I recommend reviewing the literature for niching methods, and fitness  
sharing.  Fitness sharing in particular has been useful to me in some  
previous experiments in which I required the GA to explore as many  
diverse high-fitness genotypes as possible within a single population.

The fitness sharing I've implemented in the past is described in  
Goldberg's book "Genetic Algorithms in Search, Optimization and  
Machine Learning".  The basic idea is you have a normal unshared  
fitness value for every individual.  For each individual, you quantify  
its genotype's similarity to every individual in the population, and  
sum these quantities.  Divide the individual's unshared fitness by the  
sum and you have a shared fitness.  This pressures the population to  
avoid clustering.


On Apr 10, 2009, at 12:23 PM, Peter Napp wrote:

> Hi,
> I am not very experienced with evolutionary computation and I am  
> wondering if I should worry about the behaviour of my setup. From  
> one generation to the next generation I get more and more identical  
> individuums. This happens very rapidly. My Popilation size is 3000.  
> For example some individdums of generation 1 appears ten times in  
> generation 2. At generation 10 I have some individuums wich appear  
> over 50 times. At generation 20 the whole population consists  
> basically of only one indiviuum. The diversity of generation 0 looks  
> ok. So I have enough structure to get enough different individuums,  
> but something happens during the breeding. I am using pretty much a  
> standard setup (all the "simple" stuff). I am little bit overwhelmed  
> by all the parameters, so maybe someone can give me an idea, if and  
> with which parameters I should play to change the described behaviour.
> Any hints are appreciated, thank you.
> Pete