Section: New Software and Platforms

Platforms

New developments around COCO

There were two public releases of the COCO software this year including quite some new features that have also been used for the Blackbox Optimization Benchmarking workshop (BBOB) which was held in Kyoto, Japan during GECCO-2019.

The most important new features are updated, streamlined plots, a Python 3 compatible postprocessing module with a corresponding restructuring of the postprocessing code, the support for zip files in the postprocessing, a simplified example experiment script for beginners and a non-anytime example experiment for benchmarking budget-dependent algorithms, improved coverage of the continuous integration testing via CircleCI and AppVeyor, and finally and most-important from a practical perspective an archive with automatized download from all 200+ algorithm data sets available in the COCO data archive. Of these, 17 algorithm data sets have been made newly available in 2018 with four scientific papers being presented at the BBOB-2019 workshop.

In the background, there have been additional (preparational) activities, in particular due to the two Inria ADT projects “COCOpysuites” and “COCOpost”. The “COCOpysuites” project aimed at a rewriting of the experimental part of COCO in python to allow for an easier development, testing, and implementation of new test suites. The “COCOpost” project aimed at a complete rewrite of the python postprocessing with a focus on new, interactive plots and a clearer structure for improved maintenance. In addition, new test suites have been developed and implemented for large-scale, constrained, multiobjective, and mixed-integer optimization. All those extensions will be made available step-by-step to the scientific community after proper alpha- and beta-testing in the coming planned releases.

Developments within the CMA-ES library

The pycma library has not seen major changes, but overall 39 commits pushed for maintenance, bug-fixes and smaller improvements (roughly 1000 lines of code). An as of yet unpublished development has been the modularization of the data logger. A surrogate fitness model module with 969 lines of code has been developed and is already operative but has also not yet been released.