40m QIL Cryo_Lab CTN SUS_Lab TCS_Lab OMC_Lab CRIME_Lab FEA ENG_Labs OptContFac Mariner WBEEShop
  Mariner elog  Not logged in ELOG logo
Entry  Thu Mar 4 17:04:52 2021, Paco, General, Design specs, Silicon TM dichroic coatings for phase I ETM_coating_candidates.pdfITM_coating_candidates.pdf
    Reply  Wed Mar 17 19:51:42 2021, Paco, General, Design specs, Silicon TM dichroic coatings for phase I ETM_R_210317_1927.pdfETM_Layers_210317_1927.pdfETM_nominal_cornerPlt.pdf
       Reply  Wed Mar 17 21:24:27 2021, gautam, General, Design specs, Silicon TM dichroic coatings for phase I 
          Reply  Wed Mar 24 17:36:46 2021, Paco, General, Design specs, Least common multiple stacks and varL cost ETM_Layers_210323_0925.pdf
             Reply  Fri Apr 2 19:59:53 2021, Paco, General, Design specs, Differential evolution strategies diffevostrategies.pdf
                Reply  Fri Jun 4 11:09:27 2021, Paco, General, Design specs, HR coating tolerance analysis 
          Reply  Wed Mar 24 17:42:50 2021, Paco, General, Design specs, Silicon TM dichroic coatings for phase I 
Message ID: 10     Entry time: Fri Apr 2 19:59:53 2021     In reply to: 8     Reply to this: 15
Author: Paco 
Type: General 
Category: Design specs 
Subject: Differential evolution strategies 

Differential evolution strategies 'benchmarking' for thin film optimization

Since I have been running the ETM/ITM coatings optimization many times, I decided to "benchmark" (really just visualize) the optimizer trajectories under different strategies offered by the scipy.optimize implementation of differential evolution. This was done by adding a callback function to keep track the convergence=val at every iteration. From the scipy.optimize.differential_evolution docs, this "val represents the fractional value of the population convergence".

Attachment 1 shows a modest collection of ~16 convergence trajectories for ETM and ITM as a function of the iteration number (limited by maxiter=2000) with the same targets, weights, number of walkers (=25), and other optimization parameters. The vertical axis plots the inverse val (so tending to small numbers represent convergence).

tl;dr: Put simply, the strategies using "binary" crossover schemes work better (i.e. faster) than "exponential" ones. Will keep choosing "best1bin" for this problem.

Attachment 1: diffevostrategies.pdf  434 kB  Uploaded Fri Apr 2 21:26:11 2021  | Hide | Hide all
diffevostrategies.pdf
ELOG V3.1.3-