MLP - what with enumeration inputs

[bim-bom]

I have an MLP for classification tasks.
When Input values are in some range - it is easy - I just normalize this
range to <-1 ; 1>
But what when some input is an enumeration (e.g. one of 10 colors)? I
can assign some number in the range <-1 ; 1> to each color, but I guess
that won't be able to train this network really well.
Are there any solutions for this problem?

[Greg Heath]

You could use categorical one-of-10 input coding.

However, I would favor using the corresponding
ordinal wavelengths.

Hope this helps.

Greg

[Phil Sherrod]

The usual approach for handling categorical input variables is to convert
the variable in to a set of binary (0/1) variables with the one
corresponding to the presented category having the value 1 and all others
being 0. Actually, you can use (n-1) binary variables with all of them
being zero representing the n'th category.

--
Phil Sherrod
http://www.dtreg.com -- Neural networks, SVM, Decision trees

[Phil Sherrod]

On 17-Dec-2009, Greg Heath <he...@alumni.brown.edu> wrote:

> You could use categorical one-of-10 input coding.
>
> However, I would favor using the corresponding
> ordinal wavelengths.

Well, if the colors have a spectral ordering, that's a good idea. But if
they are just color codings for arbitrary categories, then you're going to
have to generate a set of category variables.