Date of Award

6-13-2013

Document Type

Thesis

Degree Name

Master of Science

Department

Department of Electrical and Computer Engineering

First Advisor

Gilbert L. Peterson, PhD.

Abstract

Artificial Immune Systems (AISs) are a type of statistical Machine Learning (ML) algorithm based on the Biological Immune System (BIS) applied to classification problems. Inspired by increased performance in other ML algorithms when combined with kernel methods, this research explores using kernel methods as the distance measure for a specific AIS algorithm, the Real-valued Negative Selection Algorithm (RNSA). This research also demonstrates that the hard binary decision from the traditional RNSA can be relaxed to a continuous output, while maintaining the ability to map back to the original RNSA decision boundary if necessary. Continuous output is used in this research to generate Receiver Operating Characteristic (ROC) curves and calculate Area Under Curves (AUCs), but can also be used as a basis of classification confidence or probability. The resulting Kernel Extended Real-valued Negative Selection Algorithm (KERNSA) offers performance improvements over a comparable RNSA implementation. Using the Sigmoid kernel in KERNSA seems particularly well suited (in terms of performance) to four out of the eighteen domains tested.

AFIT Designator

AFIT-ENG-13-J-07

DTIC Accession Number

ADA580069

Download

Share

COinS