What is/can be provided on the ppdata webpage?

[brockho]

Finding the right link in the ppdata archive for displaying an algorithm of interest is difficult in general and only easy if you know the year of publication for the algorithm. One suggestion from today is to provide additional ppdata output for each algorithm (or for a small list of algorithms) for the "larger" test suites (aka bbob, bbob-noisy and bbob-biobj for the time being). Those can then be linked directly on the algorithm lists of the COCO data archive.

Suggestions, to be put here for discussion:

• 2009/2016 algorithms in background
• A few (always the same) reference algorithms in the foreground:
  • RS-5-initIn0, IPOP-ACTCMA-ES, BFGS-scipy, NelderDoerr for bbob
  • BFGS, IPOP-ACTCMA-ES for bbob-noisy
  • TPB, NSGA-II-platypus for bbob-biobj

[nikohansen]

• for bbob
  • BFGS-scipy does not evaluate all zeros in the beginning
  • we should rather use 2020/CMA-ES instead of IPOP-ACTCMA-ES? They look very similar anyways.
  • display also 2020/SLSQP-11 (which is 2-3 times faster than BFGS) or instead of BFGS?

[nikohansen]

For selecting a Nelder Mead version: these are our data. We could select the standard (red cross) and additionally one of the two others which are available in 40-D which would not be NELDERDOERR)? Displaying only one of the highly tweaked versions would probably lead to an unrealistic expectation.

[nikohansen]

We decided to use adapt-Nelder... as the best compromise between (relatively good) performance and (relatively little) algorithm modifications. Namely, the three factors for expansion, contraction, and shrinking are set depending on the search space dimension like $c \pm b/n$ for some $c, b\in[0.5,2]$ depending on the factor.

[nikohansen]

• re. bbob/SLSQP: it seems that '20/slsqp-11' is the best choice, see https://numbbo.github.io/cocopp-playground/ppdata-slsqp
• re. bbob/BFGS: we should chose the simplest of these https://numbbo.github.io/cocopp-playground/ppdata-bfgs-selected ? @brockho do you have more insight here, I suspect it is the one you recommended? Looking at the ill conditioned function and/or F10 and F11, the -P- versions seem preferable though.