Re: [jgap-users] local parallelization

[Klaus M. <jg...@kl...>]

Hello Andreas,
 
thanx a lot for pointing out your issue in detail.
 
Regarding fitness function: I cannot say without knowing more details. Is a
source code (maybe incomplete or in progress) for your implementation
available?
 
The understanding in JGAP is as follows:
+ The generation number is not copied as a new population is seen to have
generation 0
+ The population size for a newly created population (resulting from a split
of a huger population) is kept the same as for the original, as each part
(split) is seen to be processed independently and in the same way as the
original population.
 
However, these are arbitrary considerations, it could maybe be better to do
differently. If you could create an example that shows which way is
producing better results (your intention or the current JGAP
implementation), it would help in deciding if changing JGAPs behaviour does
make sense.
 
I have not played through the process in the past few weeks, so it may be
that I have overseen something. If so, please drop me a note.
 
Best

Klaus
www.klaus-meffert.com <http://www.klaus-meffert.com/> 

 



  _____  

From: Andreas Gantner [mailto:ga...@fl...] 
Sent: Monday, November 21, 2011 10:42 AM
To: jga...@li...
Subject: [jgap-users] local parallelization



Hello jgap-team,

 

I´m currently trying to do a CFD-based genetic optimization process with
jgap which should be locally parallelized (on one PC).

So I took as example the SimpleJobConsumer. 

As fitnessfunction I defined an extended BulkFitnessFunction because the
evaluation method needs further information which is stored in my Population
class (the working directory where the data for the cfd simulation can be
found and should be accomplished)

In the main iteration I´m splitting the population and evolve those splitted
populations in single threads (see  SimpleJobConsumer).

Afterwards the results are merged into the Population of the genotype
instance. 

I found out the Configuration clone Method used in
Genotype.getEvolves(IPopulationSplitter a_splitter) doesn´t copy the current
generation number. Additionally the Configuration clone also uses the
complete population size, and not the splitted one (popSizeSplit=
popSizeComplete/numSplitter).

So I don´t know if there might be a problem in the GABreeder? Does this
behavior affect my optimization process at all. Might that be a bug? 

Or is my approach principally wrong? Should I use a “classic”
FitnessFunction instead of a BulkFitnessfunction?

 

Thank you for your response!

 

Andreas Gantner

 

PS: I also don´t understand the original merge-method in the class Genotype.
The merger passed isn´t used at all and only the best three individuals are
chosen. Is there also a bug or why didn´t you use it? For the present I
created my own method.