Date of Award

Fall 1995

Document Type

Thesis - Restricted

Degree Name

Master of Science (MS)

Department

Computing

First Advisor

Feng, Xin

Second Advisor

Belfore, Lee

Third Advisor

Heinen, James A.

Abstract

This thesis is concerned with a numerical approximation technique for feedforward artificial neural network (FFANN) weight computation. In most cases, a feedforward artificial neural network is trained for computing the weights, but the training takes time and may not converge [7] if the initial weights and learning rate are chosen incorrectly. Various methods have been proposed for fast training [3] [4] [6], global minimum finding [5] [8], new types of ANNs [12] [13] [14], and initial weight choice [10] [40]. Only a few papers have been published on restricted direct weight computation [11] [28]. Based on the fact that most FFANNs have many solutions and some have an infinite number of solutions, the approximation technique is developed for finding one of the best solutions. To avoid the linear dependent property in each layer of FFANNs and to compute weights, a decoupled FFANN (DFFANN) structure is developed in Chapter 3. Training results in Chapter 5 show that DFFANNs have smaller errors than FFANNs after the same epochs of training. According to the theory of Halbert White [1], the ANNs are expressed as approximation functions. Using the approximation technique and these expressions, the author established a new direct weight computation method in Chapter 4. Five benchmark problems are used as experiments for testing the direct weight computation methods and DFFANNs in Chapter 5. A new learning algorithm is proposed in Chapter 6. Because of the strong nonlinear property of FFANNs, the direct weight computation is a very difficult research topic. However, trying to approximate optimal weights by means of the numerical approximation technique offered rather good results. In some situations, searching the weight; by traditional training is difficult. But the numerical approximation can reach the weights near the optimal solution, and the computed weights can be used as the initial values for training. The contributions of this research are developing a decoupled FFANN structure, establishing a new direct weight computation method and testing them in five benchmark problems.

Share

COinS

Restricted Access Item

Having trouble?