In A. Abraham, B. de Baets, M. Köppen, and B. Nickolay, eds., Applied Soft Computing Technologies: The Challenge of Complexity, pages 517-536, Springer-Verlag, 2006.
Designing Neural Networks Using Gene Expression Programming
Special genetic operators
The evolution of such complex entities composed of different domains and different alphabets requires a special set of genetic operators so that each domain remains intact. The operators of the basic gene expression algorithm
(Ferreira 2001) are easily transposed to neural-net encoding chromosomes, and all of them can be used as long as the boundaries of each domain are maintained and alphabets are not mixed up. Mutation was extended to all the domains and continues to be the most important genetic operator. IS and RIS transposition were also implemented in GEP-nets and their action is obviously restricted to heads and tails. However, a special insertion operator was created that operates within Dw and Dt,
ensuring the efficient circulation of weights and thresholds in the
population. Another special operator, weights and thresholds’ mutation, was also created
in order to directly introduce variation in the set of available weights and thresholds (see
section 3.3).
The extension of recombination and gene transposition to GEP-nets is straightforward, as their actions never result in mixed domains or alphabets. However, for them to work efficiently (i.e., allow an
efficient learning), we must be careful in determining which weights and/or thresholds go to which region after the splitting of the chromosomes, otherwise the system is incapable of evolving efficiently. In the case of gene recombination and gene transposition, keeping track of a gene’s weights and thresholds is no difficult task, and these operators work very
well in GEP-nets. But in one-point and two-point recombination where
chromosomes can be split anywhere, it is impossible to keep track of the weights and thresholds. In fact, if applied straightforwardly, these operators
would produce such evolutionary monsters that they would be of little use in multigenic chromosomes. Therefore, for multigenic systems, a special intragenic
two-point recombination was created so that the recombination is
restricted to a particular gene.