Scaling graphs over all functions

[brockho]

When looking at the new bbob-largescale results, we realized that it will be good to have a plot that can display the scaling behavior of multiple algorithms with dimension over a set of functions.

So far, we have scaling graphs for single functions, that display the aRT for given targets (or a single target) versus dimension.

The new suggestion, to come up with a similar plot but with data aggregated over several functions, is the following:

• start from the aggregated ECDFs (of aRT values) for the given functions of interest in the given dimension(s)
• for each algorithm, we extract some quantiles of the aRT from the ECDF, for example the 20%-, 50%, and 80%-quantiles
• then we simply collect these quantiles for all dimensions and plot those quantiles over dimension in a single graph for all algorithms

[nikohansen]

The idea is to have Box-Whisker plots vs dimension, where the statistics are taken from the ECDFs. One challenge is that not all quantiles need to be finite/present, that is the display must be able to deal with such infinite/missing quantiles.

The underlying ECDF data do not need to be aRT, they could be all the single runs or they could be from simulated restarts. We could use the same generated data that are already used in the ECDF plots.