THA2WE01
SLAC, Menlo Park, California, USA
UCSC, Santa Cruz, California, USA
| Paper | Title | Page |
|---|---|---|
THA2WE01 |
Bayesian Optimization for Online FEL Tuning at LCLS | |
|
||
| The Linac Coherent Light Source changes configurations 2 to 5 times per day, necessitating fast tuning strategies to reduce setup time for successive experiments. To this end, we employ a Bayesian approach to transport optics tuning to quickly optimize groups of quadrupole magnets. The power of Bayesian optimization lies in its ability to employ a probability distribution to represent the most likely region of a control feature space to optimize an objective. A Gaussian process allows us to employ kernel learning to modify the Bayesian likelihood of the machine response from observed data and learned characteristics of the machine response with respect to the controlled parameters. We build Bayesian priors and response correlations from historical LCLS run data of FEL pulse energy versus quadrupole magnet strengths, and use this to simultaneously optimize quadrupoles. Here, we introduce Bayesian optimization with Gaussian processes, and then describe our approach to training the optimizer on historical LCLS data. | ||
|
|
Slides THA2WE01 [2.511 MB] | |
| Export • | reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml) | |