Hi,
I'm working on the ex optimization when passing through a 4 bends chicane.As our initial energy spread is too large,and we want to put into some
other magnets,may be sextupoles or quadrupoles,to get a better performance.
I want to know is there some key parameters to get a better solution in optimization?(n_passes,n_evaluations,etc...)How to definite the appropriate values for them?
And you had claimed that "simplex" may converge to a local solution,is there any methods to prevent this?
Thanks a lot.
Parameters to choose in optimization
Moderators: cyao, michael_borland
-
- Posts: 1933
- Joined: 19 May 2008, 09:33
- Location: Argonne National Laboratory
- Contact:
Re: Parameters to choose in optimization
The best parameters will depend to some extent on the problem, but here are some suggestions
This will perform 3 simplex passes with decreasing initial simplex size; each will consist of at most ~1500 evaluations of the function. After each set of 3 passes, a restart will be performed starting from the best result, with the original simplex size. This helps to avoid local minima. 10 restarts will be performed.
Monitor the optimizationFunction and bestOptimizationFunction values in the "final" output file (requested in the run_setup command). It you see that it is not improving, you can using Ctrl-C or the interrupt_file feature to terminate the optimization and inspect the results.
Other important parameters are the target and tolerance values. Set the target to the value you need for the sum of the optimization terms. Set the target to much less than this, maybe by a factor of 1e6.
--Michael
Code: Select all
n_evaluations = 1500
n_passes = 3
n_restarts = 10
Monitor the optimizationFunction and bestOptimizationFunction values in the "final" output file (requested in the run_setup command). It you see that it is not improving, you can using Ctrl-C or the interrupt_file feature to terminate the optimization and inspect the results.
Other important parameters are the target and tolerance values. Set the target to the value you need for the sum of the optimization terms. Set the target to much less than this, maybe by a factor of 1e6.
--Michael