🚀 go-pugleaf

On Mon, 28 Nov 2005 22:48:53 -0000, Oliver Wong <owong@castortech.com>  
wrote:
>     Really? I thought a typical implementation would try to preserve the
> qualities of the most-fit strategy found. One way was the mutate the good
> strategies less (that's the one I mentioned above), and the other was to
> generate more "offsprings" from the good strategies.

Sorry, I misunderstood you (I was considering only mutation and not  
reproduction by cross-over) and was wrong to state that mutation occurs  
uniformly across a population.  What I meant was that *after* you have  
selected some candidates as the basis of your next generation, they each  
have an equal chance of being mutated.  This is true, for truncation  
selection at least, but doesn't consider that with some selection  
strategies, such as fitness-proportionate selection, a candidate may be  
selected more than once and therefore has more than one equal chance of  
being mutated.

I thought that what you were suggesting was that the mutation probability  
should be adjusted proportionate to the fitness.  You were right, of  
course, about selecting the fittest candidates for reproduction.  Without  
discriminating against weak candidates there is no "natural" selection and  
therefore no pressure for the population to improve.  However, you have to  
be careful not to completely discard weak candidates as they may have  
aspects that don't score highly by themselves but prove useful when  
combined with other material from other candidates.

Dan.

-- 
Daniel Dyer
http://www.dandyer.co.uk