L. Nunes de Castro, F.J. Von Zuben, and H. Knidel (Eds.): ICARIS 2007, LNCS 4628, pp. 264–275, 2007. © Springer-Verlag Berlin Heidelberg 2007

A Computational Model for the Cognitive Immune System Theory Based on Learning Classifier Systems

Daniel Voigt, Henry Wirth, and Werner Dilger

Chemnitz University of Technology D-09107 Chemnitz, Germany

{davo,henw,dilger}@informatik.tu-chemnitz.de

Abstract. In the past there have been several approaches to use Learning Clas-sifier Systems (LCS) as a tool for modelling the functioning of the immune sys-tem. In this paper we propose a modification of the classic LCS that can be used for modelling the Cognitive Immune System Theory introduced by I. Cohen. It has been pointed out before that this alternative view of the immune system and its agents provides promising functional perspectives to the field of artificial immune systems (AIS). The characteristic features of Cohen's theory, namely degeneracy of recognition and context of immune reactions, and how they can be realized in our modified LCS are described. Moreover, we introduce the re-presentations of the immune agents, the interactions that take place among them and the applied evolutionary mechanisms.

Keywords: Cognitive Immune System, Modelling, Learning Classifier Systems, Degeneracy, Cytokines.

1 Introduction

Most of the computational systems developed in AIS are based on the two leading theories in the field of immune system research, namely Burnet's Clonal Selection Theory [1] and Jerne's Network Theory [2]. But there have always been divergent views on immune activity – even though some of them turned out to be more bio-logically plausible than others (see [3]). In recent years, I. Cohen has suggested an alternative approach to understanding the functioning of the immune system as a whole which is based on the Network Theory but goes far beyond it (see [4]). He considers the immune system to be a cognitive system as it senses certain molecular aspects of its environment, creates an internal representation of it, and makes deci-sions about the actions that are required to keep the homeostasis of the individual. Characteristic features of his theory are the degeneracy of recognition events, which contrasts sharply with the assumption of monospecificity, and the emphasis on im-mune activity that is embedded in a context created by interacting immune agents. It has been pointed out that the field of AIS can benefit from computational models that are derived from such new immune theories (see [5]).

Several authors (see [6], [7], [8] and [9]) have demonstrated how LCS can be used as a framework for implementing immune-inspired computational models with

A Computational Model for the Cognitive Immune System Theory 265

promising problem solving capabilities. In this paper we present a modification of the LCS and show how the important aspects of the Cognitive Immune System Theory can be modelled in this computational system. Section 2 summarizes the classic form of the LCS. In section 3 a brief introduction to the basics of the Cognitive Immune System Theory is given. Section 4 describes how the features of the theory are implemented in our modified LCS. The final section 5 contains some remarks about possible fields of application and first experiences with the implemented computational system.

2 Learning Classifier Systems

This section briefly describes the basic components and mechanisms of the machine learning paradigm that was introduced by J. Holland and is summarized under the term LCS (see [10]). There have been several variations of the underlying architec-tures and algorithms and so it is difficult to pick out one basic LCS standard form. In our modelling approach we focus on a form which is almost identical to the one out-lined by Holland in [11]. We took this classic LCS as a starting point and modified its internal structure and processes according to the principles of Cohen's Cognitive Im-mune System Theory. Figure 1 shows the overall LCS and its computation loop.

ClassifierPopulation

Effectors

Environment

AuctionRound

GeneticAlgorithm

BucketBrigade

FeedbackOutput

MessageList

Detectors

1

Input

ExternalMessages

Effector RelatedMessages

InternalMessages

SatisfiedClassifiers Generates

NewClassifiers

AdjustsStrength

4

2

3

5

6Matching

Learning Classifier System (LCS)

Fig. 1. The structure and internal mechanisms of Holland’s classic LCS (based on [12])

A LCS consists of a population of binary rule strings called classifiers. A single classifier is a compound of two rule parts: the condition and the action. The condition part describes a subset of all possible message strings that will potentially trigger the

