'GENOTYPES' FOR NEURAL NETWORKS

Genetic algorithms applied to populations of neural networks can  be viewed 
as an alternative methods for obtaining desired performances from  networks 
with  respect  to  more  traditional  learning  methods.  Or  they  can  be 
interpreted as models of processes of neural evolution occurring in nature. 
If one  takes this second stance, aspects of evolution which are ignored in 
most applications  of  genetic algorithms  can  become important objects of 
study. One such  aspect  relates  to  the  distinction between genotype and 
phenotype. In  almost  all  applications  of  genetic  algorithms to neural 
networks  what  is  genetically inherited  is the phenotypic network itself 
whereas  in  nature genetic information specifies a set of instructions for 
constructing  the  phenotypic network,  with  the  mapping from genotype to 
phenotype very  indirect and nonlinear. This chapter describes a scheme for 
encoding 'genotypes'  for  neural networks which map in complex ways to the 
corresponding   phenotypic   network.   An   important    aspect   of   the 
genotype/phenotype   mapping   is   its   temporal   character.  Biological 
development (or maturation)  is  the sequence of different phenotypic forms 
which  are  sequentially  generated by the genotype in interaction with the 
environment.  While  most   simulations   with  complex  genotype/phenotype 
mappings ignore  this aspect and generate a single mature network at birth, 
the chapter  describes simulations with networks which change (grow) during 
the  life  of  the organism. The distinction between genotype and phenotype 
and  the  realization  of  the  mapping as a temporally distributed process 
make  it  possible  to  examine  important   questions   of  evolution  and 
development and of their interaction.