Evolutionary Programming and Neural Computation (Computer and Information Science)

点击翻译按钮获取中文摘要

The goal of evolutionary programming is to develop systems which learn how to perform critical computing tasks through a variation and selection algorithm. The appproach taken in this research is based on the recognition that some structural organizations are well-suited to learning through evolution, while others are not. A system that achieves evolvability by simulating the structure-fashion relations that facilitate evolution in biological systems has been developed. Currently the system consists of simulated networks of dynamic neurons whose input-output behavior is controlled by membrane proteins. The variation and selection learning algorithm acts on the distribution of proteins. The organization is well-suited to evolution because improvement can always be achieved through single, reasonably likely application of the variation operator. The continuous dynamics displayed by the neurons also enhances the generalization capability of the system. The system is called a molecular design since it was inspired by biophysical and biochemical properties of neurons. However, the new efforts will be concerned with virtual implementations that would be useful in and of themselves and which could provide some insight into natural information processing. Specific objectives are: to incorporate richer nonlinear and cell automation dynamics, to add a memory manipulation capability, to implement the system as a preprocessor for high-level vision and process control tasks, and to create systems capable of long-term evolutionary improvement by applying the variation and selection operators to the dynamics.