266 D. Voigt, H. Wirth, and W. Dilger

execution of the accompanying action part. Furthermore, to each classifier a strength value is assigned that represents an estimate about the respective rule string's average performance in selecting appropriate actions in the past. At the beginning of the LCS loop, the rule strings of the classifier population are randomly initialized and a pre-defined strength value is assigned to all classifiers. Then the message list is cleared and the system enters the following computation loop: (1) The detectors add binary message strings to the message list that describe certain observations on the environ-ment. (2) Each classifier's condition part is compared to each message located in the message list using a matching algorithm. All completely satisfied classifiers are added to the match set. (3) Subsequently, for each classifier of the match set a bid is com-puted that takes into account the classifier's current strength value and the specificity of its condition string. Based on these bids an auction round is held that stochastically determines the winner set by carrying out a roulette wheel selection among the classi-fiers of the match set. After that, the message list is cleared and all classifiers of the winner set publish their respective action string as a message in the message list. (4) The effectors read the relevant messages and perform the accompanying actions in the environment. (5) If the LCS receives feedback from the environment, the so-called bucket brigade algorithm adapts the strength values of certain classifiers. Thus, rule strings that triggered appropriate actions in the past are rewarded with positive feed-back and become more likely to win an auction round in the future. As a result of this reinforcement strategy the LCS is able to learn from experiences in the past. (6) If necessary, weak classifiers are replaced by new ones by applying a standard Genetic Algorithm (GA) to the classifier population, where a rule string's strength value is interpreted as its fitness value. (7) The internal LCS loop is restarted from step 1.

3 Cognitive Immune System Theory

In this section a short introduction to Cohen's ideas regarding the cognitive immune system is given. In particular, we focus on the relevant concepts for understanding our modelling approach presented in section 4. For in-depth information on Cohen's im-mune theory see [4].

The immune system is an embedded biological system that maintains the body and protects it from harmful influences. These factors may come from the outside or the inside of the organism and are summarized under the term antigens. While operating, the immune system incorporates the states of the body tissues and thus provides an immune response that is based on context. This differs from the classical point of view that the main purpose of the immune system is to defend the body by discrimi-nating between self and non-self and triggering a monospecific response (see Burnet's Clonal Selection Theory in [1]). Cohen characterizes the immune system as a cogni-tive system on the basis of three features: it is able to make decisions, it creates inter-nal images of its environment, and it is able to learn in a self-organized manner.

It is comprised of a multitude of specialized cells and organs. The immune organs divide the body into distinct compartments that are flown through by the immune cells. Certain compartments of the organism are responsible for producing immune cells, others for transporting them, and again others contain the manifold interactions between the cells. On their way through the tissues the cells gather information on the

A Computational Model for the Cognitive Immune System Theory 267

tissues' state, the presence of antigens, and the activity of other immune cells. This in-formation is available through molecular shapes that are presented by all antigens and tissue cells. All immune cells bear certain receptor molecules on their surfaces that enable them to bind other molecules complementarily. These recognized molecules are summarized as ligands. An important statement of Cohen's theory is that a recep-tor is able to bind more than one single ligand – a property referred to as degeneracy. The term affinity describes the specific binding energy between receptor and ligand that arises from their degree of molecular complementarity – the higher the affinity the higher the probability of a successful binding event. In addition, the cellular rec-ognition is also influenced by the concentration of the respective ligand. So a relative-ly low ligand concentration with high affinity already causes the receptor to bind, but the low affinity of a ligand can be compensated by a high concentration as well.

After completing this distributed recognition phase in the body tissues, the immune cells gather in the lymph nodes and exchange information by producing interaction molecules called cytokines. These molecules provide the context for the immune cells to react according to the state of the organism. So on the one hand, the cells react to observations in the tissues by producing certain cytokines, and on the other hand, they react to the reactions of other immune cells by recognizing their produced cytokines. Cohen terms this special interaction between immune cells co-respondence.

As a result of this second-order decision-making process an apt immune response is elicited. Specialized effector cells move back to the body tissues and execute the made immune decisions. So despite the degenerate perceptions of its individual cells, the immune system is able to produce a specific response as a whole. The three phases of immune cell activity (recognition, decision-making, execution) follow each other continuously and thus can be described as the functional loop of the immune system. In each phase spatial proximity is an important prerequisite for the interac-tions between immune cells and antigens and between immune cells alone.

