Title: P1254413683TZtyA
1Soft Computing
BISC The Berkeley Initiative in Soft Computing Electrical Engineering and Computer Sciences Department BISC The Berkeley Initiative in Soft Computing Electrical Engineering and Computer Sciences Department
Masoud Nikravesh BISC Program, EECSUCB National Energy Research Scientific Computing Center (NERSC) Lawrence Berkeley National Laboratory (LBNL) U.S. Department of Energy (DOE) http//wwwbisc.cs.berkeley.edu/ Email Nikravesh_at_cs.berkeley.edu Tel (510) 6434522 Fax (510) 6425775 For Internal Use Only. Not for Distribution Senior Research Fellow, British Telecom (BTexact Technology)
Masoud Nikravesh BISC Program, EECSUCB National Energy Research Scientific Computing Center (NERSC) Lawrence Berkeley National Laboratory (LBNL) U.S. Department of Energy (DOE) http//wwwbisc.cs.berkeley.edu/ Email Nikravesh_at_cs.berkeley.edu Tel (510) 6434522 Fax (510) 6425775 For Internal Use Only. Not for Distribution Senior Research Fellow, British Telecom (BTexact Technology)
Fuzzy Set 1965 Fuzzy Logic 1973 Soft Decision 1981 BISC 1990 HumanMachine Perception 2000  Fuzzy Set 1965 Fuzzy Logic 1973 Soft Decision 1981 BISC 1990 HumanMachine Perception 2000 
2Outline
 Intelligent Systems/Historical Perspective
 Foundation of Soft Computing
 Evolution of Soft Computing
 Neural Network / Neuro Computing
 Fuzzy Logic / Fuzzy Computing
 Genetic Algorithm / Evolutionary Computing
 Hybrid Systems
 BISC Decision Support System
 Demo
 Conclusions
3Computation
 Traditional Sense Manipulation of Numbers
 Human Uses Word for Computation and Reasoning
 Conclusions lt Word lt Natural Language
4Intelligent System
The role model for intelligent system is Human
Mind.
 Dreyfus
 Minds do not use a theory about the everyday
world  Knowhow vs know that
 Winograd
 Intelligent systems act, dont think
5Artificial Intelligence
 Knowledge Representation
 Predicates
 Production rules
 Semantic networks
 Frames
 Inference Engine
 Learning
 Common Sense Heuristics
 Uncertainty
6Artificial Intelligence
 Applications
 Expert tasks
 the algorithm does not exist
 A medical encyclopedia is not equivalent to a
physician  Heuristics
 There is an algorithm but it is useless
 Uncertainty
 The algorithm is not possible
 Complex problems
 The algorithm is too complicated
 Technologies
 Expert systems
 Natural language processing
 Symbolic processing
 Knowledge engineering
7COMMON SENSE
 Deduction is a method of exact inference
(classical logic)  All Greeks are humans and Socrates is a Greek,
therefore Socrates is a human  Induction infers generalizations from a set of
events (science)  Water boils at 100 degrees
 Abduction infers plausible causes of an effect
(medicine)  You have the symptoms of a flue
8COMMON SENSE
 Uncertainty
 Maybe I will go shopping
 Probability
 Probability measures how often an event occurs
 Principle of incompatibility (Pierre Duhem)
 The certainty that a proposition is true
decreases with any increase of its precision  The power of a vague assertion rests in its being
vague (I am not tall)  A very precise assertion is almost never certain
(I am 1.71cm tall)
9COMMON SENSE
 The Frame Problem
 Classical logic deducts all that is possible from
