The bbob-largescale Test Suite

The bbob-largescale test suite contains 24 single-objective functions in continuous domain, extending the well-known single-objective noiseless bbob test suite to dimensions 20 to 640. The core idea is to make bbob’s rotational transformations in search space computationally cheaper in higher dimension while retaining some desired properties by employing permuted block-diagonal matrices.

You may cite this work in a scientific context using

@article{bboblaregscale2020,
    author = {Konstantinos Varelas and Ouassim Ait El Hara and, Dimo Brockhoff and Nikolaus Hansen and Duc Manh Nguyen and Tea Tu{\v{s}}ar and Anne Auger},
    title = {Benchmarking large-scale continuous optimizers: The `bbob-largescale` testbed, a COCO software guide and beyond},
    year = {2020},
    journal={Applied Soft Computing},
    volume={97},
    pages={106737},
    year={2020},
    publisher={Elsevier}
    url = {https://www.sciencedirect.com/science/article/pii/S156849462030675X}
}

Quick Access

• function definitions
• some available results
• data archive

The Basic Idea

The bbob-largescale test suite provides 24 functions in six dimensions (20, 40, 80, 160, 320 and 640) within the COCO framework. It is derived from the existing single-objective, unconstrained bbob test suite with modifications that allow the user to benchmark algorithms on high dimensional problems efficiently.

The main difference between the bbob-largescale and the original bbob test suite is that the full matrices R in the rotational transformations x \rightarrow z = Rx of bbob are replaced by block-diagonal rotation matrices

R = P_{\text{left}}BP_{\text{right}}

with P_{\text{left}} and P_{\text{right}} being two permutation matrices and B being a block-diagonal matrix of the form:

B = \left( \begin{array}{cccc} B_1 & 0 & \ldots & 0\\ 0 & B_2 & \ldots & 0\\ 0 & 0 & \ddots & 0\\ 0 & 0 & 0 & B_{n_b} \end{array} \right)

where n_b is the number of blocks and B_i, 1 \leq i \leq n_b, are square matrices of sizes s_i \times s_i satisfying s_i \geq 1 and \sum_{i=1}^{n_b} s_i = n.

For more details, see the concrete function definitions.

Compatibility

The bbob-largescale suite is compatible with the bbob suite in the sense that in dimensions 20 and 40, both test suites contain the same functions (with full rotation matrices) and that algorithm data sets from both suites can be postprocessed (and viewed) together.