By means of several evolutionary mechanisms the more or less randomly generated population of immune cells is structured by sorting out inefficient elements. Only those immune cells are selected and included in the repertoire that show an adequate level of recognition and response in regard to certain molecular inputs. The cells that undermine the immune system's body maintaining function are deleted. As a result of this self-organized adaptation process an immunocompetent cell repertoire emerges and the immune system learns to handle the antigenic influences efficiently.

4 Our Modelling Approach

In the following section, the two systems whose basic elements and mechanisms have been introduced before are brought together: we take Holland's classic LCS as a start-ing point for outlining a computational model of Cohen's immune theory. Due to the fact that some of the characteristic features of Cohen’s theory cannot be adequately modelled in a classic LCS, we modify the LCS with regard to Cohen’s immunological principles. We term this version of a LCS a Cognitive Immune System (CIS).

At first, the necessary concepts and representations that are used in the CIS model are introduced. The interactions between the CIS elements and the resulting

268 D. Voigt, H. Wirth, and W. Dilger

dynamics of the computational system are presented afterwards. Then, we describe the implementation of the evolutionary mechanisms that are used for adapting the system's behaviour. Finally, the whole CIS is summarized by presenting its full computation loop and comparing its constituent parts to the classic LCS form.

4.1 Representations

Firstly, we assume that the classifier population of the LCS is the counterpart to the immune cell repertoire. So an immune cell of the CIS is equivalent to a single classi-fier rule string. We introduce two necessary symbol sets: the set C that contains the

cytokines {c1, ..., cl} where +∈ l , and the set A that contains the actions {a1, ..., am}

where +∈m . The former describes the cytokines that potentially can be produced and sensed in the CIS, and the latter describes the actions that can be performed by the immune cells. Both sets are determined by the user of the system before starting and remain unchanged during the whole computation loop. It is for the user to decide which internal coding mechanism is applied to the elements of the sets – so for exam-ple, symbolic or numerical representations can be used here.

In order to model the signal processing chain of the CIS according to the principles of Cohen's theory, we modify the architecture of a classic LCS as follows: the mes-sage list is split into three distinct parts. So we obtain an antigen message list, a cyto-kine message list and an action message list. The main inputs of the immune system consist of antigens and cytokines that are produced by tissue cells. These molecules can be found in the CIS in form of messages. An antigen is represented by a string

from the alphabet {0, 1} that has the length +∈k – although other representations can be used here (e.g. symbolic). Such a binary antigen string can be interpreted as an abstract description of a molecular shape. It is placed by the detectors in the antigen message list during the computation loop. A tissue-produced cytokine is represented by an element of the set C, which has already been defined above. The internal repre-sentation of a cytokine is published by the detectors into the cytokine message list. So the two kinds of messages (antigens and tissue-produced cytokines) can be seen as counterparts to the external messages of the classic LCS. The outputs of the CIS con-sist of elements of the set A, that also has been defined above. These abstract descrip-tions of actions are published in the action message list.

In our modelling approach the different classes of immune cells are integrated into one hybrid immune cell type. A single CIS immune cell is made of five parts: an anti-

gen receptor string RAntigen ∈ {0, 1}k, where +∈k , that describes the cell's possible perceptions; a positive cytokine receptor RCytokine+ C∈ , that describes the type of cytokine that stimulates the cell; a negative cytokine receptor RCytokine– C∈ , that describes the type of cytokine that inhibits the cell; a cytokine response message MCytokine C∈ , that is published by the cell in case of activation; and an action message MAction A∈ , that is suggested by the cell for execution. In addition to the receptor and the response parts, each CIS immune cell is assigned a lifetime L ∈ , that gives an estimate about the respective cell's remaining lifetime in the repertoire.