all that is available  In the real world the amount of information that
is available is infinite  It is not possible to represent what does not
change in the universe as a result of an action  Infinite things change, because one can go into
greater and greater detail of description  The number of preconditions to the execution of
any action is also infinite, as the number of
things that can go wrong is infinite
10COMMON SENSE
 Classical Logic is inadequate for ordinary life
 Intuitionism
 Non Monotonic Logic
 Second thoughts
 Plausible reasoning
 Quick, efficient response to problems when an
exact solution is not necessary
Heuristics Rules of thumbs George Polya
Heuretics
The World Of Objects The Measure
Space Qualitative Reasoning
11COMMON SENSE
 Fuzzy Logic
 Not just zero and one, true and false
 Things can belong to more than one category, and
they can even belong to opposite categories, and
that they can belong to a category only partially  The degree of membership can assume any value
between zero and one
12Cost
As complexity rises, precise statements lose
meaning, and meaningful statements lose
precision. (L.A. Zadeh)
Principle of incompatibility (Pierre Duhem) The
certainty that a proposition is true decreases
with any increase of its precision The power of a
vague assertion rests in its being vague (I am
not tall) A very precise assertion is almost
never certain (I am 1.71cm tall)
Uncertainty
Precision
13TURINGs TEST
 Turing A computer can be said to be intelligent
if its answers are indistinguishable from the
answers of a human being
Computer
14A new age
1929 HUBBLES EXPANDING UNIVERSE 1930
TELEVISION 1940 MODERN SYNTHESIS 1940
SEMIOTICS 1944 DNA 1945 ATOMIC BOMB 1946
COMPUTER 1948 TRANSISTOR 1949 HEBBS LAW 1953
DOUBLEHELIX 1955 ARTIFICIAL INTELLIGENCE 1958
LINGUISTICS 1961 SPACE TRAVEL 1967 QUARKS 1974
SUPERSTRING 1978 NEURAL DARWINISM 1982 PERSONAL
COMPUTER 1988 SELFORGANIZING SYSTEMS
 1859 THEORY OF EVOLUTION
 1854 RIEMANNS GEOMETRY
 1862 AUTOMOBILE
 1865 HEREDITARITY
 1870 THERMODYNAMICS
 1876 TELEPHONE
 1877 GRAMOPHONE
 1888 ELECTROMAGNETISM
 1894 CINEMA
 1900 PLANCKS QUANTUM
 1903 AIRPLANE
 1905 EINSTEINS SPECIAL RELATIVITY
 1907 RADIO
 1916 EINSTEINS GENERAL RELATIVITY
 1920 POPULATION GENETICS
 1923 WAVEPARTICLE DUALISM
 1926 QUANTUM MECHANICS
15Different Historical Paths
 PHILOSOPHY (SINCE BEGINNING)
 PSYCHOLOGY (SINCE BEGINNING)
 MATH (SINCE FREGE)
 BIOLOGY (SINCE DISCOVERY OF NEURONS)
 COMPUTER SCIENCE (SINCE A.I., 1955)
 LINGUISTICS (SINCE CHOMSKY, 1960s)
 PHYSICS (RECENTLY, 1980s)
Thinking About Thought The Nature of Mind Piero
Scaruffi (Oct 8  Dec 10, 2002)
16The Nature of MindThe Contribution of
Information Science
 The mind as a symbol processor
 Formal study of human knowledge
 Knowledge processing
 Commonsense knowledge
 Neural Networks
Thinking About Thought The Nature of Mind Piero
Scaruffi (Oct 8  Dec 10, 2002)
17The Nature of MindThe Contribution of Linguistics
 Competence over performance
 Pragmatics
 Metaphor
Thinking About Thought The Nature of Mind Piero
Scaruffi (Oct 8  Dec 10, 2002)
18The Nature of MindThe Contribution of Psychology
 The mind as a processor of concepts
 Reconstructive memory
 Memory is learning and is reasoning.
