Modelling the world with statistically neutral PBGAs. Enhancement and real applications

This paper is concerned with extending and enhancing the use of promoter genes and introns in the encoding of variable length artificial neural network structures for their evolution. These structures have become necessary due to the requirements imposed by the problem we are tackling, that is, the real time evolution of world and internal models for robots operating in changing environments. Promoter Based Genetic Algorithms (PBGA) contemplate the evolution of the architecture and weight values of artificial neural networks regulating the expression of the different genes in the chromosome in a statistically neutral manner. A non direct genotype-phenotype transformation is thus obtained which becomes very efficient in dynamic environments. We study new features in the algorithm that permit achieving very good solutions in modelling the world for real robot applications without predetermining number of neural nets that will collaborate in order to achieve the world model.