ù

ù

ù

ù

ú

A Computational Model for the Cognitive Immune System Theory 269

4.2 Interactions

After the CIS detectors have placed the descriptions of the antigens and the tissue-produced cytokines in the corresponding message lists, a matching procedure bet-ween the elements of the immune repertoire and the available immune signals takes place. In correspondence with the matching mechanism of the LCS, the elements' mutual interaction potential is identified by comparing each element of the CIS popu-lation to each element of the antigen message list and to the elements of the cytokine message list. But the CIS matching procedure differs in important aspects from the one of the LCS which only considers completely satisfied rule strings and makes use of special wildcard symbols. Since this matching criterion is much too strict to meet with the degenerate antigenic perception of Cohen’s theory, we suggest a matching mechanism that is based on the concept of a binary Hamming shape-space (see [13]). For computing the antigen affinity AAntigen between an available antigen message MAntigen and the antigen receptor RAntigen of an arbitrary immune cell, we use:

⎪⎩

⎪⎨⎧ ≠

== ∑= else0

if1where,

1

1

iAntigen

iAntigeni

k

i

iAntigen

RM

kA αα

( iAntigenM and i

AntigenR refer to the thi bit of the respective string)

(1)

where k is the total number of bits in an antigen message or antigen receptor string. The sum yields the number of bits where the antigen message and the corresponding antigen receptor differ, and thus describes the strings' ability to match complementa-rily. The antigen affinity is subsequently obtained by normalizing the value to the interval [0, 1]. Since only a pair of identical binary strings has no affinity at all, there is a certain degenerate interaction potential between almost all elements of the im-mune repertoire and the antigen message list.

Furthermore, the matching mechanism yields the cytokine affinity ACytokine between all available cytokine messages and a single cytokine receptor RCytokine (either positive or negative) as follows:

⎪⎩

⎪⎨⎧ =

== ∑= else0

if1where,

1

1

Cytokinej

Cytokinejn

j

jCytokine

RM

nA ββ

( jCytokineM refers to the thj cytokine message of the list)

(2)

where n is the total amount of available cytokines in the cytokine message list. The matching procedure only takes into account whether a successful recognition event on part of the immune cell receptor takes place or not. Instead of testing the degenerate matching of complementary molecular shapes, as in the case of the antigen affinity, the cytokine matching procedure focusses on the recognition event itself. Hence, each of the two cytokine receptors of an immune cell matches only one cytokine message.

The overall affinity of an immune cell in regard to an actual immune situation is the result of integrating three partial affinities: firstly, the antigen affinity is computed for the immune cell's antigen receptor (using Eq. (1)), and secondly, the cytokine af-finities are computed for the immune cell's positive and negative cytokine receptors

270 D. Voigt, H. Wirth, and W. Dilger

respectively (using Eq. (2)). In case of an occurring empty message list, the corre-sponding partial affinity value is set to zero. On account of the degenerate matching procedure between antigens and antigen receptors, a selection from all possible anti-gen-receptor-pairs has to be made to determine the set of potentially active immune cells. Since the antigen concentration also influences the interaction potential of the pair, we suggest the use of an affinity-based roulette wheel selection like in the auc-tion round of the classic LCS. By means of this algorithm a single antigen message string from the antigen message list is selected. This antigen message is taken as the basis for computing the overall affinity AOverall according to:

⎪⎩

⎪⎨⎧ >

=

−+= −+

else0

0if2where

)(

xx

f(x)

AAAfA CytokineCytokineAntigenOverall

(3)

Thus, the term AAntigen describes a basic activity with regard to an antigen, that is stimulated or inhibited by the following terms ACytokine+ and ACytokine– respectively. The function f maps AOverall to the interval [0, 1]. The overall affinity is computed for each cell of the immune repertoire. The resulting values can be interpreted as the potential activation of the immune cells in regard to the current immune situation that is repre-sented by the elements of the message lists. In order to determine from the set of all immune cells the subset that actually becomes activated, in accordance to the LCS an immune auction round is held. This restriction of potential cell activity can be seen as a computational equivalent to the spatial distribution of the immune cells in the body. According to this, an immune cell only gets activated if it has the necessary proximity to the target molecule (under the assumption that there is sufficient affinity). There-fore, the auction round can be interpreted as the CIS counterpart to this anatomical restraint. In this selection process the immune cells place their respective overall af-finity AOverall as their bids. Subsequently, the set of auction winners – and thus the set of active immune cells – is obtained by turning the virtual roulette wheel. In addition to this, the lifetime of the immune cells that remain inactive in this round is reduced by a certain predefined amount. If the value of an immune cell's remaining lifetime falls below a certain threshold, it is removed from the repertoire. The purpose of this automatic reduction of lifetime is to increase the selection pressure in the immune cell population (see section 4.3).

