An Overview of Core CI Technologies

1
An Overview of Core CI Technologies

• Luigi Barone, Rm 2.12
• Lyndon While, Rm 1.14

2
Neural Networks (NNs)

• Reading
• S. Russell and P. Norvig, Section 20.5,
  Artificial Intelligence A Modern Approach,
  Prentice Hall, 2002
• G. McNeil and D. Anderson, Artificial Neural
  Networks Technology, The Data Analysis Center
  for Software Technical Report, 1992

3
The Nature-Inspired Metaphor

• Inspired by the brain
• neurons are cells that performs aggregation and
  dissemination of electrical signals
• computational power and intelligence emerges from
  the vast interconnected network of neurons
• NNs act as function approximators or pattern
  recognisers which learn from observed data

Diagrams taken from a report on neural networks
by C. Stergiou and D. Siganos
4
The Neuron Model

• A neuron combines values via its input and
  activation functions
• The bias determines the threshold needed for a
  positive response
• Single-layer neural networks (perceptrons) can
  represent only linearly separable functions

5
Multi-Layered Neural Networks

• A network is formed by the connections (links) of
  many nodes inputs to outputs through one or
  more hidden layers
• Link weights control the behaviour of the
  function represented by the NN
• i.e., adjusting the weights changes the encoded
  function

6
Multi-Layered Neural Networks

• Hidden layers increase the power of the NN
• perceptrons capture only linearly separable
  functions
• a NN with a single, sufficiently large, hidden
  layer can represent any continuous function with
  arbitrary accuracy
• two hidden layers are needed to represent
  discontinuous functions
• at the cost of extra complexity and training
  time
• There are two main types of multi-layered NNs
• feed-forward simple acyclic structure the
  stateless encoding a function of just its current
  input
• recurrent cyclic feedback loops are allowed
  the stateful encoding supports short-term memory

7
Training Neural Networks

• Training corresponds to adjusting link weights to
  minimise some measure of error (the cost
  function)
• i.e., learning is an optimisation search in
  weight space
• Any search algorithm can be used the most common
  using gradient descent (back propagation)
• Common learning paradigms
• supervised learning training by comparison with
  known input/output examples (a training set)
• unsupervised learning no a-priori training set
  is provided the system discovers patterns in the
  input
• reinforcement learning training by using
  environmental feedback to assess the quality of
  actions

8
Evolutionary Algorithms (EAs)

• Reading
• X. Yao, Evolutionary computation a gentle
  introduction, Evolutionary Optimization, 2002
• T. Bäck, U. Hammel, and H.-P. Schwefel,
  Evolutionary computation comments on the
  history and current state, IEEE Transactions on
  Evolutionary Computation 1(1), 1997

9
The Nature-Inspired Metaphor

• Inspired by Darwinian natural selection
• population of individuals exists in some
  environment
• competition for limited resources creates
  selection pressure the fitter individuals that
  are better adapted to the environment are
  rewarded more so than the less fit
• fitter individuals act as parents for the next
  generation
• offspring inherit properties of their parents via
  genetic inheritance, typically sexually through
  crossover with some (small) differences
  (mutation)
• over time, natural selection drives an increase
  in fitness
• EAs are population-based generate-and-test
  stochastic search algorithms

10
Terminology

• Gene the basic heredity unit in the
  representation of individuals
• Genotype the entire genetic makeup of an
  individual the set of genes it possess
• Phenotype the physical manifestation of the
  genotype in the environment
• Fitness evaluation of the phenotype at solving
  the problem of interest
• Evolution occurs in genotype space based on
  fitness performance in phenotype space

11
The Evolutionary Cycle
Parent selection
Parents
Genetic variation (crossover mutation)
Population
Offspring
Survivor selection
12
An Example
X
X
2
4
5
6
8
5
13
EA Variants

• There are many different EA variants/flavours
• differences are mainly cosmetic and irrelevant
  they share more similarities than differences!
• Variations predominately differ on
  representation
• genetic algorithms (GAs) binary strings
• evolution strategies (ESs) real-valued vectors
• genetic programming (GP) expression trees
• evolutionary programming (EP) finite state
  machines
• but also on their historical origins/aims, and
  their emphasis on different variation operators
  and selection schemes

14
Designing an EA

• Best approach for designing an EA
• create a fitness function with an appropriate
  search gradient
• choose a representation to suit the problem,
  ensuring (all) interesting feasible solutions can
  be represented
• choose variation operators to suit the
  representation, being mindful of the search bias
  of the operators
• choose selection operators to ensure efficiency
  while avoiding premature convergence to local
  optima
• tune parameters and operators to the specific
  problem
• Need to balance exploration and exploitation
• variation operations create diversity and novelty
• selection rewards quality and decreases diversity

