Neural Networks for RF and Microwave Design
An electrical engineering textbook on practical applications of artificial intelligence technology methods to solving complex radio-frequency and microwave design problems.
TABLE OF CONTENTS Neural Networks for RF and Microwave Design
Chapter 9: Knowledge-Based ANN Models
This chapter [1]describes some of the recent advances in theneural network area where existing RF/microwave knowledge is combined withneural networks. The chapter starts with a brief description of knowledgeissues in conventional neural network applications (e.g., patternclassification) and in microwave design. We then study various effects ofadding knowledge on the performance of the neural models, such asgeneralization ability, extrapolation ability, and model reliability, versusdifferent sizes of training data through a knowledge-based neural network(KBNN) technique and demonstrate examples of comparisons with conventional MLP(without any knowledge base). We also describe several ways of combiningexisting circuit models with neural networks, including the source differencemethod, the prior knowledge input method, and the space-mapped neural models.Finally, an advanced hierarchical neural network structure for the task ofneural model library development is discussed.
9.1 Introduction
9.1.1 Motivation
With continuing developments in applications of neural networks tomicrowave design, there is a growing need for reduction in the cost of modeldevelopment and improvement in model reliability. The commonly used MLP modelbelongs to the type of black-box models structurally embedding noproblem-dependent information/knowledge. Therefore, it derives the entireinformation about theRF/microwave behaviors from the training data.Consequently, a large amount of training data is needed to ensure modelaccuracy. In microwave applications, training data is obtained either bysimulation of original EM/device-physics problems, or by measurements.Generating a large amount of training data could be very expensive, becausesimulation/measurement may have to be performed at many points in the modelinput parameter space (e.g., for various combinations ofgeometrical/process/bias parameters). Without sufficient training data, theresulting neural models may not...
Copyright ARTECH HOUSE Inc. 2000 under license agreement with Books24x7