As a result of their activation the immune cells produce co-respondence signals. These intercellular signals are modelled in the CIS as internal cytokine messages that are published by the active immune cells after completing the auction round. Hence, an intercellular cytokine message is the CIS counterpart to an internal LCS message. After the cytokine message list has been cleared, all activated immune cells publish their accompanying cytokine response messages in the cytokine message list. Thus, the immune cells jointly create a new cytokine pattern. In the course of the next CIS loop, this existing pattern is supplemented by the tissue-produced cytokines which are sensed through the system's detectors. Therefore, an updated cytokine context is

A Computational Model for the Cognitive Immune System Theory 271

created that is able to influence future CIS processing cycles. By means of the cytokine message list the immune cells exchange information among each other.

Besides that, the co-respondence process serves as an internal feedback mechanism within the immune repertoire. Through their two cytokine receptors the immune cells are able to receive positive and negative feedback at the same time. Therefore, a high affinity value alone is not a guarantee for leading to an immune cell's activation. In case of a simultaneously occurring inhibiting cytokine message in the cytokine mes-sage list, the overall affinity of the immune cell would be decreased by a certain amount (depending on the cytokine's concentration) and the activation probability of the immune cell would be reduced as well. The same applies to the situation where certain cytokines can compensate an immune cell's lack of antigen affinity and conse-quently cause an increase of the cell's activation probability. Even the presence of an-tigens is not a requirement – the activation of an immune cell can already be achieved by the mere influence of a sufficient amount of stimulating cytokines. Generally, in the CIS the immune cells' antigenic perceptions arise from the context of currently available cytokine messages.

According to Cohen, the co-respondence process is a joint effort between all active immune cells. As a result it yields an immune decision that influences the effector cells' subsequent actions. In the CIS this is modelled by the active immune cells that publish their respective action message in the corresponding message list and thus suggest immune actions. So the support for an arbitrary action is defined as the pro-portion of messages in the action message list that actually suggest this specific action as the next one to be performed by the effectors. In order to determine the next per-formed action, a support-based roulette wheel selection takes place. The action message that is selected by this mechanism is passed to the effectors of the CIS for execution in the tissues.

4.3 Evolutionary Mechanisms

In order to implement the evolutionary mechanisms of Cohen's immune theory, we complement the interaction procedures with an algorithmic component that models the bone marrow's function as the immune organ where immature immune cells are produced. The constituent parts of a new immune cell are produced as follows: as an analogy to the immune cells' ability to manufacture antigen receptors somatically, the CIS antigen receptor string is created randomly from the symbols of the binary alpha-bet; because of the genetic restrictions of the cytokine repertoire, the symbols that are assigned to the immune cell's two cytokine receptors and the respective cytokine re-sponse message are derived from the predefined set C (see section 4.1); accordingly, the action message is derived from the set A; the immune cell's lifetime value is set to a predefined constant.

In correspondence to the selection step that takes place in the thymus, all immature immune cells are subject to a testing mechanism that determines whether these cells are added to the immune repertoire or not. For this test we introduce the user-defined

set S of self messages {s1, ..., so} where s ∈ {0, 1}k and +∈ok, . These self mes-

sages describe a subset of the potential messages that is regarded to be part of the body's self, and thus can be interpreted as a filter that prevents certain new immune cells from being inserted into the repertoire. The immunocompetence of an immature

ù

272 D. Voigt, H. Wirth, and W. Dilger

immune cell is determined by comparing the antigen receptor of the cell to each self message and computing their mutual affinity (using Eq. (1)). According to Cohen’s immune theory, an immune cell is rejected if its antigen affinity is too high or too low. As a result of this selection step, only those cells survive that show a moderate affinity to the set of self messages.

The next selection step of the immune system is implemented by means of the CIS auction round that determines the immune cells that actually become activated. If the inhibiting effects of the cytokine messages are left aside for a moment, this selection mechanism particularly favours the immune cells that have a high antigen affinity. So the activated immune cells have the chance to prove their usefulness regarding the present immune situation and to increase their respective lifetimes by suggesting ap-propriate actions (see below). The inactive immune cells do not have this opportunity. Because of their low antigen affinity they cannot compete with the high affinity im-mune cells and as a result are displaced from the repertoire. This selection process is boosted by automatically decreasing the lifetime of inactive immune cells.

