Evaluator for symbolic data analysis that calculates the mean of logarithmic residuals

[gkronber]

In preliminary experiments, we found that this improves success rates for data sets with a few very large/small target values (e.g. Korns benchmark instances containing inverses of near zero values). In these cases the squared error or absolute error puts too much emphasis on modeling the outlier values. In such instances small variations are not stemming from noise but are actually the signal that should be modeled and a log-error evaluator would put more weight on such small values.

[gkronber]

Created a feature development branch with r10202

[gkronber]

r10203: added log residual evaluator for symbolic regression

[mkommend]

Despite the source code looking fine, I am unsure if this functionality should be included in trunk right now. First of all the logarithm is not defined for 0, which is not handled in the evaluator (skipping invalid values?). Second, the current implementation calculates avg(log(abs(y'-y)) which tends to -Inf for very small values (this may be an issue if one data sample was hit with a very high precision). Third, there is no warning / indication that despite linear scaling is requested, it is not performed.

[gkronber]

r10305: changed log residual evaluator to calculate avg(log( 1 + abs(y - pred)))

[mkommend]

Please put a comment / description that linear scaling is never performed for this evaluator. Otherwise everything looks fine.

[gkronber]

r10482: added a notice that the log residual evaluator does not apply linear scaling

[gkronber]

r10483: merged trunk changes to prepare for reintegration

[gkronber]

r10484: reintegrated branch for log residual evaluator

[gkronber]

r10485: deleted branch for log residual evaluator

[gkronber]

Please review r10482 and r10484.

[gkronber]

r10486: fixed build fail

[gkronber]

Depends on #2093

[mkommend]

r10509: Merged r10483 and r10486 into stable.