Tutorial on Artificial Immune Systems

1
Tutorial on Artificial Immune Systems

• Dr Jon Timmis
• Department of Electronics and Department of
  Computer Science
• University of York, UK
• http//www-users.cs.york.ac.uk/jtimmis

2
Part 1 Introduction and basic immunology
3
Artificial Immune Systems A Definition

• AIS are adaptive systems inspired by theoretical
  immunology and observed immune functions,
  principles and models, which are applied to
  complex problem domains
• De Castro and Timmis,2002

4
Scope of AIS
20
10
Clustering/classification
Anomaly detection
Computer security
Learning
Optimisation
Bioinformatics
Web mining
Image proc.
Robotics
Control
0
5
From a computational perspective
Computational Properties
Systems that are

• Unique to individuals
• Distributed
• Imperfect Detection
• Anomaly Detection
• Learning/Adaptation
• Memory
• Feature Extraction
• Diverse
• ..and more

• Robust
• Scalable
• Flexible
• Exhibit graceful degradation
• Homeostatic

6
Thinking about AIS

• Biology
• Modelling
• The biology
• Abstraction
• General frameworks
• Algorithms
• Realisation in engineered systems
• We will look at each stage

7
A Conceptual Framework
DC activation, T-cell clonality
Bio-inspired algorithms
Stepney et al, 2005
8
Immunology
9
What is the Immune System ?
a complex system of cellular and molecular
components having the primary function of
distinguishing self from not self and defense
against foreign organisms or substances
(Dorland's Illustrated Medical Dictionary)
The immune system is a cognitive system whose
primary role is to provide body maintenance
(Cohen)
Immune system was evolutionary selected as a
consequence of its first and primordial function
to provide an ideal inter-cellular communication
pathway (Stewart)
10
What is the Immune System ?

• The are many different viewpoints
• These views are not mutually exclusive
• Lots of common ingredients

11
Classical Immunity

• The purpose of the immune system is defence
• Innate and acquired immunity
• Innate is the first line of defense. Germ line
  encoded (passed from parents) and is quite
  static (but not totally static)
• Adaptive (acquired). Somatic (cellular) and is
  acquired by the host over the life time. Very
  dynamic.
• These two interact and affect each other

12
Multiple layers of the immune system
13
Innate Immunity

• May take days to remove an infection, if it
  fails, then the adaptive response may take over
• Macrophages and neurophils are actors
• Bind to common (known) things. This knowledge
  has been evolved and passed from generation to
  generation.
• Other actors such as TLRs and dendritic cells
  (next lecture) are essential for recognition

14
Adaptive Immune System
15
A Lymph Node

• Goldsby et al., Immunology, 5th edition, 2003

16
Lymphocytes

• Carry antigen receptors that are specific
• They are produced in the bone marrow through
  random re-arrangement
• B and T Cells are the main actors of the adaptive
  immune system

17
B Cell Pattern Recognition

• B cells have receptors called antibodies
• The immune recognition is based on the
• complementarity between the binding region of
• the receptor and a portion of the antigen called
• the epitope.
• Recognition is not just by a single antibody,
• but a collection of them
• Learn not through a single agent, but
• multiple ones

18
B Cell Receptor

• V-region antigenic recognition and binding
• C-region effector functions
• Heavy and light chains
• VDJC immunoglobulin (encoded by genes)
• Genes direct the development of an individual.
  Encode proteins, which are made up of amino acids

19
Processes within the Immune System (very
basically)

• Negative Selection
• Censoring of T-cells in the thymus gland of
  T-cells that recognise self
• Defining normal system behavior
• Clonal Selection
• Proliferation and differentiation of cells when
  they have recognised something
• Generalise and learn
• Self vs Non-Self

20
Clonal Selection
21
Clonal Selection
22
Clonal Selection

• Each lymphocyte bears a single type of receptor
  with a unique specificity
• Interaction between a foreign molecule and a
  lymphocyte receptor capable of binding that
  molecule with high affinity leads to lymphocyte
  activation
• Effector cells derived from an activated
  lymphocyte bear receptors identical to those of
  parent cells
• Lymphocytes bearing self molecules are deleted at
  an early stage

De Castro and Timmis,2002
23
History of Immune Models
24
Immune Responses
25
Affinity Maturation

• Responses mediated by T cells improve with
  experience
• Mutation on receptors (hypermutation and receptor
  editing)
• During the clonal expansion, mutation can lead to
  increased affinity, these new ones are selected
  to enter a pool of memory cells
• Can also lead to bad ones and these are deleted

26
Summary

• Innate and adaptive immunity
• Focused on adaptive here
• Lymphocytes
• Negative selection
• Clonal selection
• Immune memory and learning

