Support more than 256 variables in linear regression models

[swinkler]

Linear regression crashes if more than 256 input features are to be used. (The number of subtrees of the tree representing the linear model is too big.)

[gkronber]

I'm tempted to reject this ticket, as it is unlikely that this would produce an useful solution. If we allow more than 256 variables I'd prefer that we produce a different kind of model (no symbolic expression tree) in such cases.

For me the actual issue is that we do not have standard feature selection methods for LR or regularized linear models.

More discussion is needed before we decide on the further steps.

[gkronber]

Ticket on LR with feature selection: #745.

[mkommend]

Test Script

[mkommend]

r15766: Adapted symbolic expression tree compilers to allow more than 256 child nodes.

The issue is not that the used LR (ALGLIB) does not allow more than 256 variables, but that our symbolic models only allow 256 child nodes in an expression tree. LR models are transformed into an symbolic model by adding all features as variables below an addition and hence this error occurred.

[mkommend]

Please review the changes in r15766 carefully as this affect many data analysis parts of HeuristicLab.

Furthermore, I have no idea how this adaptations affect the performance of the interpreters, but I suspect / hope not much.

[gkronber]

The test run times on the builder do not show a direct effect of r15766.

[gkronber]

Reviewed r15766. Looks good.

[gkronber]

r15836: merged r15766 from trunk to stable.