Document Type

Thesis

Date of Award

1999

Keywords

algorithms, neural networks, computer science

Degree Name

Master of Arts (MA)

Department

Computer Science

First Advisor

Walker Land

Second Advisor

Leslie Lander

Abstract

Much research has been done to find faster training algorithms for neural networks. There are many promising training algorithms available, yet it may take days for a neural network to train if it has a sufficiently large problem to solve. Simulated annealing algorithms are extremely fast compared to more traditional algorithms, and are being further developed to reduce training time. The stochastic smoothing with gradient hints algorithm (SSG) is noteworthy because it has much potential in outperforming even the best training algorithms (like the Levenberg- Marquardt algorithm). The SSG algorithm performs very well during the initial phases of training, but then its performance decreases, showing no significant improvements in the later stages of training. Ideally, it would make significant improvements throughout the entire training process. There have been versions of the SSG algorithm that have successfully worked; however, these algorithms have been designed with a specific data set/problem in mind, and are not suited for general problem solution. The objective is to design an SSG algorithm that is very general and works for a variety of problems, not just for certain ones. The importance of making the SSG method work is considerable; it would make available an extremely fast and versatile generalized algorithm that could be used in a variety of situations. Therefore this algorithm, if it were to converge quickly and consistently, would be a very valuable contribution to the existing library of algorithms. This thesis presents the analysis and suggested improvements to the SSG algorithm, with the objective of developing a versatile and fast-training algorithm.

Download

Share

COinS