Novel method of virtual embryogenesis for structuring Artificial Neural Network controllers

The organization of an Artificial Neural Network (ANN; e.g. the organization in layers, the number of cells per layer and the degree of connectivity between the cells) has a big influence on its abilities (e.g. learning ability). In this article, we present a novel method to organize the nodes and links of an ANN in a biologically motivated manner using virtual embryogenesis (VE). The VE mimics processes observable in biology, like interaction of cells via chemical substances or tissue differentiation. In our system, a virtual embryo consists of individual cells controlled by a genome. These cells can develop to nodes in the ANN during the embryogenetic process. The embryo is implemented as a spatially and temporally discrete multi-agent model. The cells in our model interact with each other via virtual physics and virtual chemistry. With the work at hand, we show that patterns developing in VE are comparable to patterns found during natural embryogenesis. We plan to combine VE with Evolutionary Algorithms to optimize the genome of the embryo. We expect the described model of VE (in combination with Evolutionary Algorithms) to lead to novel, evolutionary shaped net structures of ANNs.