Hands-On Neuroevolution with Python

Indirect genome encoding

To overcome size issues with direct encoding, Kenneth O. Stanley proposed an indirect encoding method, which is inspired by how the phenotype is encoded by the genome in the DNA. It is based on the fact that the physical world is built around geometry and regularities (structural patterns), where natural symmetries are found everywhere. Thus, the encoding size of any physical process can be significantly reduced through the reuse of a specific set of encoding blocks for the same structure that repeats many times. The proposed method, called Hypercube-based NeuroEvolution of Augmenting Topologies (HyperNEAT), is designed to build large-scale neural networks by exploiting geometrical regularities. HyperNEAT employs ...