27
Part 2 Further Immunology and Modelling
28
What is the Immune System ?

• The are many different viewpoints
• These views are not mutually exclusive
• Lots of common ingredients

29
Problems with the classical view

• What happens if self changes?
• What about things that are not harmful
• Babies, food, lots of stuff around us.
• The Danger model was proposed

30
Danger theory (Matzinger 1994)

• it is not non-self, but danger that the IS
  recognises
• dangerous invaders cause cell death or stress
• these cells generate danger signal molecules
• unlike natural cell death
• these stimulate an immune response local to the
  danger
• to identify the culprit

31
Danger Theory
4

1. Antigen enters system and into tissue
2. Tissue affected and necrosis causes danger
   signals
3. Signals located by DCs which interact T-cells
4. Adaptive response initiated against antigen

Bad antigens (self or non-self)
Lymphocytes
3
1
2
DCs
Tissues
Adapted from Zhang (2007)
32
What is the difference?
Matzinger, P. Science 296 (301-305). 2002.
33
Self-Assertion

• Basic premise is that the immune system is
  complete unto itself
• There is no difference between an antigen and
  antibody and any node of the network can bind and
  be bound by others
• The immune system is not a defense system, but a
  maintenance system

34
Immune Network Theory

• Idiotypic network (Jerne, 1974)
• B cells co-stimulate each other
• Treat each other a bit like antigens
• Creates an immunological memory

35
Shape Space Formalism

• Repertoire of the immune system is complete
  (Perelson, 1989)
• Extensive regions of complementarity
• Some threshold of recognition
• Pattern matching and data reduction

V

V
e
e
V
e
e




V
e
e


36
Self-Assertion

• The immune system builds up an internal image of
  the world
• Argues that danger is just another word for self
• The immune system still needs to recognise what
  is wrong (dichotomy still exists)
• Here they say the immune system always acts the
  same, just depends on the context
• Same external impact can make the system behave
  is two different ways depending on when it occurs

37
Self Assertion

• Recall from the clonal selection a one-to-one
  mapping
• However, in this less classical view, all cells
  bind to all cells.
• there is no essential difference between the
  recognized and the recognizer, since any
  given antibody might serve either, or both,
  functions. Immune regulation is based on the
  reactivity of antibody with its own repertoire
  forming a set of self-reactive, self-reflective,
  self-defining immune activities. Bersini, 2002

38
Emergence of Self
Bersini, 2002
39
Modelling the Immune System
40
UML

• Simple modelling language that allow us to
  capture
• Class information (and relationships between
  objects)
• State
• Transitions
• And many more

41
Relationships

• the antigen-specific activation of these
  effector T cells is aided by co-receptors on the
  T-cell surface that distinguish between two
  classes of MHC molecule cytotoxic cells express
  CD8 co-receptor, which binds MHC class I
  molecules, whereas MHC class II molecules
  specific T cells express the CD4 co-receptor,
  which has specificity for MHC Class II molecules
  Janeway

42
Relationships
Bersini, 2006
43
State Chart - Clonal Selection
Bersini, 2006
44
Sequence Diagram
Bersini, 2006
45
Summary

• All is not self
• An initial look at modeling the immune system
  using UML

46
Part 3 The Cognitive Model of the Immune System
47
Cognitive immunology?

• Stage 1 Cohen's view of the Immune System
• Maintenance through Cognition
• Stage 2 Applying the conceptual framework
  approach
• Computational model of receptor degeneracy
• First stage to design a Cohen inspired AIS

48
Views of the Immune System

• Classical immunology
• A top down / reductionist view
• Matzinger
• Varela
• Cohen
• A bottom-up / complex systems view

49
Definitions

• Some useful definitions
• Immune cells T cells, B cells and APC
• Receptors proteins on the surface of cells that
  can bind to other molecules
• Antigen any molecule that can bind to the
  unique receptors of T and B cells can be self
  or non-self
• Immune molecules any molecule used by immune
  cells for immune purposes

50
Cohen Overview (1)

• Immune System
• Complex, reactive and adaptive system
• Carries out body maintenance
• Operates via cognitive strategy similar to brain
• T immune system as a black box
• Input molecular shapes sensed by immune cell
  receptors
• Output inflammation

51
Cohen Overview (2)

• Inflammation
• Range of processes e.g. cell growth, replication,
  death, movements and differentiation
• Results in body maintenance
• Body Maintenance
• Detect state of body tissues and elicit
  appropriate response to keep body fit
• Defence against pathogen as a special case

52
Immune Specificity (1)

• Specificity refers to the ability of the immune
  system to distinguish
• Antigen receptors of immune cells are degenerate
• A receptor can recognise different shapes
• No way of making sure only one immune cell has a
  receptor for a particular antigen
• Specificity thus not simply a function of
  individual cells with a 'specific' receptors

