🚀 go-pugleaf

recent work by N. Radcliffe, "Genetic Set Recombination and Its
Application to Neural Network Topology Optimization" in Neural
Computing and Apllication, 1993 1:67-90, suggests that desirable
properties of GAs include "proper assortment" and "respect for
formae membership".  both these definitions build upon Radcliffe's
extension of Holland's schema analysis, called formae analysis,
and require, respectively, that the offspring of two genomes in
a GA have properties which in some way combine the desirable
properties of the parents.

this is nothing new, perhaps.  but what is interesting is his
arrival at the Random, Respectful Recombination operation which
involves selecting randomly and uniformly from the so-called
similarity set of the two parents to produce the offspring.  this
operation bears a striking resemblance to the Gaussian modification
scheme proposed by Fogel and Porto starting in the mid-60s and
elucidated in Artificial Intelligence Through Simulated Evolution,
an out-of-print and rare book, as well as in recent work (see
Fogel, D. and Atmar, J.W. "Comparing Genetic Operators with
Gaussian Mutations in Simulated Evolutionary Processes Using 
Linear Systems,"  Biol. Cybernetics, 63, 111-114, 1990.)  In 
their work, the value of standard GA mutation operators has been
repeatedly called into question, both in terms of its robust
ability to find globally optimal solutions, and in its speed
of convergence.

the fundamental criticism levied by the so-called evolutionary
simulation-people against GAs as a methodology is precisely
the question of the accuracy of the simulation.  the assumption
that representations which code individual properties in a
linear genome and operations on those representations can
approach the kind of robust optimization seen in nature is
at question.  specifically, they have suggested that complex
behavior in natural systems is almost always the result of 
complex epistatic and pleiotrophic interactions of suites of
genes, rarely of individual genes in isolation.  GAs lack
the ability to effectively simulate this property, and 
ultimately require random mutation to avoid suboptimal
convergence and complex considerations of representation to
provide comparable convergence rates.

any comments???

mark

--
                          WORLD                                    WORLD VIEW
        
      ** *   **     **  ***  ** **** **** 
       **** * *    ** ***************** * 
       *****      *  * ** ************ *                               
        *               ************** *
       ****       ****** ** ** ******      
       *****     *******     *   ***     
        **          ***           *   **
        *           ***            ***  
        *            *            *****  *
                                   * *   *

                                          - mark -