15
Learning Classifier Systems (LCSs)

• Reading
• R. Urbanomwicz and J. Moore, Learning classifier
  systems a complete introduction, review, and
  roadmap, Journal of Artificial Evolution and
  Applications, 2009
• O. Sigaud and S. Wilson, Learning classifier
  systems a survey, Soft Computing A Fusion of
  Foundations, Methodologies and Applications
  11(11), 2007
• M. Butz and S. Wilson, An algorithmic
  description of XCS, Advances in Learning
  Classifier Systems, 2001

16
The Nature-Inspired Metaphor

• Inspired by a model of human learning
• frequent update of the efficacy of existing rules
• occasional modification of governing rules
• ability to create, remove, and generalise rules
• LCSs simulate adaptive expert systems adapting
  both the value of individual rules and the
  structural composition of rules in the rule set
• LCSs are hybrid machine learning techniques,
  combining reinforcement learning and EAs
• reinforcement learning used to update rule
  quality
• an EA used to update the composition of the rule
  set

17
Algorithm Structure

• An LCS maintains a population of
  condition-action-prediction rules called
  classifiers
• the condition defines when the rule matches
• the action defines what action the system should
  take
• the prediction indicates the expected reward of
  the action
• Each step (input), the LCS
• forms a match set of classifiers whose conditions
  are satisfied by the input
• chooses the action from the match set with the
  highest average reward, weighted by classifier
  fitness (reliability)
• forms the action set the subset of classifiers
  from the match set who suggest the chosen action
• executes the action and observes the returned
  payoff

18
Algorithm Structure

• Simple reinforcement learning is used to update
  prediction and fitness values for each classifier
  in the action set
• A steady-state EA is used to evolve the
  composition of the classifiers in the LCS
• the EA executing at regular intervals to replace
  the weakest members of the population
• the EA operating on the condition and action
  parts of classifiers
• Extra phases for rule subsumption
  (generalisation) and rule creation (covering) are
  used to ensure a minimal covering set of
  classifiers is maintained

19
An Example
Diagram taken from a seminar on using LCSs for
fraud detection by M. Behdad
20
LCS Variants

• There are two main styles of LCS algorithms
• Pittsburgh-style each population member
  represents a separate rule set rule sets form
  permanent teams
• Michigan-style only a single population of rules
  is maintained rules form ad-hoc teams as
  required
• LCS variants differ on the definition of fitness
• strength-based (ZCS) classifier fitness is based
  on the predicted reward of the classifier and not
  its accuracy
• accuracy-based (XCS) classifier fitness is based
  on the accuracy of the classifier and not its
  predicted reward, thus promoting the evolution of
  accurate classifiers
• XCS generally has better performance, although
  understanding when remains an open question

21
Artificial Immune Systems (AISs)

• Reading
• J. Timmis, P. Andrews, N. Owens, and E. Clark,
  An interdisciplinary perspective on artificial
  immune systems, Evolutionary Intelligence 1(1),
  2008
• S. Forrest and C. Beauchemin, Computer
  immunology, Immunological Reviews 236(1), 2007
• E. Hart and J. Timmis, Application areas of AIS
  the past, the present and the future, Applied
  Soft Computing 8(1), 2008

22
The Nature-Inspired Metaphor

• Inspired by the mammalian immune system
• define pathogens as the infectious
  micro-organisms (e.g. viruses, bacteria, etc)
  that can invade a vertebrate
• define antigens as the chemical markers (e.g.
  toxins, enzymes, etc) of the invading
  micro-organisms
• there are two basic types of immune system
  response
• innate immunity a genetic-inherited static
  response against any invading pathogen cells
  bind to common molecular patterns found only in
  micro-organisms
• adaptive (acquired) immunity an acquired dynamic
  response directed against antigen-specific
  invaders cells are modified by exposure to such
  invaders
• Key principle distinguishing self from non-self

23
The Nature-Inspired Metaphor
Diagram taken from a lecture on collaborative
bioinspired algorithms by J. Timmis
24
Immune-Inspired Algorithms

• Common immune-inspired algorithms
• clonal selection inspired by the Darwinian
  evolutionary process occurring with the immune
  system that rewards (replicates clones) cells
  that match an activating antigen
• negative selection inspired by the selection
  process that occurs during cell maturation that
  identifies and deletes self-reacting cells, thus
  leaving only cells capable of binding to non-self
  antigens
• immune network the immune system is considered
  to be a network of self-interactions, thus
  allowing behaviours such as learning and memory
  to emerge
• dendritic cell inspired by the operation of
  dendritic cells that detect danger signals
  emitted by dying self-cells, thus allowing
  tolerance to non-dangerous non-self cells

25
Questions?