53
Immune Specificity (2)

• Cohen Specificity
• Diagnosis of varied body conditions and the
  production of an appropriate specific
  inflammatory response
• Emerges due to two properties
• Co-respondence
• Patterns of elements

54
Co-Respondence

• Immune cells respond to different aspects of
  target
• Immune cells respond to each other
• Immune dialogue through immune molecules
• Picture of target emerges from co-operation
• Immune agents form networks
• Feedback both positive and negative
• Integration of innate and adaptive immunity

55
Patterns of Elements

• Complex arrangements of immune agent populations
• Individual non-specific to target, pattern is
  unique
• Overlapping reactions of degenerate receptors
• Example The Eye
• We possess millions of colour receptors, of which
  there are only 3 highly degenerate types
• The brain can, however, distinguish thousands of
  different colours

56
A Cognitive Strategy (1)

• A way of adjusting to the environment
• Both the brain and immune system are cognitive
  systems
• Cognitive systems utilise 3 elements
• An internal history
• Self-organisation
• Deterministic decisions
• Internal History
• Build internal images that map the environment in
  which they exist

57
A Cognitive Strategy (2)

• Self-Organisation
• Learning and acquisition of memory
• Deterministic Decisions
• System must have different options available to
  it and an internal history
• Decisions emerge from match of environmental
  conditions with an internal image
• Elements of internal history so complex, the
  choice appears unpredictable

58
Immunologists Disagree (1)

• There is an obvious and dangerous potential
• for the immune system to kill its host but
• it is equally obvious that the best minds in
  immunology are far from agreement on
• how the immune system manages to avoid
• this problem
• Langman, R. E. and Cohn, M., Editorial Summary,
  Seminars in Immunology, vol. 12, pp. 343-344,
  2000

59
Immunologists Disagree (2)

• All is not lost
• Alternative immune ideas for AIS
• Explore different computational properties
• Different theories may require different
  modelling / algorithmic techniques
• AIS dont have to be based on the correct
  immune theory

60
Stage Two

• A Computational Model of
• Immune Receptor Degeneracy
• Applying The Conceptual Framework Approach

61
Motivation

• Explore and exploit previously unused immune
  ideas/theories to inspire new and unique AIS
• Cohen's view of the immune system as a complex
  adaptive system for body maintenance
• Utilise the Conceptual Framework Approach as a
  methodology for developing such an AIS
• Investigate biological processes free of
  engineering application bias before an algorithm
  is built

62
Conceptual Framework Approach
63
Model Overview (1)

• Investigate issues surrounding the degeneracy of
  antigen receptors and how this might affect AIS
• A model of lymph node paracortex and the
  interactions between T helper cells and APCs
• Naive T helper cell activation by APCs is the
  initial recognition event of adaptive immune
  response
• Activated T helper cells required for all
  adaptive immune responses
• Assume the T helper cell receptors are degenerate

64
Model Overview (2)

• The model is an abstract computational model
• Modelling stage of Conceptual Framework Approach
• Are there benefits to degenerate detectors?
• What issues do degenerate detectors cause?
• How might these be overcome?
• Envisage an AIS where recognition will emerge
  from the collective response of a set of
  detectors
• See Cohens Patterns of Response

65
Why Degeneracy?

• Present at all levels of biological organisation
• Genetic code, body movements, language
• Powerful property
• Previous example of the eye
• Degenerate systems highly adaptable to changes in
  environment
• Useful generalisation property for recognition?

66
A Lymph Node

• Goldsby et al., Immunology, 5th edition, 2003

67
Lymph Nodes

• Split into 3 concentric regions cortex,
  paracortex and medulla
• Each supports a different environment
• Paracortex supports naive T helpers and APCs
• Segmentation due to presence of signalling
  molecules (chemokines)
• Cells with receptors migrate towards
  concentration
• Naive T helpers and activated APCs have receptor
  for a chemokine only produced in paracortex area

68
Lymph Node Model Overview

• 2 overlapping cellular layers
• Chemical space models chemokine concentration
  produced in paracortex
• Agent space cells can contain one of 3
  specialised immune agents based on APCs, foreign
  antigen and T helper cells
• Chemical space used by agents in movement rule
• Layers are 2-D grids of cells sharing co-ordinate
  system and dimensions
• Model updates at discrete time steps

69
Chemical Space

• A user defined area in the centre of chemical
  space is set to be paracortex region
• Cells in this region mimic chemokine production
• Cells updated at each time step according to a
  chemical diffusion rule
• Overall chemical concentration smooths but leaves
  local randomness
• Stable chemical gradient emerges from centre of
  paracortex region to boundary of chemical space

