Main content

Regression and Classification from Extinction

Show full item record

Title: Regression and Classification from Extinction
Author: Brown, Joseph Alexander
Department: School of Computer Science
Program: Computer Science
Advisor: Ashlock, Daniel
Abstract: Evolutionary Algorithms use the principles of natural selection and biological evolution to act as search and optimization tools. Two novel Spatially Structured Evolutionary Algorithms: the Multiple Worlds Model (MWM) and Multiple Agent Genetic Networks (MAGnet) are presented. These evolutionary algorithms create evolved unsupervised classifiers for data. Both have a property of subpopulation collapse, where a population/node receives little or no fitness implying the number of classes is too large. This property has the best biological analog of extinction. MWM has a number of evolving populations of candidate solutions. The novel fitness function selects one member from each population, and fitness is divided between. Each of these populations meets with the biological definition of a separate species; each is a group of organisms which produces offspring within their type, but not outside of it. This fitness function creates an unsupervised classification by partitioning the data, based on which population is of highest fitness, and creates an evolved classifier for that partition. MAGnet involves a number of evolving agents spread about a graph, the nodes of which contain individual data members or problem instances. The agents will in turn test their fitness on each of the neighbouring nodes in the graph, moving to the one where they have the highest fitness. During this move they may choose to take one of these problem instances with them. The agent then undergoes evolutionary operations based on which neighbours are on the node. The locations of the problem instances over time are sorted by the evolving agents, and the agents on a node act as a classifier.
Date: 2014-01

Files in this item

Files Size Format View Description
Brown_Joseph_201401_PhD.pdf 14.65Mb PDF View/Open Thesis

This item appears in the following Collection(s)

Show full item record