Meta-Genetic Programming: Co-evolving the Operators of Variation - Bruce Edmonds
This is not the first time that co-evolving the genetic operators has been proposed. Teller [14] suggests the co-evolution of programs to perform operations of variation on other programs. In that example both the base and operator programs were networks and part of the rationale for using evolved operators is that there is no obvious operator that would correspond to a GP "crossover" operator. MGP is conceptually much simpler and hence may enable more opportunities for meaningful analysis. Furthermore because it is uses a straightforward GP algorithm on the operator population, some of the analysis and heuristics learnt about using GP algorithms can be brought to bear. Also comparisons with other techniques are easier.
Peter Angeline [2] investigated the possibility of a "self-adaptive" crossover operator. In this the basic operator action is fixed (as a crossover) but probabilistic guidance is used to help the operator choose the crossover nodes so that the operation is more productive.
MGP can be compared to approaches in GAs, such as [12] in which augment a standard GA with a series of crossover masks each of which has a weight associated, which changes according to their success, and which affects their later utilisation. In MGP, however, there are a potentially infinite number of such operators
Generated with CERN WebMaker