70
Chemical Space Behaviour

• Initialisation 100 iterations

71
Agent Space

• Antigen agents have a shape represented as a
  string of bits
• APC agents can ingest antigen agents and present
  it to T helper agents
• T helper agents can be either naive or activated
  and have a unique bit string receptor
• At each time step agents move
• Interacts between agents occurs when they are in
  proximity and depends on agent type and state

72
Agent Space Behaviour

• Initialisation 40 iterations

73
Simulation Observations

• At the end of a simulation run, receptors of
  activated T helper agents can be analysed and
  compared to the antigen string
• Dependent on choice of suitable parameters,
  different antigen strings invoke activation of a
  different (and mostly unique) set of T helper
  cells
• Different numbers of T helper cells may become
  activated to different antigen strings

74
Computational Lessons

• Based on simple immune agent details we can build
  a model expressing emergent population behaviours
• A computational model is easily open to
  experimentation and analysis
• An AIS with degenerate receptors must
• Be able to distinguish in some way between
  receptor sets to be useful
• Cope with highly reactive or inactive receptors

75
Part 4 Framework, Representation and Algorithms
76
A Framework for AIS
Solution
Algorithms
Shape-Space Binary Integer Real-valued
Symbolic
Affinity
AIS
Representation
Application
De Castro and Timmis, 2002
77
A Framework for AIS
Solution
Algorithms
Euclidean Manhattan Hamming
Affinity
AIS
Representation
Application
78
A Framework for AIS
Solution
Bone Marrow Models Clonal Selection Negative
Selection Positive Selection Immune Network
Models
Algorithms
Affinity
AIS
Representation
Application
79
Shape-Space

• An antibody can recognise any antigen whose
  complement lies within a small surrounding region
  of width ?(the cross-reactivity threshold)
• This results in a volume ve known as the
  recognition region of the antibody

V
ve
?
?
S
ve
?
ve
Perelson,1989
The Representation Layer
80
Affinity Layer

• Computationally, the degree of interaction of an
  antibody-antigen or antibody-antibody can be
  evaluated by a distance or affinity measure
• The choice of affinity measure is crucial
• It alters the shape-space topology
• It will introduce an inductive bias into the
  algorithm
• It needs to take into account the data-set used
  and the problem you are trying to solve

The Affinity Layer
81
Affinity

• Affinity through shape similarity. On the left, a
  region where all antigens present the same
  affinity with the given antibody. On the right,
  antigens in the region b have a higher affinity
  than those in a

The Affinity Layer
82
Hamming Shape Space

• 1 if Abi ! Agi 0 otherwise (XOR operator)

The Affinity Layer
83
Hamming Shape Space

• (a) Hamming distance
• (b) r-contigous bits rule

The Affinity Layer
84
Mutation - Binary

• Single point mutation
• Multi-point mutation

85
Affinity Proportional Mutation

• Affinity maturation is controlled
• Proportional to antigenic affinity
• ?(D) exp(-?D)
• ? mutation rate
• D affinity
• ? control parameter

86
The Algorithms Layer

• Bone Marrow models (Hightower, Oprea, Kim)
• Clonal Selection
• Clonalg(De Castro), B-Cell (Kelsey)
• Negative Selection
• Forrest, Dasgputa,Kim,.
• Network Models
• Continuous modelsJerne,Farmer
• Discrete models RAIN (Timmis), AiNET (De Castro)

The Algorithms Layer
87
Clonal Selection CLONALG

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Cycle

The Algorithms Layer
88
Clonalg

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Cycle

• Create a random population of individuals (P)

The Algorithms Layer
89
Clonalg

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Cycle

• For each antigenic pattern in the data-set S do

The Algorithms Layer
90
Clonal Selection

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Cycle

• Present it to the population P and determine its
  affinity with each element of the population

The Algorithms Layer
91
Clonal Selection

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Cycle

• Select n highest affinity elements of P
• Generate clones proportional to their affinity
  with the antigen
• (higher affinitymore clones)

The Algorithms Layer
92
Clonal Selection

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Cycle

• Mutate each clone
• High affinitylow mutation rate and vice-versa
• Add mutated individuals to population P
• Reselect best individual to be kept as memory m
  of the antigen presented

The Algorithms Layer
93
Clonal Selection

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Cycle

• Replace a number r of individuals with low
  affinity with randomly generated new ones

The Algorithms Layer
94
Clonal Selection

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Cycle

• Repeat step 2 until a certain stopping criteria
  is met

The Algorithms Layer
95
Naive Application of Clonal Selection

• Generate a set of detectors capable of
  identifying simple digits
• Represented as a simple bitmap

96
Representation

• Each individual is a bitstring
• Use hamming distance as affinity metric

97
Evolution of Detectors

• Clones
• Mutated clones

98
A Slight Aside
99
Inductive Bias

• Can affect choice of representation and affinity
  functions
• Emma will talk more about this in the context of
  immune networks
• It is any (explicit or implicit) bias favouring
  one hypothesis over another
• All learning algorithms have it, otherwise it
  would only perform rote learning
• Bias is domain dependant
• Therefore, can not always say A is better than B

