Application areas of AIS: The past, the present and the future
Introduction
The artificial immune system (AIS) community has been vibrant and active for a number of years now, producing a prolific amount of research ranging from modelling the natural immune system, solving artificial or bench-mark problems, to tackling real-world applications, using an equally diverse set of immune-inspired algorithms. Whilst it is natural, and indeed healthy, for a somewhat scattergun approach to be taken in the early days of developing any new paradigm, in the sense that high-level, often naïve metaphors are selected and applied to problem areas that have often been tackled with other paradigms, there comes a point at which research effort needs to have a more coherent focus in order to more clearly define the field, and allow it to go forward and be fully exploited. We argue that this point has now been reached in the AIS community—with a solid foundation of published work to build on, the time has come to try and define the role that AIS can play and the type of applications that will really allow its potential to be realised.
Without a doubt there have been a lot of successful applications of AIS, and these should not be ignored. However, at this point, there are still few exemplars that really stand out as instances of successfully applying an AIS to a hard, real-world problems, or of AIS being used in earnest in industry. This is in contrast for example to the field of Evolutionary Algorithms, where at the most recent flagship conference in the field, GECCO 2004 [4], there were 38 papers describing the applications of EAs to real-world problems, and the EVONET repository [2] is able to list 39 examples of Evolution at Work, i.e. practical applications of EAs. On the one hand, this is somewhat of an unfair comparison, given the relative time-periods that the two fields have been active, however it illustrates the importance of focusing research effort in the next few years in order to provide hard evidence of a distinctive niche for AIS.
For any new paradigm to prove itself is always a difficult task—there is a lot of good competition from existing tried and tested algorithms. There has perhaps been a natural tendency for AIS to be compared to other biologically inspired paradigms such as Evolutionary Algorithms, Neural-networks, and to other more traditional classification or clustering algorithms. Scientifically, it is essential that such comparisons to be made; however, we argue that it is not sufficient for AIS simply to outperform other algorithms on any given set of problem instances to be declared useful. For a start, test instances (particularly benchmarks) are not necessarily difficult, and any number of other problem instances can be generated on which performance will be unknown. Secondly, in the light of the no-free lunch theorem [86], we cannot expect any one algorithm to outperform all others given all possible problem instances. We argue that for a paradigm to be truly successful, it should contain features that are not present in other paradigms and/or it should give rise to emergent system properties which are not obtained through other paradigms. Thus, the algorithm can be described as distinctive. This is in contrast to work in [34] which suggests that an appropriate manner in which to assess AIS algorithms is by their distinctiveness and effectiveness. Garrett [34] defines distinctiveness in terms of an algorithm containing novel symbols, expressions or processes. We counter that this definition is not strong enough; some of the essential characteristics of immune algorithms are emergent properties that arise from the interaction of a number of processes which in themselves may not be described as distinct. In addition, whilst it is important that an algorithm is effective, we disagree with the definition proposed in [34] which measures effectiveness in terms of an algorithm performing better than other algorithms on shared benchmarks tests, quicker than other algorithms or providing a unique method of obtaining results. The problems associated with benchmarking data-sets have already been outlined and hence we do not believe that only being better or quicker will allow AIS to prove itself. Furthermore, in a climate in which there are any number of algorithms (biological and otherwise) inspired by an equally diverse range of processes, it seems critical that an algorithm must achieve something that cannot be achieved by any other means in order to earn its place in history.
In this position paper, we hope to extract some general features of problems that we believe will allow AIS to really bring some benefit, and thus distinguish it from other techniques. We suggest that the way forward for AIS is in part a focussed attempt to carefully select application areas based on mapping problem features to mechanisms exhibited by the IS, taking the problem-oriented perspective outlined by example in [73], [32], [13], and discussed further in Section 4.2. However, we emphasise that application development needs to be under-pinned with a continuing line of research into the theoretical basis of AIS and with the overriding need for extraction of novel and accurate metaphors from immunology.2
Section snippets
Survey of existing application areas
In order to place the following discussions in context, we first present a general review of application areas to which AIS has currently been applied. The following brief summary is based in part on a bibliography produced by De Castro [23], used in a tutorial at ICARIS 2004 [24] on Engineering Applications of AIS. The information contained in this tutorial has been expanded to include references from ICARIS 2004 [3] and ICARIS 2005 [49]. A useful summary of application areas can also be found
“Was it worth it”—a look at the added value of the AIS
It is now pertinent to re-evaluate the application of immune algorithms to the above application areas, and question whether there is really any added value in applying AIS to the three areas listed above. Again we re-iterate that there is no doubt that AIS has been successful in these areas; however, we question as to whether they AIS brings any benefits that could not have been gained from applying a different sort of algorithm. Recall the seminal list of features of an AIS, originally due to
A new approach to AIS
The above discussion has shed a rather gloomy light on future of AIS in solving real-world applications. Perhaps this is a suitable point to take a step backwards and first re-evaluate our approach to designing AIS algorithms, as well as attempting to define what kind of applications they may be suitable for. With this in mind, we take brief look at both sides of the coin and take first an algorithm-oriented and then a problem-oriented view of the situation.