 Fundamental unity of cognition
Thinking About Thought The Nature of Mind Piero
Scaruffi (Oct 8  Dec 10, 2002)
19The Nature of Mind The Contribution of
Neurophysiology
 The brain is an evolutionary system
 Mind shaped mainly by genes and experience
 Neurallevel competition
 Connectionism
Thinking About Thought The Nature of Mind Piero
Scaruffi (Oct 8  Dec 10, 2002)
20The Nature of Mind The Contribution of Physics
 Living beings create order from disorder
 Nonequilibrium thermodynamics
 Selforganizing systems
 The mind as a selforganizing system
 Theories of consciousness based on quantum
relativity physics
Thinking About Thought The Nature of Mind Piero
Scaruffi (Oct 8  Dec 10, 2002)
21MACHINE INTELLIGENCE History
 Make it idiot proof and someone will make a
better idiot
22 DAVID HILBERT (1928)
 MATHEMATICS BLIND MANIPULATION OF SYMBOLS
 FORMAL SYSTEM A SET OF AXIOMS AND A SET OF
INFERENCE RULES  PROPOSITIONS AND PREDICATES
 DEDUCTION EXACT REASONING
 KURT GOEDEL (1931)
 A CONCEPT OF TRUTH CANNOT BE DEFINED WITHIN A
FORMAL SYSTEM
 ALFRED TARSKI (1935)
 DEFINITION OF TRUTH A STATEMENT IS TRUE IF IT
CORRESPONDS TO REALITY (CORRESPONDENCE THEORY OF
TRUTH)
ALFRED TARSKI (1935) BUILD MODELS OF THE WORLD
WHICH YIELD INTERPRETATIONS OF SENTENCES IN THAT
WORLD TRUTH CAN ONLY BE RELATIVE TO
SOMETHING METATHEORY
23 ALAN TURING (1936)
 COMPUTATION THE FORMAL MANIPULATION OF SYMBOLS
THROUGH THE APPLICATION OF FORMAL RULES  HILBERTS PROGRAM REDUCED TO MANIPULATION OF
SYMBOLS  LOGIC SYMBOL PROCESSING
 EACH PREDICATE IS DEFINED BY A FUNCTION, EACH
FUNCTION IS DEFINED BY AN ALGORITHM
 NORBERT WIENER (1947)
 CYBERNETICS
 BRIDGE BETWEEN MACHINES AND NATURE, BETWEEN
ARTIFICIAL SYSTEMS AND NATURAL SYSTEMS  FEEDBACK, HOMEOSTASIS, MESSAGE, NOISE,
INFORMATION  PARADIGM SHIFT FROM THE WORLD OF CONTINUOUS LAWS
TO THE WORLD OF ALGORITHMS, DIGITAL VS ANALOG
WORLD
24 CLAUDE SHANNON AND WARREN WEAVER (1949)
 INFORMATION THEORY
 ENTROPY A MEASURE OF DISORDER A MEASURE OF
THE LACK OF INFORMATION  LEON BRILLOUINS NEGENTROPY PRINCIPLE OF
INFORMATION
ANDREI KOLMOGOROV (1960) ALGORITHMIC INFORMATION
THEORY COMPLEXITY QUANTITY OF
INFORMATION CAPACITY OF THE HUMAN BRAIN 10 TO
THE 15TH POWER MAXIMUM AMOUNT OF INFORMATION
STORED IN A HUMAN BEING 10 TO THE 45TH ENTROPY
OF A HUMAN BEING 10 TO THE 23TH
 LOTFI A. ZADEH (1965)
 FUZZY SET
 Stated informally, the essence of this
principle is that as the complexity of a
system increases, our ability to make precise and
yet significant statements about its behavior
diminishes until a threshold is reached beyond
which precision and significance (or
relevance) become almost mutually exclusive
characteristics.
25MACHINE INTELLIGENCE History
 1936 TURING MACHINE
 1940 VON NEUMANNS DISTINCTION BETWEEN DATA AND
INSTRUCTIONS  1943 FIRST COMPUTER
 1943 MCCULLOUCH PITTS NEURON
 1947 VON NEUMANNs SELFREPRODUCING AUTOMATA
 1948 WIENERS CYBERNETICS
 1950 TURINGS TEST
 1956 DARTMOUTH CONFERENCE ON ARTIFICIAL
INTELLIGENCE  1957 NEWELL SIMONS GENERAL PROBLEM SOLVER
 1957 ROSENBLATTS PERCEPTRON
 1958 SELFRIDGES PANDEMONIUM
 1957 CHOMSKYS GRAMMAR
 1959 SAMUELS CHECKERS
 1960 PUTNAMS COMPUTATIONAL FUNCTIONALISM
 1960 WIDROWS ADALINE
 1965 FEIGENBAUMS DENDRAL
26MACHINE INTELLIGENCE History
 1965 ZADEHS FUZZY LOGIC
 1966 WEIZENBAUMS ELIZA
 1967 HAYESROTHS HEARSAY
 1967 FILLMORES CASE FRAME GRAMMAR
 1969 MINSKY PAPERTS PAPER ON NEURAL NETWORKS
 1970 WOODS ATN
 1972 BUCHANANS MYCIN
 1972 WINOGRADS SHRDLU
 1974 MINSKYS FRAME
 1975 SCHANKS SCRIPT
 1975 HOLLANDS GENETIC ALGORITHMS
 1979 CLANCEYS GUIDON
 1980 SEARLES CHINESE ROOM ARTICLE
 1980 MCDERMOTTS XCON
 1982 HOPFIELDS NEURAL NET
 1986 RUMELHART MCCLELLANDS PDP
 1990 ZADEHS SOFT COMPUTING
 2000 ZADEHS COMPUTING WITH WORDS AND
PERCEPTIONS PNL
27What is Soft Computing
BISC The Berkeley Initiative in Soft Computing Electrical Engineering and Computer Sciences Department BISC The Berkeley Initiative in Soft Computing Electrical Engineering and Computer Sciences Department
The basic ideas underlying soft computing in its current incarnation have links to many earlier influences, among them Prof. Zadehs 1965 paper on fuzzy sets the 1973 paper on the analysis of complex systems and decision processes and the 1979 report (1981 paper) on possibility theory and soft data analysis. The principal constituents of soft computing (SC) are fuzzy logic (FL), neural network theory (NN) and probabilistic reasoning (PR), with the latter subsuming belief networks, evolutionary computing including DNA computing, chaos theory and parts of learning theory.
