduction to data-driven surrogate models for NMPC.
- Comparison between system identification and reinforcement learning for training dynamic surrogate models.
- Focus on learning model-free control policies through deep reinforcement learning with continuous action spaces.
- Discussion on post-optimal sensitivity analysis of convex problems and its application in deep learning projects.
- Methodology section detailing the end-to-end refinement of Koopman models for MPC applications through RL.
- Numerical experiments section discussing case studies based on a continuous stirred-tank reactor model, including NMPC and eNMPC scenarios.
- Results showing the superior performance of end-to-end trained Koopman models in both NMPC and eNMPC applications.
Статистика
Data-driven surrogate models reduce computational burden in (e)NMPC.
End-to-end training improves controller performance without retraining.