page 1  (8 pages)
2to next section

Proceedings of the International Conference on Artificial Neural Networks and Genetic Algorithms, pp. 468-71, 1995

USING EVOLUTIONARY COMPUTATION TO GENERATE TRAINING SET DATA FOR NEURAL NETWORKS?

Dan Ventura
Tim Andersen
Tony R. Martinez

Provo, Utah 84602Computer Science Department, Brigham Young University e-mail: dan@axon.cs.byu.edu, tim@axon.cs.byu, martinez@cs.byu.edu

Most neural networks require a set of training examples in order to attempt to approximate a problem function. For many real-world problems, however, such a set of examples is unavailable. Such a problem involving feedback optimization of a computer network routing system has motivated a general method of generating artificial training sets using evolutionary computation. This paper describes the method and demonstrates its utility by presenting promising results from applying it to an artificial problem similar to a realworld network routing optimization problem.

Introduction

Many inductive learning algorithms based on neural networks, machine learning, and other approaches have been developed and have been shown to perform reasonably well on a variety of problems [2][4]. Typically, neural networks (NN) perform inductive learning through the presentation of preclassified examples; however,

one of the largest obstacles faced in applying these algorithms to realworld problems is the lack of such a set of training examples. Many times collecting data for a training set is the most difficult facet of a problem.
This paper presents such a real-world problem -- one for which no training data exists and for which gathering such data is at best extremely expensive both in time and in resources. To remedy the lack of training set data, a method using evolutionary computation (EC) [3][8] is described in which the survivors of the evolution become the training examples for a neural network. The synthesis of EC with NN provides both initial unsupervised random exploration of the solution space as well as supervised generalization on those initial solutions. Work involving a combination of EC and NN is becoming more prevalent; the reader is referred to [1][5][6][7] for examples.

? This research was funded in part by a grant from Novell, Inc.