Kent Academic Repository
Full text document (pdf)
Copyright & reuse
Content in the Kent Academic Repository is made available for research purposes. Unless otherwise stated all
content is protected by copyright and in the absence of an open licence (eg Creative Commons), permissions
for further reuse of content should be sought from the publisher, author or other copyright holder.
Versions of research
The version in the Kent Academic Repository may differ from the final published version.
Users are advised to check http://kar.kent.ac.uk for the status of the paper. Users should always cite the
published version of record.
Enquiries
For any further enquiries regarding the licence status of this document, please contact:
researchsupport@kent.ac.uk
If you believe this document infringes copyright then please contact the KAR admin team with the take-down
information provided at http://kar.kent.ac.uk/contact.html
Citation for published version
Watkins, Andrew and Timmis, Jon and Boggess, Lois C. (2004) Artificial Immune Recognition
System (AIRS): AN Immune Inspired Supervised Machine Learning Algorithm. Genetic Programming
and Evolvable Machines, 5 (3). pp. 291-317. ISSN 1389-2576.
DOI
https://doi.org/10.1023/B:GENP.0000030197.83685.94
Link to record in KAR
https://kar.kent.ac.uk/14200/
Document Version
UNSPECIFIED
Artificial Immune Recognition System (AIRS): An Immune-
Inspired Supervised Learning Algorithm
Andrew Watkins
∗
and Jon Timmis
(abw5,jt6@kent.ac.uk)
Computing Laboratory, University of Kent. CT2 7NF. UK.
Lois Boggess
(lboggess@cse.msstate.edu)
Department of Computer Science and Engineering, Mississippi State University,
USA
Abstract. This paper presents the inception and subsequent revisions of an immune-
inspired supervised learning algorithm, Artificial Immune Recognition System (AIRS).
It presents the immunological components that inspired the algorithm and describes
the initial algorithm in detail. The discussion then moves to revisions of the basic
algorithm that remove certain unnecessary complications of the original version.
Experimental results for both versions of the algorithm and are discussed and these
results indicate that the revisions to the algorithm do not sacrifice accuracy while
increasing the data reduction capabilities of AIRS.
Keywords: supervised learning, artificial immune systems, classification, neural
networks
1. Introduction
In recent years there has been considerable interest in exploring and
exploiting the potential of Artificial Immune Systems for applications
in computer science and engineering. These systems are inspired by
various aspects of the immune systems of mammals. Some of these as-
pects, such as the distinction between self and non-self and the concept
of negative selection, have a natural and intuitive fit for applications
involving computer security, network intrusion detection [16], [20], [23],
change detection, and the like. Moreover, research into natural immune
systems suggests the existence of learning properties which may be used
to advantage in machine learning systems [2]. With the exception of [7],
until very recently Artificial Immune System (AIS) research into ma-
chine learning has focused on the development of unsupervised learning
and clustering [10], [29] rather than supervised learning and reinforce-
ment learning
1
. In contrast, this paper presents an AIS classifier which
∗
Also, Department of Computer Science and Engineering, Mississippi State
University, USA.
1
While others have discussed the similarity of Artificial Immune Networks and
other connectionist models such as Artificial Neural Netorks [15], there seems to have
c
° 2003 Kluwer Academic Publishers. Printed in the Netherlands.
airs.tex; 11/06/2003; 12:44; p.1
2 Watkins, Timmis, and Boggess
initially was intended to show that a subset of metaphors from the
AIS literature could produce an effective supervised learning system
[31]. The classifier that resulted performs well on a variety of publicly
available test problems used by the machine learning community [32]
and has been the object of study both to further refine its algorithms
and to better understand the source of its classification power. This
paper describes the original algorithm developed in [31] and proposes
a modification which enhances the efficiency of the resulting classi-
fier while retaining classification accuracy. An empirical review of the
effects of this modification is presented. This paper is an expanded
version of [34].
2. Bio-Inspired Computing and Machine Learning
The means of constructing and developing machine learning systems
have been extremely diverse. These have ranged from the manipulation
of symbolic data in order to develop a concept learning system to the
heavy use of sub-symbolic representations of data in the development
of function approximation algorithms (e.g., back-propagation neural
networks). There are several good general surveys of the field of machine
learning and its techniques (foremost being [26]), and no attempt will be
made to duplicate that effort here. However, instead, the focus will be
on one particular emerging field of computing science that can, perhaps,
lend insight into the development of machine learning systems.
Since there seems to be a desire to build computational systems
that exhibit some (if not all) of human cognitive abilities, one of the
first clear examples of this has been in the proliferation of the field of
artificial neural networks (ANNs). This field, as can be surmised from
its name, has looked to the workings of the human brain as inspiration
for the development of computing systems. Since these workings are not
completely understood and since the goals of building our intelligent
systems are often problem-driven, ANNs have by no means been an
attempt to directly model all of the processes occurring within human
brains. Instead, however, certain observed phenomena have been sim-
plified and encoded in order to replicate some of the mechanisms of
our thought processes that might best be suited to manipulating data
and learning general patterns or trends in this data. Of course, pattern
recognition has not been the only application of ANN techniques. They
have shown to be successful for a wide variety of learning and control
problems [3], [19].
been no direct translation of these ideas to a well-tested immune-inspired supervised
learning system.
airs.tex; 11/06/2003; 12:44; p.2
Artificial Immune Recongition System 3
A second, obvious example, of a computing paradigm inspired by ob-
servations of natural phenomena is that of Genetic Algorithms (GAs).