100
Inductive Bias

• Representation Bias
• Associated with the knowledge representation of
  the algorithm
• Preference Bias
• Associated with the evaluation function employed
• Use distance to measure affinity (this is
  important for AIS in particular)

101
Choice of Representation

• Assume the general case
• Ab  ?Ab1, Ab2, ..., AbL?
• Ag  ?Ag1, Ag2, ..., AgL?
• Binary representation
• Matching by bits
• Continuous (numeric)
• Real or Integer, typically Euclidian
• Categorical (nominal)
• E.g female or male of the attribute Gender.
  Typically no notion of order

102
Choice of Representation (2)

• Hybrid
• Possible and desirable, as is quite common to
  have data of different types at the same time
• More common in data mining literature, not really
  done (or appreciated) in AIS literature
• Typically, AIS seem to adapt the data to fit
  their particular algorithm rather than adapting
  the algorithm to making it more specific
• From a problem orientated point of view, this
  could lead to throwing away useful data without
  realising it!

103
Choice of Affinity Functions

• Choice of function should take into account the
  data being mined as they will all have a bias
• Binary Representation
• Typically employ Hamming or r-contiguous rule
• Argued that r-contiguous is more biologically
  plausible, therefore, use it not so.
• This assumes an ordering within the data that may
  not exist and will introduce a positional bias
• In the data mining, quite common not to have
  unordered sets, representing the data when doing
  classification.
• Therefore, a measure that takes into account
  position is not needed.

104
Choice of Affinity Functions (2)

• Continuous Representation
• Vast majority of AIS use Euclidean .. Because ?
• Also is Manhattan. They will produce different
  results .. They have different inductive biases
  and are more effective for different data sets
• Dist(Ab, Ag) ( ? (Abi Agi)2 )1/2
  (Euclidan)
• Dist(Ab, Ag) ? Abi Agi (Mahantten)
  
  
  
• How do they differ?
  

105
Differences

• Which of the two antibodies is closer?
• Euc. Man.
• Ab1 5.66 8
• Ab2 6.08 7
• It depends ..

Ab1
4
Ab2
1
Ag
4
6
Freitas and Timmis, 2003
106
Why?

• Euclidean is more sensitive to noisy data
• A single error in the coordinate of a vector
  could be seriously amplified by the metric
• Manhattan is more robust to noisy data and the
  differences tend not to be amplified
• So, results will be different and computational
  complexity is also different
• A rationale behind the choice is needed.

107
Negative Selection Algorithms

• Define Self as a normal pattern of activity or
  stable behavior of a system/process
• A collection of logically split segments
  (equal-size) of pattern sequence.
• Represent the collection as a multiset S of
  strings of length l over a finite alphabet.
• Generate a set R of detectors, each of which
  fails to match any string in S.
• Monitor new observations (of S) for changes by
  continually testing the detectors matching
  against representatives of S. If any detector
  ever matches, a change ( or deviation) must have
  occurred in system behavior.

The Algorithms Layer
108
Illustration of NS Algorithm
Match 1011 1000
Dont Match 1011 1101
r2
The Algorithms Layer
109
Classic Application of Negative Selection

• Domain of computer security
• Concept of self/non-self recognition
• Use of negative selection process to produce a
  set of detectors
• T-cells and their co-stimulation
• Antibody/antigen binding

110
Computer security Mapping from IS to AIS
111
Computer security Engineering the AIS

• Self defined as normally occurring connections
  over time (depends on frequency) in Hamming shape
  space.
• Needs to learn to distinguish between
  self/non-self
• Define self-set
• dynamically
• Detector produced via
• negative selection has
• tolerisation period

112
Network security Problem description

• Local area broadcast LAN
• Need to protect LAN from unwanted intrusions
• AIS needs to monitor traffic and flag intrusions
• Connection is a data path triple
• ltsource, destination, portgt

113
Computer security Performance evaluation

• Reported by Hofmeyr Forrest, 2000
• Tested on subnetwork of 50 computers
• Data collected over 50 days, 2.3 million TCP
  connections, filtered down to 1.5 mill.
• Bitstrings of length 49 (3900 unique)
• Non self set only 7 intrusions
• Low false positive rate (2 per day)
• Work by Kim Bentley, Stibor et al has shown
  some significant problems with negative selection

114
Immune Networks(self-assertion)
115
aiNET

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Clonal suppression
• Network interactions
• Network suppression
• Diversity
• Cycle

• Create a random population of individuals (P)

De Castro and Von Zuben, 2000
The Algorithms Layer
116
aiNET

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Clonal suppression
• Network interactions
• Network suppression
• Diversity
• Cycle

