Exploring the MLDA benchmark on the nevergrad platform

Jérémy Rapin, Marcus Gallagher, Pascal Kerschke, Mike Preuss, Olivier Teytaud

Abstract

This work presents the integration of the recently released benchmark suite MLDA into Nevergrad, a likewise recently released platform for derivative-free optimization. Benchmarking evolutionary and other optimization methods on this collection enables us to learn how algorithms deal with problems that are often treated by means of standard methods like clustering or gradient descent. As available computation power nowadays allows for running much 'slower' methods without noticing a performance difference it is an open question which of these standard methods may be replaced by derivative-free and (in terms of quality) better performing optimization algorithms. Additionally, most MLDA problems are suitable for landscape analysis and other means of understanding problem difficulty or algorithm behavior, due to their tangible nature. We present the open-source reimplementation of MLDA inside the Nevergrad platform and further discuss some first findings, which result from exploratory experiments with this platform. These include superior performance of advanced quasi-random sequences in some highly multimodal cases (even in non-parallel optimization), great performance of CMA for the perceptron and the Sammon tasks, success of DE on clustering problems, and straight forward implementations of highly competitive algorithm selection models by means of competence maps.
BibTeX
@inproceedings{rapin2019exploring,
  author = {Jérémy Rapin and Marcus Gallagher and Pascal Kerschke and Mike Preuss and Olivier Teytaud},
  title = {{Exploring the MLDA benchmark on the nevergrad platform}},
  booktitle = {GECCO (Companion)},
  year = {2019},
  doi = {10.1145/3319619.3326830},
}

Loading related content...

Figure