The basic ideas underlying soft computing in its current incarnation have links to many earlier influences, among them Prof. Zadehs 1965 paper on fuzzy sets the 1973 paper on the analysis of complex systems and decision processes and the 1979 report (1981 paper) on possibility theory and soft data analysis. The principal constituents of soft computing (SC) are fuzzy logic (FL), neural network theory (NN) and probabilistic reasoning (PR), with the latter subsuming belief networks, evolutionary computing including DNA computing, chaos theory and parts of learning theory.
Fuzzy Set 1965 Fuzzy Logic 1973 Soft Decision 1981 BISC 1990 HumanMachine Perception 2000  Fuzzy Set 1965 Fuzzy Logic 1973 Soft Decision 1981 BISC 1990 HumanMachine Perception 2000 
28SOFT COMPUTING
SC
29SOFT COMPUTING
Soft computing is consortium of computing
methodologies which collectively provide a
foundation for the Conception, Design and
Deployment of Intelligent Systems. L.A.
Zadeh ...in contrast to traditional hard
computing, soft computing exploits the tolerance
for imprecision, uncertainty, and partial truth
to achieve tractability, robustness, low
solutioncost, and better rapport with
reality L.A. Zadeh The role model for
Soft Computing is the Human Mind.
30SOFT COMPUTING
 NeuroComputing (NC)
 Fuzzy Logic (GL)
 Genetic Computing (GC)
 Probabilistic Reasoning (PR)
 Chaotic Systems (CS), Belief Networks (BN),
Learning Theory (LT)  Related Technologies
 Statistics (Stat.)
 Artificial Intelligence (AI)
 CaseBased Reasoning (CBR)
 RuleBased Expert Systems (RBR)
 Machine Learning (Induction Trees)
 Bayesian Belief Networks (BBN)
31SOFT COMPUTING
 Neural Networks
 create complicated models without knowing their
structure  gradually adapt existing models using training
data  Fuzzy Logic
 Fuzzy Rules are easy and intuitively
understandable  Genetic Algorithms
 find parameters through evolution(usually when a
direct algorithm is unknown)
32Neural Networks
 Ensemble of simple processing units
 Connection weights define functionality
 Derive weights from training data(usually
gradient descent based algorithms)
33Fuzzy Logic
 Allow partial membership to sets
 Express knowledge through linguistic terms and
rules (Computing with Words)  Derive sets of Fuzzy Rules from data (usually
based on heuristics)
34Evolutionary Algorithms
 Finding an optimal structure (parameters) for a
model is often complicated (due to large search
space, complex structure)  Find structure (parameters) through evolution
(generate population, evaluate, breed new pop.)
35Why Fuzzy Logic
 Uncertainty in the data and laws of nature
 Imprecision due to measurement human error
 Incomplete and sparse information
 Subjective and Linguistic rules
 So far as the laws of mathematics refer to
reality, they are not certain and so far as they
are certain, they dont refer to reality  Albert Einstein
36Words are less precise than numbers!
