Penn ML Benchmark Problem Instances

[bburlacu]

The EpistasisLab at the University of Pennsylvania maintains a curated set of benchmark datasets for classification and regression [1].

A large part of the regression datasets have already been used for an in-depth comparison of symbolic regression methods [2].

This ticket will track integration of these benchmark problems in HeuristicLab via a new problem instance provider.

[1] ​https://github.com/EpistasisLab/penn-ml-benchmarks

[2] ​https://arxiv.org/abs/1804.09331

[bburlacu]

r15951: Add PennML problems and implement instance provider. A reusable descriptor was also implemented, taking advantage of the structure of the data (by convention, target is named "target" and is always the last column). The instance provider generates the descriptors on the fly from the archive.

[bburlacu]

r15952: Fix namespace.

[gkronber]

I made a quick test using our GBT implementation with standard settings and 3000 iterations.

See the two attachments.

[bburlacu]

Random Forest results

[gkronber]

Reviewed r15951 and r15952. Thanks!

[gkronber]

r16163: merged r15951, r15952 from trunk to stable

[gkronber]

r15951 depends on r15855 (#2908)