As an analogy to the affinity maturation, a subset of the activated immune cells is reproduced. This is realized by copying the constituent parts of an immune cell and thus obtaining a set of identical cell clones. While the daughter cells' respective anti-gen receptor string is mutated by inverting random position bits, the other parts of the immune cells remain unchanged. So the result of this cellular modification process is a set of immune cells that only differ in their antigen receptor string. Subsequently, the antigen affinities between the mutated daughter cells and the antigen message that caused their mother's activation are computed (according to Eq. (1)). By means of an affinity-based roulette wheel selection a single immune cell is selected from the set of the mother and the daughter cells. This immune cell replaces the mother cell in the immune repertoire; all other cell clones are rejected. As a result of this selection step the immune cells' antigenic perceptions may be improved.

The last selection step (which Cohen describes only allusively) is realized in the CIS through the workings of the classic LCS bucket brigade algorithm (for details see [11]). By means of the positive or negative feedback that is received from the tissues as a result of the performed immune actions, the bucket brigade adjusts the lifetime values of the cells in the immune repertoire. An immune cell that is able to gather suf-ficient amounts of lifetime can be seen as the CIS equivalent to a memory cell.

4.4 Summary

The functional components of the resulting CIS and its computation loop are sum-marized in Fig. 2. The first step of the CIS is to clear all three message lists and to produce an initial population of immune cells by means of the bone marrow and the thymus components. Apart from certain structural modifications, the sequence of the internal CIS computation steps is similar to the classic LCS: (1) The detectors place messages in the corresponding message lists that describe the observed state of the tissues. (2) Each element of the immune cell repertoire is compared to the available messages of the antigen and the cytokine message list. The respective affinities are computed. (3) These potential activities are taken as a basis for an auction round that selects the immune cells that actually become activated in this round. (4) An affinity maturation mechanism is applied to a certain subset of the active immune cells. In this

A Computational Model for the Cognitive Immune System Theory 273

evolutionary step the immune cells' antigenic perception may be improved by cell cloning, receptor mutation and affinity-based selection. (5) Furthermore, all activated immune cells publish their respective cytokine and action messages in the corre-sponding message lists.

AntigenMessage

List

ImmuneCell

Repertoire

Detectors Effectors

Tissues

CytokineMessage

List

ActionMessage

List

Antigens Cytokines Actions

AuctionRound

Input Output

BoneMarrowThymus

CellDeath

BucketBrigade

AffinityMaturation

Matching

1

25

Cytokines Actions

6

34

NewImmuneCells

97

Feed-back

Adjusts Lifetime

Inefficient Immune Cells8

Cognitive ImmuneSystem (CIS)

Fig. 2. The structure and internal mechanisms of the resulting CIS

(6) If there is sufficient action support, the effectors perform the suggested immune actions in the body tissues. On the one hand, from these executed actions arise new CIS inputs (see step 1), and on the other hand, the tissues react by delivering rein-forcement information to the system. (7) In order to improve the overall behaviour of the CIS this feedback is used by the bucket brigade algorithm to adapt the lifetime values of the cells in the immune repertoire. (8) As a result of the several evolutionary steps, certain inefficient immune cells are removed from the repertoire. (9) New im-mune cells are added to the repertoire and the computation loop continues at step 1.

A comparison between the resulting CIS and the classic LCS shows, that certain LCS concepts have been included almost unchanged (e.g. message list and bucket brigade algorithm), while others have been partially reworked (e.g. rule syntax and

274 D. Voigt, H. Wirth, and W. Dilger

Table 1. Corresponding parts of the classic LCS and the CIS

Learning Classifier System (LCS) Cognitive Immune System (CIS) environment tissues (input/output space)

message list respective message lists for antigens,

cytokines and actions

external message antigen message,

tissue-produced cytokine message internal message cell-produced cytokine message

classifier population immune cell repertoire classifier rule string immune cell (modified syntax)

classifier condition part antigen receptor,

cytokine receptors

classifier action part cytokine and immune

action response

classifier strength affinity (in auction round), lifetime (in bucket brigade)

rule chaining co-respondence

GA bone marrow, thymus, activation,

lifetime reduction, affinity maturation

matching procedure), and again others have been completely replaced by immune-inspired mechanisms (e.g. the GA). A selection of the corresponding parts of the two computational systems is shown in table 1.

