Genetic Optimization of Cellular Encodings for Neural Networks

点击翻译按钮获取中文摘要

9312748 Whitley A grammar tree encodes a cellular developmental process that can develop whole families of Boolean neural networks. The development process resembles biological cell division. A single cell executes a program specified by a grammar tree and undergoes cell division to develop a complete neural network capable of computing specific target functions. A genetic algorithm is used to find a grammar tree that yields both architecture and weights specifying a particular neural network for solving specific Boolean functions, such as parity and symmetry problems. Because cellular encodings include recursive descriptions, the resulting "genetic code" is much smaller than the neural networks that are developed from these codes. Preliminary research has also shown the effectiveness of very simple Hebbian learning for speeding up the evolution of grammar trees. The proposed research will continue to develop grammar trees that specify neural networks for a wider variety of applications as well as to develop more sophisticated learning models to use in combination with cellular development. The specification of nonBoolean networks must be investigated so as generalized the types of networks that can be generated by cellular development. Genetic algorithms are being used to evolve the grammar tree for the cellular development process; as a result, learning algorithms that have minimal computational requirements will be investigated