[astro] Re: update about the conjugate gradient descent

[Travis Desell]

I did some more testing today, and for 22 parameters, CGD is taking 
about 19-56 parameters (depending on
how fine grained i specify),  and GD on the other hand is taking over 
200.  I'm using the Polak-Ribière method
of calculating the gradient.  I think this seems to be about right.  
I'll bring in some graphs monday about
how fast each is converging.

On another note, Nate, I was wondering if you were having problems with 
your changes to the astronomy MPI
code hanging.  When i run it, it gets about 25-50% done with a 
calculation of the integral then freezes.  I'm not quite
sure whats going on and it wasn't doing this before...

--Travis


Malik Magdon-Ismail wrote:
> Hi,
>
> Make sure that the reset is >26.
>
> -Malik
>
> --------------------------------------------------------------------------
> Have Faith in Probability
> -Malik Magdon-Ismail
>   Associate Professor of Computer Science
>   Rensselaer Polytechnic Institute (RPI)
>
>   Office: Rm 312 Lally           mail:   Rm 207, Lally Building,
>   email:  magdon at cs.rpi.edu              CS Department, RPI,
>   tel:    518-276-4857                   110 8th Street,
>   fax:    518-276-4033                   Troy, NY 12180-3590.
>   http://www.cs.rpi.edu/~magdon
> --------------------------------------------------------------------------
>
> On Thu, 14 Jun 2007, Travis Desell wrote:
>
>   
>> I've implemented a sum of squares function for the MPI version of our
>> MLE code, and am running
>> it with conjugate gradient descent.  For 22 parameters, it's taking
>> around 56 iterations to optimize -
>> so I think that means we might be having a problem with it.  That could
>> explain why it's not converging
>> faster than normal gradient descent.
>>
>> -- 
>> ----------------------------------------------------------------------------------------------------------
>> Travis Desell
>> <deselt @ cs.rpi.edu>
>> 1-518-878-7424
>> Worldwide Computing Laboratory ( http://wcl.cs.rpi.edu/ )
>> Rensselaer Polytechnic Institute, 110 8th Street, Troy NY 12180, USA
>> ----------------------------------------------------------------------------------------------------------
>>
>>
>>
>>     
>
>
>   


-- 
----------------------------------------------------------------------------------------------------------
Travis Desell
<deselt @ cs.rpi.edu>
1-518-878-7424
Worldwide Computing Laboratory ( http://wcl.cs.rpi.edu/ )
Rensselaer Polytechnic Institute, 110 8th Street, Troy NY 12180, USA
----------------------------------------------------------------------------------------------------------