Learning-Based Event-Triggered MPC With Gaussian Processes Under Terminal Constraints.

Area of Science:

• Control Systems Engineering
• Machine Learning
• Robotics

Background:

• Event-triggered control reduces computational load by executing tasks only when necessary.
• Model predictive control (MPC) offers optimal control but requires accurate system models.
• Unknown system dynamics pose a challenge for traditional MPC and event-triggered strategies.

Purpose of the Study:

• To develop a novel learning-based event-triggered MPC for systems with initially unknown dynamics.
• To ensure recursive feasibility and stability of the control system.
• To reduce the frequency of control task executions while maintaining performance.

Main Methods:

• Utilizing Gaussian process (GP) regression for predictive state estimation.
• Formulating optimal control problems (OCPs) based on GP predictions and terminal constraints.
• Deriving an event-triggered condition based on recursive feasibility to minimize OCP solutions.
• Analyzing the convergence properties of the closed-loop system.

Main Results:

• The proposed method effectively handles initially unknown system dynamics.
• The event-triggered condition ensures that OCPs are solved only when the prediction error surpasses a defined threshold.
• Convergence analysis demonstrates that the system state enters the terminal set in finite time under certain uncertainty conditions.
• Validation through a tracking control problem confirms the practical effectiveness of the approach.

Conclusions:

• The learning-based event-triggered MPC provides an efficient and stable control solution for systems with unknown dynamics.
• This approach significantly reduces computational burden compared to traditional MPC.
• The method shows promise for real-world applications requiring adaptive and resource-efficient control.