Using Modular Neural Networks with Local Representations to Control Dynamic Systems.

摘要

The objective of this research is to develop an artificial neural network with very fast learning. Major areas of activity during this period include developing cross validation methods to refine parameters such as the distance metric, developing parallel versions of the learning algorithms which allow implementations to be scaled up by simply adding additional processing hardware, with a negligible penalty in processing time, and performing numerical experiments on simulated data to test the approach. We have also compared our approach with other neural network approaches, and found that it provides equal or better performance. We have implemented versions of this approach on several platforms: serial computers (standard Sun workstations), digital signal processors (Intel i860), a parallel computer (Connection Machine), and using special purpose circuitry. We are now convinced we can perform training and access sufficiently quickly to allow real time learning.