Putting the user in the loop: on-line user adaption of genetic algorithms

The practical and theoretical success of any evolutionary computation (EC) application depends on the selection of an appropriate combination of representation, search operators and parameters. How these are actually settled upon remains one of the more troublesome aspects of EC. Typically, they are the end result of a tedious meta-search by the user. We consider the issue in terms of the genetic algorithm (GA) specifically with regard to genetic linkage. A central tenet of GA theory is the construction, preservation and mixing of building blocks (BBs). This is commonly attempted by judicious ordering of the variables on the chromosome to create ´tightly-encoded´ building blocks. However, this approach makes several demands on the user, including prior knowledge of variable dependencies and an understanding of the inherent bias of the standard recombination operators. A technique is proposed with the aim of allowing the user to explicitly and transparently direct the recombinative bias of the search at run-time, based on problem-specific insights they gain from a high-level visualisation.