 When information is too imprecise
 Close to reality
 Complex problem
 As complexity rises, precise statements lose
meaning, and meaningful statements lose
precision.  Lotfi A. Zadeh
37Why Neural Network
 Structure Free Nonlinear Mapping
 Multivariable Systems
 Trains Easily Based on Historical Data
 Parallel Processing Fault Tolerance
 Much Like Human Brain
38Why Evolutionary Computing
 For Multiobjectives and MultiCriteria
Optimization Purposes  Resolving Conflict
 Capability to learn. Adapt, and to be selfaware
 Darwinians law
39Why Fuzzy Evolutionary Computing
 To Extract Fuzzy Rules
 The Tune Fuzzy Membership
40Neural Network Fuzzy Logic Models
 A Neural Network capturing presence of Fuzzy
Rules  Ideal for Knowledge Acquisition / Discovery
 Introduces fuzzy weights to NN internal
structure
41SOFT COMPUTINGNeural Network
NNet
42Biological Neuron
43Biological Neuron
44(No Transcript)
45Biological vs. Artificial Neuron
46Analogy between biological and artificial neural
networks
47Artificial Neuron
48Schematic Diagram for Single Neuron
b
1
w22
w1
yf(s)
sxiwi
x1
y
wk
xk
y f b w1 x1 w2 x2 wk xk . w22
49Activation Functions
f(s)
sgn(s)
1
s
1
50Multilayer perceptron with two hidden layers
i g n a l s
Input Signals
O u t p u t S
First
Second
Input
hidden
hidden
Output
layer
layer
layer
layer
Artificial Neural Network (Feedforward)
51Multi Layer Perceptron ANN
52Mapping Functions
 One hidden layer (two adaptive layers) networks
can approximate any functional continuous mapping
from one finite dimensional space to another  Onetoone mapping
 Manytoone mapping
 Manytomany mapping
53Artificial Neural Network vs. Human Brain
 Largest neural computer
 20,000 neurons
 Worms brain
 1,000 neurons
 But the worms brain outperforms neural computers
 Its the connections, not the neurons!
 Human brain
 100,000,000,000 neurons
 200,000,000,000,000 connections
54Brain vs. Computer Processing
Processing Speed Milliseconds VS
Nanoseconds. Processing Order Massively
parallel.VS serially. Abundance and
Complexity 1011 and 1014 of neurons operate in
parallel in the brain at any given moment,
each with between 103 and 104 abutting
connections per neuron. Knowledge Storage
Adaptable VS New information destroys old
information. Fault Tolerance Knowledge is
retained through the redundant, distributed
encoding information VS the corruption of a
conventional computers memory is irretrievevable
and leads to failure as well.
Cesare Pianese
55NEURAL NETWORK
PERCEPTRON
56History of Neurocomputing
Cesare Pianese
57 First Attempts
 Simple neurons which are binary devices with
fixed thresholds simple logic functions like
and, or McCulloch and Pitts (1943)  Promising Emerging Technology
 Perceptron three layers network which can learn
to connect or associate a given input to a random
output  Rosenblatt (1958)  ADALINE (ADAptive LInear Element) an analogue
electronic device which uses leastmeansquares
(LMS) learning rule Widrow Hoff (1960)  1962 Rosenblatt proved the convergence of the
perceptron training rule.  Period of Frustration Disrepute
 Minsky Paperts book in 1969 in which they
generalised the limitations of single layer
Perceptrons to multilayered systems.  ...our intuitive judgment that the extension (to
multilayer systems) is sterile  Innovation
 Grossbergs (Steve Grossberg and Gail Carpenter
in 1988) ART (Adaptive Resonance Theory) networks
based on biologically plausible models.  Anderson and Kohonen developed associative
techniques  Klopf (A. Henry Klopf) in 1972, developed a basis
for learning in artificial neurons based on a
biological principle for neuronal learning called
heterostasis.  Werbos (Paul Werbos 1974) developed and used the
backpropagation learning method  19701985 Very little research on Neural Nets
 Fukushimas (F. Kunihiko) cognitron (a step wise
trained multilayered neural network for
interpretation of handwritten characters).  ReEmergence
 1986 Invention of Backpropagation Rumelhart and
McClelland, but also Parker and earlier on
Werbos which can learn from nonlinearlyseparable
data sets.  Since 1985 A lot of research in Neural Nets!
