Very good idea. I think "cuda" has a java plugin
For the upcoming 1 teraflop tesla card. Have run up to 6 of the 500 teraflop
cards at the same time
Tim

-----Original Message-----
From: ECJ Evolutionary Computation Toolkit
[mailto:[log in to unmask]] On Behalf Of Denis Robilliard
Sent: Monday, September 29, 2008 3:06 AM
To: [log in to unmask]
Subject: Re: runtime complilation for expensive eval ?

You could also port evaluation on an Nvidia G8xxx graphics card (high 
end cards are 10x more powerful  than a standard CPU): ECJ-compatible 
sample code is available at
http://www-lil.univ-littoral.fr/~robillia/GPUregression.html

see also:
 www.gpgpu.com

Denis

Ondrej Pacovsky a écrit :
> David R White wrote:
>> Hi Ondrej,
>>
>> In the past I've written individuals out to a C source file, compiled 
>> them using GCC and run them through a processor simulator.  It is 
>> fairly straightforward to implement, I use makeCTree and override 
>> this method where necessary.
> This is quite interesting, is the source available somewhere ?
>> It is quite slow, but there are plenty of ways to speed up the 
>> process of compiling and evaluating, depending on the nature of the 
>> problem you're trying to solve.  I've also recently started using 
>> Master-Slave evaluation to run my experiments across machines, which 
>> is a very scalable way of dealing with this situation.  Whether it 
>> will be fast enough for your work depends on the problem you're 
>> solving (e.g. can you compile the individual and evaluate it in a 
>> single execution? or would you have to re-compile or execute multiple 
>> times?) and the number of generations and population size.
> I need to re-compile each individual each time it changes (usually 
> every generation). But my fitness eval means calling >10^3 times the 
> same individual, so the cost of compilation could be overcome if the 
> single tree eval call is significantly faster after compilation.
>
> Ondrej
>
>>
>> Cheers
>>
>> David
>>
>> Ondrej Pacovsky wrote:
>>> Hi,
>>>
>>> I was wondering whether someone tried converting the GP individual to
>>> java (or other) code (perhaps by the ECJ to Java converter) and
>>> compiling it before actually running the evaluation. This is of course
>>> quite slow, but for symbolic regression on many training values 
>>> could be
>>> interesting. Thinking of 10^3 and more evals per individual per 
>>> generation.
>>>
>>> -- Ondrej
>>
>

-- 
Denis Robilliard
L.I.L.
Université du Littoral
50 rue F. Buisson
62100 Calais
03.21.45.56.67