Some Regression Problem Instances are not correct

[sforsten]

Some of the regression problem instances don't work at all (Korn 7, 9, 15), while others seem to have miss some points in the training or test partition and Keijzer 1-3 have the same name.

[sforsten]

r8900:

• renamed Keijzer 1-3
• corrected training and test partition of Keijzer 9 and 10
• changed interval for some variables in Korn 7, 9, 15 to avoid infinity and NaN values

[mkommend]

Please add the following problem to the real world problem instances ​http://symbolicregression.com/?q=emulatorProblem.

[gkronber]

Discussion is needed. Please do not yet add the emulatorProblem...

[gkronber]

r8999: adapted Vladislavleva regression problem instances to match the definition of the "GP Benchmarks" paper

[mkommend]

I had a quick look over r8999 and I would bet that the range of the Kotanchek and SineCosine problem are not correct, because I could not imagine that a test partition of ~ 100,000 samples was intended.

[gkronber]

Replying to mkommend:

I had a quick look over r8999 and I would bet that the range of the Kotanchek and SineCosine problem are not correct, because I could not imagine that a test partition of ~ 100,000 samples was intended.

Well it is defined in the GP benchmarks paper in this way and implemented in ECJ like that. So it seems this is intentional. Additionally, the large test partition should not be a problem as it is only needed once for test performance evaluation.

The only other problem with a large test set is Keijzer-15 (441*441 samples. All other instances have smaller test sets.

[gkronber]

Review comments about r8900: Korns 7, 9, and 15 do work but the range of the target variable is very large due to exp(50) occurring in the target expression. Since this is how the problem is defined we should keep it that way for now. I'm awaiting further input from the gp-benchmarks mailing list.

[gkronber]

r9007 cross-checked all regression problem instances with the GP benchmarks paper and adapted where I thought necessary.

[gkronber]

r9008: fixed ranges of Vladislavleva-5

[gkronber]

r9013: fixed range for Vladislavleva-1 as suggested on the GP-benchmarks mailing list.

[gkronber]

r9091: added a new artificial benchmark problem for regression specifically for testing feature selection algorithms

... and reverted in r9092 (will be tracked in: #1999)

[gkronber]

Please review r9007:9008 and r9013 (r9091 has been reverted) I'm still waiting for the original authors answer on the Vladislavleva-5 ranges, but as this is only a notational difference (the data being equal) I think we can leave it this way.