| |
 |
 |
|
Last update: February 19, 2014
|
|
|
 |
 |
|
|
Rounding Thresholds
The sophisticated classification algorithms of GeneXproTools allow the combination
of powerful fitness functions with a wide range of evolvable rounding thresholds.
The evolvable rounding thresholds used in classification are:
For classification,
GeneXproTools also supports two kinds of fixed thresholds:
In addition, certain fitness functions, namely
Maximum Likelihood,
Hinge Loss,
and Logistic RMSE, require a fixed zero threshold.
The evolvable thresholds are evaluated for each individual model during evolution,
which means that they are evaluated for the actual training data that was being used
at the time the model was created and, therefore, it could have been the
entire training dataset or a sub-sample of the training dataset. Either way,
the evolvable threshold becomes an intrinsic
model parameter and
can only be changed by evolution (for example, if a model is used as seed) or
manually through the Update Threshold functionality of GeneXproTools.
Updating the rounding threshold of a model might be useful either because the
training data changed, or the model structure changed or the threshold type
changed. Moreover, by changing the rounding threshold of a model or an
ensemble of models, you are creating slightly different models that can be used
to create good ensembles.
Logistic regression also uses an evolvable rounding threshold, the
logistic rounding
threshold, which is also fine-tuned by evolution in order to improve
the quality of the evolving models. Like the evolvable thresholds of classification,
the logistic threshold can also be updated using the
Update Threshold functionality of GeneXproTools.
See Also:
Related Tutorials:
Related Videos:
|
|
|
|
|
