I thank you very much for your response Rick R.,
So, my assumption of one model is incorrect. I believe in my case it is not (still?) a big problem, because my model (NN) runs in just a few seconds. But still I am wondering if, for a bigger simulations model(CRNN in my case), could be possible to utilize just one model to evaluate on a serial fashion , every individual (just the parameters encoded) of every generation.
However, now I know I need to start thinking on another way
I believe that under this Swarm libraries the method to create the populations should pass the parameters (the weights encoded as binary strings ) to every new created individual (NN). May be there is some one else who has some experience utilizing this libraries??
David Rick Riolo <address@hidden> wrote:
i'm not quite sure what you are asking, but here is
one typical way that people use a GA (more generally, EA)
with a NN:
With an EA you generally want to have a population size P >> 1.
Each individual encodes the information for a set of weights
for one NN. (i'm assuming the NN architecture is fixed.)
You use each individual NN to do whatever its supposed to do,
and rate it, i.e., assign it a fitness.
You do that for all P individuals (NNets).
Then you run a "generation" of the EA, to generate the
new population of individuals (NNets).
Repeat.
There are other similar approaches, of course.
(eg steady state GAs, etc).
I am not familiar the GA or NN libs that come with Swarm,
so i don't know about implementation.
- r
Rick Riolo address@hidden
Center for the Study of Complex Systems (CSCS)
4477
Randall Lab
University of Michigan Ann Arbor MI 48109-1120
Phone: 734 763 3323 Fax: 734 763 9267
http://cscs.umich.edu/~rlr
On Mon, 17 May 2004, DAVID CAMACHO wrote:
> Date: Mon, 17 May 2004 00:37:51 -0700 (PDT)
> From: DAVID CAMACHO
> Reply-To: address@hidden
> To: address@hidden
> Subject: [Swarm-Modelling] GA question
>
>
> Hi everyone, I am developing an agent model based on the CRNN (continuous time recurrent neural networks, implemented by J.J.Merello in Neurolib-2.1) the actual model runs fine. But as mentioned for some authors, the best way to improve their behaviour (changing weights and bias terms) is through a Genetic Algorithm (there are also one GA, Breeder-2.1 developed by JJMerello) but I don´t know how to implement it. I have read the respective documentation, witch is clear, but still some very basic things are not clear to me, because I have no clue on GA.As far as I
understood, the model to be optimised, in this case the NN, should be created by the GA. As this model is fixed, (I don´t want to optimised the architecture, it is full connection, nor the number of neurons) I suppose the GA should create just one copy of the model. Which in turn should be utilized by the fitness function to evaluate the entire population of chromosomes (which represent the weights to be optim!
is!
> ed). If
> this one copy supposition is right, still is not clear to me, if the model should be reinitialised every time for the new set of weights, or a new model as to be created.On the other hand, is all this is right, I suppose I have to run the model in batch mode, but it is the most important part for me, not the GA. So how can I visualize the result?I hope some one can give me some directions,Thanks,David Camacho
>
>
>
> ---------------------------------
> Do you Yahoo!?
> SBC Yahoo! - Internet access at a
great low price.
_______________________________________________
Modelling mailing list
address@hidden
http://www.swarm.org/mailman/listinfo/modelling