• For each antigenic pattern in the data-set S do

The Algorithms Layer
117
aiNET

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Clonal suppression
• Network interactions
• Network suppression
• Diversity
• Cycle

• Present it to the population P and determine its
  affinity with each element of the population

The Algorithms Layer
118
aiNET

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Clonal suppression
• Network interactions
• Network suppression
• Diversity
• Cycle

• Select n highest affinity elements of P
• Generate clones proportional to their affinity
  with the antigen
• (higher affinitymore clones)

The Algorithms Layer
119
aiNET

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Clonal suppression
• Network interactions
• Network suppression
• Diversity
• Cycle

• Mutate each clone
• High affinitylow mutation rate and vice-versa
• Select h highest affinity cells and place into
  memory set

The Algorithms Layer
120
aiNET

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Clonal suppression
• Network interactions
• Network suppression
• Diversity
• Cycle

• Eliminate all memory clones whose affinity with
  the antigen is less than a predefined threshold

The Algorithms Layer
121
aiNET

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Clonal suppression
• Network interactions
• Network suppression
• Diversity
• Cycle

• Determine similarity between each pair of network
  antibodies

The Algorithms Layer
122
aiNET

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Clonal suppression
• Network interactions
• Network suppression
• Diversity
• Cycle

• Eliminate all network antibodies whose affinity
  is less than a pre-defined threshold

The Algorithms Layer
123
aiNET

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Clonal suppression
• Network interactions
• Network suppression
• Diversity
• Cycle

• Introduce a random number of new antibodies into P

The Algorithms Layer
124
aiNET

• Initialisation
• Antigenic presentation
• Affinity evaluation
• Clonal selection and expansion
• Affinity maturation
• Metadynamics
• Clonal suppression
• Network interactions
• Network suppression
• Diversity
• Cycle

• Repeat 2 - 5 for a pre-defined number of
  iterations

The Algorithms Layer
125
aiNET on Data Mining
Training Pattern

• Limited visualisation
• Interpret via MST or dendrogram
• Compression rate of 81
• Successfully identifies the clusters

Result immune network
126
aiNET on multimodal optimisation
Initial population
Final population
127
Results Multi Function
aiNET
CLONALG
128
Dynamic Immune Networks

• Neal 2002
• Continous version of an immune network
• Notion of resources to control population
  (distributed to each node)
• Mortality rate for each node

129
Dynamic Immune Networks
130
Summary

• AIS Framework
• Clonal selection
• Negative selection
• Nice idea, but there are some issues
• Immune networks
• Static
• Dynamic

131
Part 5 Clonal Selection Algorithms for Learning
132
Quick Primer

• Supervised machine learning
• When you know the class of an instance before you
  start
• Wish to build a model of that data so you can
  then classify instances you have not seen before
• Many approaches including
• Neural networks, rule induction, Bayesian, ILP

133
Learning in the Immune System

• Recall the learning and memory capabilities in
  the clonal selection process
• Exploit this local v global search in a learning
  context
• Evolve a set of detectors that can generalise
  well enough to classify unseen data items

134
AIRS Algorithm Watkins, 2003

• Artificial Immune Recognition System (AIRS)
• Clonal selection based
• Uses the concepts of ARBs (Artificial
  Recognition Balls)
• Resource based competition for survival in order
  to control the population
• One-shot learning system

135
Memory Cell Identification
A
Memory Cell Pool
136
MCmatch Found
A
1
Memory Cell Pool
MCmatch
137
ARB Generation
A
1
Memory Cell Pool
MCmatch
Mutated Offspring
2
138
Exposure of ARBs to Antigen
A
1
Memory Cell Pool
MCmatch
Mutated Offspring
3
2
139
Development of a Candidate Memory Cell
A
1
Memory Cell Pool
MCmatch
Mutated Offspring
3
2
140
Comparison of MCcandidate and MCmatch
A
1
Memory Cell Pool
MCmatch
A
4
Mutated Offspring
3
2
MC candidate
141
Memory Cell Introduction
A
1
Memory Cell Pool
MCmatch
A
4
5
Mutated Offspring
3
2
MCcandidate
142
Results
Classification Accuracy
Iris 96.0 ?1.9
Ionosphere 95.6 ?1.7
Diabetes 74.2 ?4.4
Sonar 84.9 ?9.1
143
Classification Accuracy

• Important to maintain accuracy
• Can gain speed up through parrellisation

Iris 96
Diabetes 74.1
Sonar 84.0
Ionosphere 94.9
144
Dynamic Learning

• AISEC Email Classification

145
Web Mining

• Linoff and Berry (2001) the process of
  extracting useful information from the text,
  images and other forms of content that make up
  the pages
• Reasons why AIS might be useful?

