The idea of --gen-best-candidate is to generate head-on a candidate that perfectly fits the data, with the obvious risk of completely overfitting it. This is useful however for experimenting and testing moses when something might be going wrong, such as moses performing poorly in-sample, etc. Additionally it might speed up learning in special circumstances, such as when the sample set is extremely large relative to the feature set, etc.
The idea of
--gen-best-candidate
is to generate head-on a candidate that perfectly fits the data, with the obvious risk of completely overfitting it. This is useful however for experimenting and testing moses when something might be going wrong, such as moses performing poorly in-sample, etc. Additionally it might speed up learning in special circumstances, such as when the sample set is extremely large relative to the feature set, etc.