[PD] New external on neural networks

[IOhannes m zmoelnig]

Davide Morelli wrote:
> 
> IOhannes:
> the learn_from_file()  function must be called only once, after the net 
> is trained you can save its state sending 2 messages:
> [filename test.net(
> then
> [save(
> the net can be loaded simply using the test.net as argument [nn 
> test.net] or by sending 2 messages:
> [filename test.net(
> and
> [load(

yes, i have seen that as soon as i have compiled and tried your external.
(btw, wouldn't it be nice to also accept [save test.net( instead of 
[filename test.net, save( ??

> 
> IOhannes and Georg:
> I would really like to have on-line training but I think is not 
> supported by FANN right now.. umph.

looking at my fann.h file (from fann-1.2.0) i see a 
fann_train()-function which should do that, no ?
(this will work only for floating-point values though)


> 
> I never created an autoconf myself although I saw many, I will try as 
> soon as I have some spare time. (I must find the time to reinstall linux 
> at home)
> It will be a perfect solution for linux and osx.  ...but what about win32?

cygwin will know how to handle it.
a plain msvc-environment is different of course, but i guess one could 
live with that:

just an idea:
+ provide a "Makefile" for msvc (which people would have to edit to 
match their system)
+ provide a configure script that overwrites the "Makefile" on systems 
that support configure

or:
  use "Makefile.nt"


this reminds me, that i should buy a written documentation (aka: 
handbook) for make.



mfg.a.dr
IOhannes