Evolving Neural Network Structures Using Axonal Growth Mechanisms

In the eld of arti cial evolution creating methods to evolve neural networks is an important goal. But how to encode the structure and properties of the neural network in the genome is still a problem. If one overloads the genome with detailed information for a network the evolutionary time increases prohibitively. If the genome is too simple, only simple problems can be solved. As Nature has found an e cient and evolvable solution to this problem, it is worthwhile imitating the mechanisms on how biological neural nets are generated. In this paper I propose a model in which arti cial genes tune the ability of axons to nd, detect and connect to speci c targets. Initial simulation results of simple tasks are evolved and the genetic tuning of the developmental processes for arti cial evolution is discussed.