Sean Luke  



Please remove me from email as I am not using your system.  

[log in to unmask]

---------- Original message ----------------------

From: Sean Luke <[log in to unmask]>

> On Jul 27, 2007, at 2:28 PM, Peter Walker wrote:

> 

> > 1)       (Traditional mutation): select a random node and remove it  

> > (along with all its children) and replace it by a randomly  

> > generated subtree and subtree root.

> 

> This is ec.gp.koza.Mutation

> 

> > The other two schemes I would like to implement are based on “point- 

> > mutation”. Both point-mutation schemes would start out by selecting  

> > a node and generate a random node (e.g. symbol1) to replace it.  

> > Symbol1 would be checked to see whether it is of the same arity as  

> > the node. If it is, then the node would be replaced by symbol1. If  

> > they are not of the same arity, then we would either:

> >

> > 2)      Pick another symbol until we find one of the same arity; OR

> >

> > 3)      Remove the node and replace it with a new subtree using  

> > symbol1 as the root.

> 

> You'll need to make a custom mutator for this that FIRST picks a  

> replacement and THEN decides on what to do (#2 or #3).  However:

> 

> - ec.gp.breed.MutateOneNodePipeline will pick another node guaranteed  

> to be of the same arity (and type equivalence) and replace the node  

> with the new node.  This is roughly equivalent to #2, as you've noted.

> 

> - #1 and #3 are roughly equivalent, it'd seem.

> 

> So I think you'll be writing a custom mutator -- take a look at the  

> code of these classes to get an idea about how to go about doing this  

> (it's nontrivial but doable!)

> 

> Sean