Genetic Synthesis of Unsupervised Learning

Algorithms

Ali DAS?DAN and Kemal OFLAZER

Department of Computer Engineering and Information Science

Bilkent University

06533 Bilkent, Ankara, TURKEY

Email : [email protected]

Abstract

This paper presents new unsupervised learning algorithms that have been synthesized using a genetic approach. A set of such learning algorithms has been compared with the classical Kohonen's Algorithm on the Self-Organizing Map and has been found to provide a better performance measure. This study indicates that there exist many unsupervised learning algorithms that lead to an organization similar to that of Kohonen's Algorithm, and that genetic algorithms can be used to search for optimal algorithms and optimal architectures for the unsupervised learning.

1 Introduction

Genetic algorithms (GAs) [2, 4] are search and optimization algorithms which emulate the mechanics of natural evolution. GAs operate on a population of individuals, each of which consists of a string, called a chromosome, that encodes a solution to the problem being solved. Each individual is used to explore a different region of the search space of the problem. At each iteration, called a generation, a new population is created using probabilistic rules to exchange the information in the chromosomes of the individuals. In this way, GAs both exploit the information already gained and explore new regions in the search space.

The design of artificial neural networks (ANNs) whose performance is optimized for a certain application is still a research issue [8]. This design problem is also a complicated one because of the large design space, the presence of many variables, and complex interactions among these variables [5]. GAs can be used to search for optimal ANNs. A number of researchers have already applied GAs to synthesize application-specific ANNs [8]. Some have used GAs to determine the connection weights in an ANN, while others have used them to find the optimal topology and optimal parameters of a certain learning rule for an ANN. Chalmers [1] has used a GA to find better learning rules for a single layer ANN. Up to now, all the applications have been done in ANNs employing a supervised learning process.