146
Continuous Learning

• Used when you want to classify changes over time
  (the notion of what is in a class)
• Levels of what you are interested in may change
  over time or the context of where you are working
  or what you are doing
• Web content mining is a perfect testbed for these
  ideas
• This study looked at email classification of
  interesting v un-interesting

147
AISEC

• Supervised classification algorithm
• E-mail classified as interesting and
  uninteresting
• Uses constant feedback from user
• Capable of continuous adaptation
• This tracks concept drift and can also handle
  concept shift
• Representation Subject, Sender and Return
  address (based on existing literature, this is
  all you really need)
• Affinity measure Proportion of words found in
  one cell compared to another (very naive)
• Mutation Keep a library of words to use

148
AISEC Design Decisions

• Representation of one data class
• B cells represent uninteresting emails
• Gene Libraries
• Two libraries of words (subject, sender) used in
  mutation to create a new B cell
• Cloning
• Random clones are not produced (make meaningless
  B cells)

149
AISEC Design

• Co-stimulation
• A B cell classifies an email as un-interesting.
  Those emails are moved to a temp. store. B Cell
  rewarded if a successful classification
• Two recognition regions
• One for training, one for classification
• Cell death
• If B cells remains unstimulated, it dies.

150
The algorithm - classification

1. System is initialised with known uninteresting
   e-mail

Memory cells
Naive cells
2. E-mail presented for classification.
Classified as uninteresting as it stimulates
close cells
151
The algorithm correct classification

1. Highly stimulated cell reproduces n-times. Less
   stimulated cell produces fewer clones but with
   higher mutation rate

Stimulation Region
Classification Region
4. Cell with highest affinity is known to be
useful therefore rewarded by becoming memory cell.
152
The algorithm cont

• Incorrect classification
• Any cell responsible for incorrect classification
  is removed (memory or otherwise)

• Cell removal
• Aged naïve cells deleted. Memory cells placed in
  already covered areas also deleted.

153
Results

• 2268 e-mails (742 uninteresting) received over 6
  months
• E-mails presented in chronological order
• Feedback given after EVERY classification
• AISEC run 10 times
• C5.0, neural network run in Clementine data
  mining package
• Bayesian algorithm used feedback to update like
  AISEC
• Traditional Learning Continuous Learning

C5.0 83.9
Naïve Bayesian 85.0
Neural Network 85.6
AISEC 86

Naïve Bayesian 88.05
AISEC 89.05(.97)
154
Results
155
Results (cont)

• Standard measures of quality
• Precision is the proportion of positive documents
  retrieved compared with the total number of
  positive documents
• Recall is the proportion of positive documents
  actually classified as positive

Precision Recall
Naïve Bayesian 93.93 67.76
AISEC 82.2 (2.63) 81.13(4.71)
156
Summary

• Explored the use of clonal selection ideas in
  learning
• Static and dynamic
• Clonal selection is very popular base for immune
  algorithms at the moment.

157
Part 6 Homeostasis (immune, neural and endocrine)
158
Physiology

• Study of how living organisms work (vanders
  human physiology)
• Many levels from cell, to system wide
  interactions
• Cells
• Tissues
• Organs

159
Cells

• Human development starts from a single cell
• This differentiates into cells for various
  function
• All cells have the same genes, but how does a
  cell know what to do and what to become?
• Muscle cells, Nerve cells, Epithelial cells
  (secretion and absorption), Connective cells

160
Tissues

• Form when cells associate with other cells
• Is an aggregate of a single type of specific cell
  (same type of tissues as cells), but is commonly
  used to describe more general things (e.g. liver
  or lung tissue)

161
Organs

• Composed of four kinds of tissue arranged in
  various proportions and patterns
• Organ systems
• Urinary system contains bladder, kidneys, tubes)

162
Homeostasis

• state of reasonably stable balance between
  physiological variables
• Not good enough, as some variables e.g. blood
  sugar levels, undergo dramatic swings during the
  day but they are still in balance
• Homeostasis is a dynamic process
• There is a baseline to which things are returned,
  but things vary within a normal range

163
Homeostasis

• time averaged means
• Homeostasis must be described differently for
  each variable
• A person may be homeostatic for one thing, but
  not another
• But many diseases can be classified as falling
  out of homeostasis for one or more variable

164
Homeostatic control Steady State
Room temperature
Heat loss from body
Body temperature
(Bodys responses)
Constriction of skin blood vessels
shivering
Curling up
Heat production
Heat loss from body
Return of body temperature toward original value
Vanders human physiology
165
Rhythms

• We just looked at a corrective response (what
  happens when the steady state has been perturbed)
• Circadian rhythm anticipatory in nature

