Speaker
Description
Deterministic feedback control is critical in scenarios where controllers must perform actions or system corrections in a predictable, consistent, and fast manner based on observations, which is crucial for real-time operations in lasers and accelerators. These complex multi-input multi-output (MIMO) systems typically lack comprehensive models linking observations to system errors, leading to reliance on stochastic or dither-and-search methods within feedback loops. These traditional approaches can be time-consuming and must be repeated with system condition changes. This paper presents a simplified reinforcement learning (RL) method that leverages differentiable reward functions to train deterministic control policies. Our approach facilitates direct, gradient-based optimization, enabling efficient and predictable corrections of system errors in real-time and significantly reducing system optimization time and cost. We validate this method through examples of feedback controls in lasers and accelerators, including an experimental demonstration in laser combining. These results highlight the method's effectiveness, offering a practical and efficient solution for fast, deterministic feedback controls in complex systems.
| In which format do you inted to submit your paper? | LaTeX |
|---|