Print

Print


Thanks very much for all answers!! Sean's offer is super!!
I usually bring best individual(result from ECJ(GP)) to Matlab and test it 
on different data sets and produce different graphs needed. 

Usually, I replicate the fitness for this best individual and then test it 
on different data set. As I feed different individual with different data, 
say each one get 200 of total 20000. I need to find out exactly which data 
set has been used by this best individual. 

However when I tried to do the above steps the last time. I found some 
problem. I could not replicate the raw fitness!!!!It looks like either 
there is some problem when I transfer Best ind. from ECJ to Matlab or 
something happened during ECJ training. It is like a mystery to me. 

I went back to the output from ECJ incuding all generations. I tested 
couple of individuals in that output. Some of the individual could get the 
fitness replicated. some not. This is the first time this strange thing 
happen. It never happened before.

Would Sean guess some possibilities?
After Sean gives that procedure which can evaluate best individual off 
line. I could test that individual, see, if I could replicate fitness, 
therefore to find out where the mistake come from. 

Looking forward to your answer.
Zheng



On Fri, 4 Sep 2009 17:38:36 +0200, Sean Luke <[log in to unmask]> wrote:

>Zheng, the standard thing I'd say would be to reduce the population
>size to 1, create a file with the best individual and load it at
>runtime into the population, and use evaluate() to test the
>individual.  But that's pretty tough for a Java newbie to grok.  What
>I ought to do is think about creating a simple test harness that makes
>this procedure easy for you -- it's a common need.  I could do it in a
>few days maybe (no promises!).  Let me mull it over.
>
>Sean
>
>On Sep 4, 2009, at 9:19 AM, Claes Gyllenswärd wrote:
>
>> 2009/9/3 Zheng Yin <[log in to unmask]>
>> Hi, Sean,
>> I have a problem, which is not really ECJ question. I am using ECJ
>> to do
>> the symbolic regression. I would like to test the best individual
>> (around
>> 400 nodes) at out of sample off line. Is there any code in ECJ I could
>> use?  I am a Java beginner.
>>
>> many thanks
>> Zheng
>>
>> While it's very possible that Sean knows a lot better than me, I
>> don't understand your question at all, so perhaps you will want to
>> clarify.
>> More specifically, you want to do something offline, but what is
>> online in the first place?