Bayesian Optimization attempts for European XFEL tuning

It reads the configuration from a dict or a json / yaml file and starts the Bayesian optimization.

Important notes for new BO Practitioner: The code in this repository serves as a good initial point to understand how BO can be implemented. For a more systematic and maintainable solution, we kindly refer to the Xopt package, which provides more BO variants and is designed to be a generic optimizer.

The results are published in the IPAC'23 proceeding: https://doi.org/10.18429/JACoW-IPAC-23-THPL028 (Pre-proceeding version)

Additional Notes

• Currently a basic plotting is implemented in utils.plot_progress, but using it in VS-Code shows some flickering (see this issue), in plain jupyter notebook it works smoothly.
• pydoocs require additional package jpeg that needs to be installed via conda conda install -c conda-forge jpeg.

Problem

The json configuration files contain the following tuning tasks

• SASE1_CAX_CAY.json SASE1 tuning with undulator air coil correctors
• SASE1_matching_quads.json SASE1 tuning with quadrupole magnets
• SASE2_CAX_CAY.json, SASE2_CAX_CAY_Launch.json SASE2 tuning with air coil correctors
• SASE2_matching_quads.json SASE2 tuning with quadrupole magnets

---

Repository Structure

• simplebo.py Implements the basic BO logic
• conf/*.json Configuration files for optimization tasks
  • test_rosenbrock.json A template for optimizating a test function
• utils.py Utility functions, e.g. the proximial acquisition function implemented in Xopt

Usage

Installing Requirements

It is recommended to use a virtual environment, e.g. conda. With the environment activated, simply do

pip install -r requirements.txt

Using BO

(For pydoocs users): Simply define your problem in a json file structured as the Ocelot optimizer configuration. Load the config into the SimpleBO object and start with optimize.

---

Tasks

TODO:

• Basic BO functionality
  • BO Loop
  • Basic logging
• Advanced control:
  • Custom start condition: random initialization, start from current setting
  • Fine tuning the Acq, Prior...
  • Step size control: hard / proximal biasing
  • More information in the logging
• Preliminary Tests:
  • BO test with simple mathematical functions
  • I/O test with pydoocs
  • Environment test on xfeluser server
  • (Check compatibility) Environment test on control room PCs
  • Provide a beam time note book with the defined procedures
• Other features
  • live progress plot
• (If enough time): try direct interface with Xopt, this will be a sustainable and more preferred way for productive tools.