presentation:
- general case
- who did what/how it was split
- how we did pipeline
- explain each of steps in pipeline
- try to put choices we were faced with and conseuqeunces/shortcomings of decisions


- General Case:
 Steering through the whole CLARA lattice with optimisation and RL techniques, achieving closest possible sim2real matching
 
- Who Did What:
  + Francessco: Sim2Real matching with tomography, beam optics, multi-fidelity BO 
  + Eiad: Beam optics and SVD to find a good reference tune and the key quadrupoles and correctors for a given target
  + Ivan: numerical optimistation techniques for S04, which was then used with Eiad & Franccessco's beam optics 
  + Manuel: Developing a robust measure for real screen image processing, using a DALIA job to process the whole 0.8TB!
  + Amy: Developed the framework and simulation environment, and built a model based RL agent against it, making use of Eiad's beam optics SVD controls 
 
  
- Pipeline
  - Simulation generates a variation on a realistic beam and lattice, with small perturbations
  - Set nominal optics from lookup table (optics matching)
  - SVD of the correctors vs BPM for the target beam gives us the most significant control variables to center the current beam at the selected optics
  - Do steering simple optimsiation - typically 1 iteration, worst case 4
  - Focus optimisation - Use SVD again on the combined correctors and quadrupoles to find the significant control variables for focussing 
  - Use the numeric optimisation or RL model to tweak the optics matched tune, using focussing only
  - Use the robust screen measure to read screen images
  - Tweak and read the target screen until a good match -^
  - Finally one last round of SVD on the correctors to steer with the focussed optics


Image processing note:
- Noisy beam image is an example of the cleaning of the data
- Weak beam image is an example of a bad clean, where data is discarded