Re: Eureqa - a new Error Metric for Binary Outputs

[Michael Schmidt]

How are you modeling now? If you use the step(x) function, you would effectively be optimizing for % correct answers, rather than some error measure. For example:

output = step( f(x) )

Though, I would usually recommend optimizing using logistic first:

output = logistic( f(x) )

Assuming output hold 0 and 1 values.

Michael

On Tue, Apr 16, 2013 at 2:38 AM, Eddy Taylor <ennor...@gmail.com> wrote:

Hi People,
I am using Eureka to predict binary outputs of 1 and 0.
Do you know the best way I can get Eureka (Desktop or API version) to understand I want my model to be wrong < 20% of the time?

The Error Metric would be calculated as such:

PercentWrong_Error = w/(c+w)

Where
-w is the number of times y=f(x,w) is false
-c is the number of times y=f(x,w) is true

--
You received this message because you are subscribed to the Google Groups "Eureqa Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to eureqa-group...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.

[Michael Schmidt]

Eddy, for this expression, it should be fine to use the default error metric setting, mean absolute error. If you ever do more complicated custom error metrics, that might go negative, you can use the "Signed difference" error metric that's an option in the Set Target view.

If you're interested in more of this fancy custom error metrics stuff, check out:

http://blog.eureqa.com/2010/07/custom-error-metrics-and-special-search.html

Michael

On Wed, May 15, 2013 at 2:54 AM, Eddy Taylor <ennor...@gmail.com> wrote:

My reply to Michael -

Ok, I will try both "if( y == f(x,t), 0, 1) = 0" and "if( abs(y - f(x,t)) < 0.1, 0, 1) = 0" in the Target Expression but where do I select the option to maximize or minimize the value of the function?