58HechtNielsens MLP/backprop
Theory of backpropagation neural network, IEE
Proc., Int. Conf. NNet, Washignton, DC, 1989
Robert HechtNielsen 180747, San Francisco HNC
Software/UCSD electrical engineering computer
science Neurocomputing Addison/Wesley, 1991
59Different NonLinearlySeparable Problems
Types of Decision Regions
ExclusiveOR Problem
Classes with Meshed regions
Most General Region Shapes
Structure
SingleLayer
Half Plane Bounded By Hyperplane
TwoLayer
Convex Open Or Closed Regions
Arbitrary (Complexity Limited by No. of Nodes)
ThreeLayer
60Architecture of Neural Networks
 Layers
 Input, hidden and output layers
 Number of units per layer
 Connections Flow of Information
 Feedforward
 Recurrent
 Fully connected
 Laterally connected
 Modular Networks
61Topology
Layers
Connection Weights
Architecture
Modular
Feedforward
Hetero associative
Recurrent
Autoassociative
Cesare Pianese
62Learning Paradigms
 Supervised learning
 network trained by showing a set of input and
output patterns  Unsupervised learning
 network is shown only the input patterns
 Reinforcement learning
 Information on quality of response is available
63Learning/Training
Supervised Learning Batch learning The
network parameters (weights and biases) are
adjusted once all the training examples have been
presented to the network. The parameters change
is based on the global error made during the
training examples (inputs) classification.
Adaptive learning The weights are modified
after each example has been presented to the
network. The network tries to satisfy all the
examples one at a time. Unsupervised
Learning The network has no feedback about the
output accuracy and learns through a
selfarrangement of its structure similar inputs
activate similar neurons while different inputs
activate other neurons.
64 Neural Network Classification
Cesare Pianese
65Neural Network Models
 Neural network Models are Characterised by
 Type of Neurons(units, nodes, neurodes)
 Connectionist architecture
 Learning algorithm
 Recall algorithm
66Multilayer Perceptron
Kohonen
Radial Basis Functions
Neural Network Models
Generalised Regression
Probabilistic
ART
Recurrent
Cesare Pianese
67 Neural Network Classification
Cesare Pianese
68Unit delay operator
Recurrent network without hidden units
Recurrent network with hidden units
Artificial Neural Network (Feedforward)
69Other Types of Neural Networks
SOM
RBFN
Committee Machines
ART1
70 Output layer
Hidden layer
Input layer
Artificial Neural Network (recurrent)
71A Common Framework for Neural Networks and
Multivariate Statistical Method
Y (X1, X2, , Xp) k k (k X1, X2, , Xp)
72Input Transformation
KernelBased Method
Linear
Nonlinear
73(No Transcript)
74(No Transcript)
75BACKPROBAGATION NEURAL NETWORK
BNNet
761974 Werbos Backprop
math economics (by lack of brain
theory) Optimization A Foundation for
Understanding Consciousness In Optimality in
Biological and Artificial Neural Networks,
Levine and Elsberry eds, Erlbaum 1997
Paul J. Werbos 040947, Philadelphia NSF,
Arlington VA 1974 Ph.D. dissertation at Harvard
University
77Typical Neural Network
 A neural network is a set of interconnected
neurons (simple processing units)  Each neuron receives signals from other neurons
and sends an output to other neurons  The signals are amplified by the strength of
the connection  The strength of the connection changes over time
according to a feedback mechanism  The net can be trained
78Threelayer backpropagation neural network
Input signals
1
x
1
y
1
1
1
2
x
2
y
2
2
2
i
w
w
j
jk
ij
y
x
k
k
i
m
n
y
l
l
x
n
Hidden
Input
Output
layer
layer
layer
Error signals
79Backpropagation  Gradient Descent
 Weight initialisation
 saturation of units
 random values
 Error surface characteristics
 local and global minima
 multidimensional plateaus
 narrow valleys or ravines
 Learning rate
 Momentum term
80Fixed step size too small
Fixed step size too large
Steepest descent with line minimization
Fixed step size too large
81DFP
start
BFGS
start
End
End
BroydenFletcherGoldfarbShanno
(Unconstrained quasiNewton
minimization)
DavidonFletcherPowell
(Unconstrained quasiNewton
minimization)
Steepest
Simplex
start
start
End
End
NelderMead
(Unconstrained simplex
minimization)
Steepest Descent
(Unconstrained minimization)
LM
GN
start
start
End
End
LevenbergMarquardt
(Least squares
minimization)
GaussNewton
(Least squares
minimization)
82NEURAL NETWORK
NNet