Looking to neo-darwinian evolutionary theory with an emphasis on re-
production, mutation, and genetic crossover, GAs have been successful
in solving certain difficult or computationally expensive optimization
problems. Darwinian evolutionary theories have also inspired other
successful evolutionary programming techniques [25]. While these two
examples have been the most prolifically discussed computational ideas
inspired by natural observations, they are not the only ones. In recent
years, we have seen the exploration of other metaphors from nature as
applied to computing problems. These include swarming insects [22],
ant colonies [5], and mammalian immune systems [9]. It is this latest
source of inspiration, mammalian immune systems, and the use of this
natural system as a guide to developing machine learning systems that
will be the particular focus of this article.
2.1. The Artificial Immune Recognition System
AIRS (Artificial Immune Recognition System) is a novel immune in-
spired supervised learning algorithm [31]. Motivation for this work came
from the author’s identification of the fact that there was a significant
lack of research that explored the use of the immune system metaphor
for supervised learning; indeed, the only work identified was that of
[7]. However, it was noted that within the AIS community there had
been a number of investigations on exploiting immune mechanisms for
unsupervised learning (that is, where the class of data is unknown
apriori) [30], [29] and [11]). Work in [10] examined the role of the
clonal selection process within the immune system [6] and went on to
develop an unsupervised learning known as CLONALG. This work was
extended by employing the metaphor of the immune network theory
[21] and then applied to data clustering. This led to the development
of the aiNet algorithm [11]. Experimentation with the aiNet algorithm
revealed that evolved artificial immune networks, when combined with
traditional statistical analysis tools, were very effective at extracting in-
teresting and useful clusters from data sets. aiNet was further extended
to multimodal optimization tasks [8]. Other work in [30] also utilized
the immune network theory metaphor for unsupervised learning, and
then augmented the work with the development of a resource limited
artificial immune network [29], which reported good benchmark results
for cluster extraction and exploration with artificial immune networks.
Indeed, this work has been further extended by [27] with the introduc-
tion of fuzzy logic and refinement of various calculations. The work in
[29] was of particular relevance to [31] and the further work described in
airs.tex; 11/06/2003; 12:44; p.3
4 Watkins, Timmis, and Boggess
this paper, which builds on this previous work, in particular the ideas
of artificial recognition balls and resource limitation from [29] and long-
lived memory cells from [11]. However, while these population control
mechanisms and data representation concepts were borrowed from this
work on immune networks, it should be stressed that AIRS is in no
way an immune network model of compuation. The rest of this section
and describes the immune metaphors that have been employed within
AIRS.
2.2. IMMUNE PRINCIPLES EMPLOYED
A little time should be taken to draw attention to the most relevant
aspects of immunology that have been utilized as inspiration for this
work. A more detailed overview of the immune system and its rela-
tionship with computer science and engineering can be found in [9].
Throughout a person’s lifetime, the body is exposed to a huge variety
of pathogenic (potentially harmful) material. The immune system con-
tains lymphocyte cells known as B- and T-cells, each of which has a
unique type of molecular receptor (location in a shape space). Receptors
in this shape space allow for the binding of the pathogenic material
(antigens), with higher affinity (complementarity) between the receptor
and antigen indicating a stronger bind. Work in [9] adopted the term
shape-space to describe the shape of the data being used, and defined
a number of affinity measures, such as Euclidean distance, which can
be used to determine the interaction between elements in the AIS.
Within AIRS (and most AIS techniques) the idea of antigen/antibody
binding is employed and is known as antigenic presentation. When
dealing with learning algorithms, this is used to implement the idea
of matching between training data (antigens) and potential solutions
(B-Cells). Work in [29] employed the idea of an artificial recognition
ball (ARB), which was inspired by work in [14] describing antigenic
interaction within an immune network. Simply put, an ARB can be
thought to represent a number of identical B-Cells and is a mechanism
employed to reduce duplication and dictate survival within the pop-
ulation. Once the affinity between a B-Cell and an antigen has been
determined, the B-Cell involved transforms into a plasma cell and ex-
periences clonal expansion. During the process of clonal expansion, the
B-Cell undergoes rapid proliferation (cloning) in proportion to how well
it matches the antigen. This response is antigen specific. These clones
then go through affinity maturation, where some undertake somatic
hypermutation (mutation here is inversely proportional to antigenic
affinity) and eventually will go through a selection process through
which a given cell may become a memory cell. These memory cells
airs.tex; 11/06/2003; 12:44; p.4
