Generation of Recurrent Neural Networks Using an Evolutionary Method Inspired by Gene Regulation in Biological Organisms

In this thesis an evolutionary method, inspired by gene regulation in bio-logical systems, is used to generate recurrent neural networks (RNNs). The method is employed to study whether it is possible to obtain RNNs that can change their structure easily, which is a desirable property e.g. when the RNNs are used as artificial brains in autonomous robots. In most simulations using genetic algorithms in connection with neural networks, a direct mapping method is used where the genome serves as a lookup table, and only a minor distinction can be made between the inherited genetic code (genotype) and the resulting organism (phenotype). By constrast, in this thesis, a method is used where the genotype-to-phenotype mapping is based on a simplified biological genetic regulatory network. Here, each gene is given an input sequence and an output sequence. If the output sequence of one gene matches all or part of the input sequence of another gene, the first gene binds to the second gene forming a connectio