166
The Endocrine System

• Responsible for
• Production and control of hormones
• Hormones are a chemical substance that has a
  specific regulatory effect on the cells upon
  which they act
• Therefore, they can affect behaviour
• They are produced not only by the endocrine
  system, but also the neural and immune system
• Production of hormone is linked to changes in
  state of the organism

167
Endocrine System

• Organism responds to changes to internal and
  external factors
• Endocrine system tries to maintain homeostasis
  through the secretion of various hormones to act
  in response to a change in physiological state

168
Glands of the Endocrine System
De Castro and Timmis, 2002
169
Glands and their Role
De Castro and Timmis, 2002
170
Endocrine System

• Usually a number of hormones in the body at any
  one time
• Binding takes place between hormones and cells
  (at the receptor level)
• Receptors are located either within the cell
  nucleus or the plasma membrane (this depends on
  type of hormone the receptor is for)
• Many may bind at the same time, giving rise to a
  complex interaction of many components!!
• Hormones typically decay over time
• Minutes or days

171
Binding of hormones
Here we have a steroid hormone which will bind
in plasma membrane
Vander et al, 1990
172
Endocrine response to stimulus

• 1. Glands and nerve cells signal endocrine glands
  in response to stimuli such as temperature
  changes, hunger, fear, and growth needs.
• 2. In response, endocrine glands release hormones
  that carry instructions to specific cells. These
  hormones travel all around the body or just to
  neighbor cells looking for special binding
  proteins, the receptors, which are located in and
  on the target cells.
• 3. Once bound, the receptor interprets the
  hormones message and carries out its
  instructions by starting one of two distinct
  cellular processes. The receptor can
• Turn on genes to make new proteins, which causes
  long-term effects such as growth or sexual and
  reproductive maturity
• Alter the activity of existing cellular proteins,
  which produce rapid responses such as a faster
  heart beat and varied blood sugar levels.

173
Immune-neural-endocrine
174
Immune-neural-endocrine
175
Going Artificial
ANN AES AIS
(1) Neuron Endocrine gland Lymphocyte
(2) Network topology Hormone interactions Affinity measures
(3) Learning algorithms Hormone structure update Immune algorithms
176
ANN

• Very simple analogue of the neural system

x u ? wi . xi i0
f(u) 1 1 e-u
177
AES

• Aim is to provide a long term regulatory control
  mechanism for the behaviour of the system
• AES consists of gland cells which secrete
  hormones in response to an external stimuli to a
  value r for a gland g. ?g rate of release for the
  hormone

x rg ?g ? xi i
0
c(t 1)g c(t)g . ?
Release mechanism
Decay mechanism
178
AES

• Membrane receptors in neurons are sensitive to
  hormones
• Gland cells secrete and record level of hormone
• Gland cells secrete specific hormones
• In the system, neuron membranes have a list of
  hormones they are sensitive too

• Different cells react to different hormones in
  different ways
• Varies according to
• Detected hormone
• Concentration
• Type of receiving cell
• Cell make up

179
AES-ANN

• x
  g
• u ? wi . xi ? Cj . Sij . Mij
• i0
  j 0

Mij 1 1 dis(i,j)
180
Target Application

• Active media robot
• 16 sonar sensors, capable of detecting 5 meters
• Sonar link to a PC (robot runs Linux)
• Can we explore a certain space?

181
Experimental Procedure

• Simple ANN links sensors to motors
• Generates simple object avoidance
• Weights chosen manually and adjusted after
  experimentation (no learning as yet)
• Bias node to ensure forward motion
• Resultant behaviour was wandering and avoidance
• Not good when environment became more cluttered
• Robot came very close and collided with objects
• Large regions left unexplored

182
Experiments (AES-ANN)

• Same ANN structure
• Single hormone gland (fixed in content)
• Neurons receive input from sensors (as does the
  gland)
• Gland inhibits or excites neurons

183
Experiments

• Method
• 5 runs for 22 minutes
• Three different controllers
• Pure ANN (our control)
• Fixed level of hormone ANN
• Variable level of hormone ANN
• Aims
• Higher hormone levels result in a more
  expeditious retreat from obstacles
• Variable hormone release effective in allow close
  approaches and the above

184
Results (Proximity)
185
Results (Hormone Levels)
Variable hormone
Fixed hormone
186
Emotional Response??

• Informal experiments
• Trap it and it looks for a way out
• We tend to assign emotional responses to things
• Gets scared ??
• Hormone levels go through the roof and it will
  calm down

187
Further results

• Various hormones for behavior
• White seeking, black seeking, caution promoting,
  tiredness,

Neal and Timmis, 2005
188
Summary

• Physiology and homeostasis
• Endocrine system
• Neural system
• ANN-AES

189
Tutorial Summary

• Looked at AIS from an interdisciplinary
  perspective
• Covered the basics of AIS
• There is a great deal more out there, but the
  extra reading should be useful.