The Adaptive Distributed Network Computing for Intelligent System Design project is undertaken on the platform of Windows NT. It is covered in my Ph.D thesis. Followed is the abstract of this project.

It is widely acknowledged that non-linear systems can be represented by various regressive methods involving polynomial, spline, or trigonometric fixed basis expansions. More recently it has been shown that, in lieu of the traditional fixed basis expansions, adaptive multi-variable non-linear basis involving neurocomputing networks are more computationally efficient regressors for purpose of function approximation and prediction. But an adaptive method which is suited (i.e., most efficacious) for learning all non-linear systems is indeed a grand challenge and of particular appeal regarding computational tractability.

To address this problem of an efficacious function learning method, a novel viewpoint referred to as functional space is proposed. Through this new viewpoint, for the first time, we give the geometric interpretation of regularization. Of significance regarding functional space is deterministic or function learning as distinguished from non-deterministic noise, and more important, how they can be clearly separated in functional space.

Based upon the theorems derived for functional space, this thesis proposes an innovative neural network structure - Orthogonal Functional Basis Neural Networks (OFBNN) and an efficient learning algorithm - Orthogonal Functional Basis Functional Mapping (OFBFM). As compared with other adaptive neurocomputing approaches, OFBNN employs techniques such as regularization and general-cross-validation for both improved computational tractability and generalization performance. OFBNN and OFBFM are currently being proposed for material process modeling problem such as in situ process control of nonlinear optical filters. Also to be discussed is a proposed approach for applying OFBNN to telecommunication network problems such as the `connection-admission-control' problem in Asynchronous Transfer Mode (ATM) networks for optimizing resources (i.e., bandwidth) while maintaining Quality of Service (QOS). The ATM application is developed over Windows NT.

Yang Cao
Research Assistant
Air Force Research Lab
Ph.D Candidate
Wright State University
Phone: (937)879-1281
Email: ycao@cs.wright.edu
Postal Address:
712 Cedar Dr. Apt 6
Fairborn OH 4532