[DATA SUBMISSION] DE-NO-r1-B, DE-j-r2-B, and DE-CoBi-r1-L on the bbob-mixint suite

ryojitanabe commented 2 months ago

Dear Organizers of COCO,

I would like to submit the benchmarking results of three optimizers (DE-NO-r1-B, DE-j-r2-B, and DE-CoBi-r1-L) on the bbob-mixint suite. Could you please upload them to the COCO data archive?

Best regards, Ryoji

Reference

Ryoji Tanabe: Benchmarking Parameter Control Methods in Differential Evolution for Mixed-Integer Black-Box Optimization, Proc. ACM Genetic and Evolutionary Computation Conference (GECCO2024), arXiv

Description of the Algorithm(s)

The above paper benchmarked parameter control methods for scale factor and crossover rate in differential evolution (DE) on the bbob-mixint suite. The results show that using a suitable parameter control method significantly improves the performance of DE on the bbob-mixint suite. I benchmarked 160 DE configurations, where I used 10 parameter control methods, 8 mutation strategies, and 2 repair methods. However, uploading the benchmarking data of the 160 optimizers to the COCO data archive is not possible. For this reason, I would like submit only the benchmarking results of the following three optimizers. Here, in the notation "DE-X-Y-Z", X represents the type of parameter control method, Y represents the type of mutation strategy, and Z represents the type of repair method.

DE-NO-r1-B: DE with no parameter control method, the rand/1 strategy, and Baldwinian repair method. This is the most classical DE algorithm.
DE-j-r2-B: DE with the parameter control method in jDE, the rand/2 strategy, and Baldwinian repair method. This DE performs the best out of the 160 DE configurations for 160 dimensions.
DE-CoBi-r1-L: DE with the parameter control method in CoBiDE, the rand/1 strategy, and Lamarckian repair method. This DE performs the best out of the 160 DE configurations for 5 dimensions.

Both DE-j-r2-B and DE-CoBi-r1-L show better performance than CMA-ES-pycma, CMA-ESwM, and cmaIH1e-1 for high dimensions and larger budgets of function evaluations.

Link to Data

https://drive.google.com/file/d/1PzABfbLLO2lcYaCU02DXdrqRROYs2MYb/view?usp=sharing

Optional: Source Code of Experiment

https://github.com/ryojitanabe/de_bbobmixint

brockho commented 1 month ago

Dear @ryojitanabe,

Many thanks for submitting your GECCO-2024 data files to us. All files are clean, postprocess without problems and the documentation is clear as well (including the link to the paper and the source code - ideal!). I put them now online on our current data archive page and they are also available already via the postprocessing (after running cocopp.archives.update_all()).

ryojitanabe commented 1 month ago

Thank you very much for the upload!

numbbo / coco