5 Conclusion

In this paper, we have presented a modification of a classic LCS that can be used as a computational model of Cohen's Cognitive Immune System Theory. In order to lay the foundations for our modelling approach, a short introduction to the basic elements and mechanisms of the LCS has been given. Then, we have briefly summarized the characteristic features of Cohen's immune theory. After that, the representations of the immune agents, their mutual interactions and the evolutionary mechanisms that they are subject to have been presented. The resulting computational system CIS has been compared to the parts of a classic LCS.

As in the case of the classic LCS, the possible field of application for the CIS can be seen in the domain of machine learning and problem solving. In particular, the CIS can be used for all tasks that involve context-based processing of signals (e.g. recog-nition of noisy patterns). Because of its simple internal representations and mecha-nisms the system can be easily adapted to a wide variety of computational problems.

The CIS has been implemented and several test runs with benchmark data sets from the LCS domain have been conducted. The preliminary results of these simple learning tasks are promising. It could be shown that the CIS is able to learn patterns from a presented training set by adapting its immune cell repertoire to the underlying structures of the data. But more complex tests regarding the computational system’s learning capabilities remain to be done.

A Computational Model for the Cognitive Immune System Theory 275

References

1. Burnet, F.M.: The Clonal Selection Theory of Acquired Immunity. Cambridge University Press, Cambridge (1959)

2. Jerne, N.K.: Towards a Network Theory of the Immune System. Ann. Immunol. (Inst. Pasteur) 125C, 373–389 (1974)

3. Bibel, D.J.: Milestones in Immunology - A Historical Exploration. Science Tech Publ. Madison, Wis. (1988)

4. Cohen, I.R.: Tending Adam’s Garden: Evolving the Cognitive Immune Self. Elsevier Academic Press, Amsterdam (2000)

5. Andrews, P.S., Timmis, J.: Inspiration for the Next Generation of Artificial Immune Systems. In: Jacob, C., Pilat, M.L., Bentley, P.J., Timmis, J.I. (eds.) ICARIS 2005. LNCS, vol. 3627, pp. 126–138. Springer, Heidelberg (2005)

6. Farmer, J.D., Packard, N.H., Perelson, A.S.: The Immune System, Adaptation & Learning. Physica D 22, 187–204 (1986)

7. Hofmeyr, S.A., Forrest, S.: Architecture for an Artificial Immune System. Evolutionary Computation 8 (2000)

8. Vargas, P.A., de Castro, L.N., Michelan, R., Von Zuben, F.J.: An Immune Learning Classifier Network for Autonomous Navigation. In: Timmis, J., Bentley, P.J., Hart, E. (eds.) ICARIS 2003. LNCS, vol. 2787, pp. 69–80. Springer, Heidelberg (2003)

9. Vargas, P.A., de Castro, L.N., Von Zuben, F.J.: Mapping Artificial Immune Systems into Learning Classifier Systems. In: Lanzi, P.L., Stolzmann, W., Wilson, S.W. (eds.) Learning Classifier Systems. LNCS (LNAI), vol. 2661, pp. 163–186. Springer, Heidelberg (2003)

10. Holland, J.H., Reitman, J.S.: Cognitive Systems Based on Adaptive Algorithms. In: Waterman, D.A., Hayes-Roth, F. (eds.): Pattern-directed inference systems. Academic Press, New York (1978) (reprinted). In: Fogel, D.B. (ed.) Evolutionary Computation. The Fossil Record. IEEE Press, New York (1998)

11. Holland, J.: Adaptation in Natural and Artificial Systems. University of Michigan Press, Ann Arbor, 1975. Republished by the MIT Press (1992)

12. Nissen, V.: Einführung in Evolutionäre Algorithmen. Vieweg, Braunschweig (1997) 13. De Castro, L.N., Timmis, J.: Artificial Immune Systems: A New Computational

Intelligence Approach. Springer, London (2002)