Suggestions as to the way forward
We have outlined what we believe to be the problems with the current applications to which AIS has been applied, from the perspective that although reasonably successful on a narrow range of problems, they do not add sufficient value over and above that which is offered by other paradigms to make them anything other than another tool in the engineers application tool-box. Although from some points of view, any tool is a worthwhile addition, we believe there is still a wealth of unexploited
Conclusions: features of AIS applications
We summarise by proposing a list of features that draw together some of the preceding discussion and that we believe point to the way forward for AIS. Some of these features are currently absent in any of the AIS literature. Others, such as life-long learning, have been modelled in a limited sense. We emphasise that it is by the combination of these principles that a distinctive niche is carved for AIS.
- (1)
They will be embodied.
- (2)
They will exhibit homeostasis.
- (3)
They will benefit from interactions
Acknowledgments
Many of the ideas in this paper have evolved from useful and stimulating discussions at the ICARIS conferences with many people in the field, and in particular at meetings of the UK based ARTIST [1]. We would like to thank ARTIST for financial support provided during the development of ideas in this paper.
References (89)
- et al.
The immune system, adaptation, and machine learning
Physica
(1986) - et al.
Immunity-based autonomous guided vehicles control
Appl. Soft Comput.
(2007) - et al.
An immunity based ant colony optimization algorithm for solving weapon-target assignment problem
Appl. Soft Comput.
(2002) - Artificial immune systems, in: Proceedings of ICARIS 2004, LNCS 3239, Springer,...
- Deb et al., Proceedings of Genetic and Evolutionary Computation Conference, Springer,...
- et al.
The danger theory and its application to artificial immune systems
- et al.
Danger theory: the link between AIS and IDS?
- M. Ayara, J. Timmis, R. de Lemos, S. Forrest, Immunising automated teller machines, in Jacob et al. [49], pp....
Coverage and generalization in an artificial immune system
Or-library: distributing test problems by electronic mail
J. Oper. Res. Soc.
Immune network and adaptive control.
Revisiting idiotypic immune networks
The Immune Learning Mechansims: Recruitment, Reinforcement and their Applications
A grasp for job shop scheduling
A modified immune network algorithm for multi-modal electromagnetic problems
IEEE Trans. Magn.
A markov chain model of the b-cell algorithm
Use of an artificial immune system for job shop scheduling
Exploring the capability of immune algorithms: a characterization of hypermutation operators
Negative selection algorithm for aircraft fault detection
The clonal selection algorithm with engineering applications
Ainet: an artificial immune network for data analysis
A formal framework for positive and negative detection schemes
IEEE Trans. Syst., Man Cybern. Part B
Computer immunology
Commun. ACM
Self-nonself discrimination in a computer
Revisiting the foundations of artificial immune systems: a problem oriented perspective
How do we evaluate artificial immune systems?
Evol. Comput.
From GAs to artificial immune systems: improving adaptation in time-dependent optimization.
An innately interesting decade of research in immunology
Nat. Med.
A randomized real-valued negative selection algorithm.
Proceedings of the 2nd International Conference on Artificial Immune Systems (ICARIS)
Lect. Notes Comput. Sci.
Introducing dendritic cells as a novel immune-inspired algorithm for anomaly detection
Exploiting the analogy between the immune system and sparse distributed memories
Genet. Prog. Evol. Mach.
Studies on the implications of shape-space models for idiotypic networks
The impact of the shape of antibody recognition regions on the emergence of idiotypic networks
J. Unconven. Comput.
Cited by (0)
- 1
Tel.: +44 1904 432348; fax: +44 1904 432335.