Re: [jgap-users] Comparing GAs to nature

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

Hello,

you mention a really interesting aspect. In my opinion it is not very
popular which makes it even more amazing in my eyes!
I absolutely agree with you. Nature de facto has a set of fitness functions
which help evolving individuals concerning different aspects.

What can we do with JGAP? Several options come to my mind:

A) provide multiple fitness functions for a single problem
B) provide a pool of shared fitness functions for a set of problems
C) provide several problems for a single fitness function

Aspect C) is not your point, really. But in combination with A) and B) there
may be some opportunities to bring it into play.
I am wondering about a good and simple example to implement in order to show
that your elabored principle can be transferred to the machine. Any ideas?

Sme other thoughts to make it more complex :-)
D) chain problems sequentially --> solution of problem 1 gets input for
problem 2 --> don't know a good example for that yet
E) coevolution: let 2 Gas/GPs compete against each other --> coevolution -->
may not be related to the current issue but maybe it is, anyway

Best

Klaus
 

> -----Original Message-----
> From: puzzled [mailto:sg...@nu...] 
> Sent: Friday, March 02, 2007 8:13 AM
> To: jga...@li...
> Subject: [jgap-users] Comparing GAs to nature
> 
> 
> Hello,
> I have a rather 'philosophical' question, and I'd thank to 
> any GA expert pointing to a related link/paper.
> 
> When designing a GA fitness function (and a gene coding 
> scheme to go along), it is often a concern to provide the 
> means for gradual, partial fitness improvement towards the goal.
> 
> However, it often happens in nature that evolution proceeds 
> 'modularly' - the pieces of a complex molecular machine are 
> not useful until the whole machine is complete, but 
> individual pieces can provide an evolutionary advantage by 
> having *some other* function by themselves.
> 
> Now, this works because nature provide an infinity of 
> "fitness functions", so to speak - opportunities for 
> individuals to have an advantage. But in the contexts of GAs, 
> the fitness function is only one. Partial solutions *must* 
> provide an advantage, to some degree, in terms of the goal 
> (and only) fitness function; they don't have the chance to 
> provide an advantage by having merit according to "some other 
> function" (often a completely unexpected one), as in nature.
> 
> Can anybody think of a way to incorporate this feature of 
> nature into GAs?
> Alternatively, what can be done for the problem of partial 
> fitness improvement?
> 
> Thanks in advance for taking care of my curiosity.
> --
>