Eureqa II: Possible bug in the calculation of R^2.

[electrolito]

My dear

I obtain R values ​​of about - 9.034 using Eureqa 2 (large and
negative value). However, the R value should always be between 0 and
1. Probably this is due to an bug in the calculation of R^2 in Eureqa
2.
You can explain this please?

I obtain R values ​​between 0 and 1, using the same data in Eureqa 1.


Thanks
Regards
Carlos

[Michael Schmidt]

In the general nonlinear case, it's common for R^2 to be negative since a nonlinear model may not be able to tune its scale and offset. The definition used is:

R^2 = 1 - SS_error/SS_total

Wikipedia discusses negative R^2 briefly in the summary:
http://en.wikipedia.org/wiki/Coefficient_of_determination

You might also see negative R^2 because Eureqa/Formulize optimizes absolute error by default rather than squared error. If R^2 is the statistic you are most interested in, definitely switch to the "Mean squared error" option in the "Set Target" view.

The Results view also displays the correlation coefficient which you might find more intuitive. This ignores scale and offset and will always range -1 to 1. Squaring this value will produce an alternative calculation of R^2.

Michael

--
Eureqa Formulize ( http://www.nutonian.com )
-------------------------------------------------
Unsubscribe: eureqa-group...@googlegroups.com
View Group: http://groups.google.com/group/eureqa-group

[electrolito]

Michael

In Eureqa 1 you calculate R^2 by squaring the correlation
coefficient??

Regards
Carlos

On 18 feb, 14:53, Michael Schmidt <michael.douglas.schm...@gmail.com>
wrote:

> In the general nonlinear case, it's common for R^2 to be negative since a
> nonlinear model may not be able to tune its scale and offset. The
> definition used is:
>
> R^2 = 1 - SS_error/SS_total
>
> Wikipedia discusses negative R^2 briefly in the summary:http://en.wikipedia.org/wiki/Coefficient_of_determination
>
> You might also see negative R^2 because Eureqa/Formulize optimizes absolute
> error by default rather than squared error. If R^2 is the statistic you are
> most interested in, definitely switch to the "Mean squared error" option in
> the "Set Target" view.
>
> The Results view also displays the correlation coefficient which you might
> find more intuitive. This ignores scale and offset and will always range -1
> to 1. Squaring this value will produce an alternative calculation of R^2.
>
> Michael
>

[Michael Schmidt]

No, it's the same calculation as Eureqa I.

--
Eureqa Formulize ( http://www.nutonian.com )