83NEURAL NETWORK ADAptive LInear Neuron ADAptive
LInear Element
ADALIN / MADLIN
841960 WIDROWS ADALINE
1960 ADALINE (ADAptive LInear Element) an
analogue electronic device which uses
leastmeansquares (LMS) learning rule Widrow
Hoff 1962 ADALINE (ADAptive LInear Neuron)
Generalization and information storage in
networks of ADALINE neurons
 Bernard Widrow and Ted Hoff introduced the
LeastMeanSquare algorithm (a.k.a.  deltarule or WidrowHoff rule) and used it to
train the ADALINE (ADAptive Linear Neuron)  The ADALINE was similar to the perceptron, except
that it used a linear activation function instead
of the threshold  The LMS algorithm is still heavily used in
adaptive signal processing
Bernard Widrow 240129, Norwich CT electrical
engineering, Stanford
85Widrow and Hoff, 1960
Bernard Widrow and Ted Hoff introduced the
LeastMeanSquare algorithm (a.k.a. deltarule or
WidrowHoff rule) and used it to train the
Adaline (ADAptive Linear Neuron) The Adaline
was similar to the perceptron, except that it
used a linear activation function instead of the
threshold The LMS algorithm is still heavily
used in adaptive signal processing
MADALINE Many ADALINEs Network of ADALINEs
86Perceptron vs. ADALINE
Percptron LTU Emperical Hebbian
Assumption ADALINE LGU GradientDecent
f(s)
sgn(s)
1
linear(s)
s
LTU sign function / (Positive/Negative) LGU
Continuous and Differentiable Activation function
including Linear function
1
MADALINE Many ADALINEs Network of ADALINEs
87Linearly nonSeparable
Linearly Separable
88NEURAL NETWORKSupport Vector Machine
SVM
89From basic trigonometry, the distance between a
point x and a plane (w,b) is
Cherkassky and Mulier, 1998
Noticing that the optimal hyperplane has infinite
solutions by simply scaling weight vector and
bias, we choose the solution for which the
discriminant function becomes one for the
training examples closest to the boundary
This is known as the canonical hyperplane
Ricardo GutierrezOsuna
90Therefore, the distance from the closest example
to the boundary is
And the margin becomes
Ricardo GutierrezOsuna
91Cherkassky and Mulier, 1998 Haykin, 1999
Schölkopf, 2002 _at_ http//kernelmachines.org/
92Burges, 1998 Kaykin, 1999
931982 Kohonens SOM
presents selforganizing maps using one and
twodimensional lattice structures. Kohonen SOMs
have received far more attention than van der
Malsburgs work and have become the benchmark for
innovations in selforganization
Unsupervised Hebbian learning Hebb gt
support/strengthen activity learns only iff
input data redundant Find/recognize
patterns/structures
Teuvo Kohonen born 110734 Finland TU Helsinki
SOFM selforganized feature mapping, or SOM
selforganized mapping
Selforganizing maps. Springer (1997)
94(No Transcript)
95(No Transcript)
96NEURAL NETWORKLearning Vector Quantization
LVQ
97Learning Vector Quantization (LVQ)
VQ can be considered a special case of SOFM. The
treatment will overlap at several points. Many
researchers have contributed to the area, notably
Kohonen and coworkers.
Although the SOM can be used for classification
as such, one has to remember that it does not
utilize class information at all, and thus its
results are inherently suboptimal. However, with
small modifications, the network can take the
class into account. The function SOM_SUPERVISED
does this. Learning vector quantization (LVQ) is
an algorithm that is very similar to the SOM in
many aspects. However, it is specifically
designed for classification.
98Learning Vector Quantization (LVQ)
Single element from a competitive layer coupling
within layer, such that the best wins, or the
winner takes all. Achieved by positive feedback
to itself and inhibition of others. The winning
node have weight vector is called VQ vectors.
 often Euclidian distance, or some squared error
measure is used to find the minimum distance
Duifhuis
99SOM, Unsupervised
Although the SOM can be used for classification
as such, one has to remember that it does not
utilize class information at all, and thus its
results are inherently suboptimal.
Learning vector quantization (LVQ) is an
algorithm that is very similar to the SOM in many
aspects. However, it is specifically designed for
classification.
LVQ, Supervised
100NEURAL NETWORKAuto Resonance Theory
ART
1011988 Grossbergs ART
 (Steve Grossberg and Gail Carpenter in 1988) ART
(Adaptive Resonance Theory) networks based on
biologically plausible models.
 Carpenter (1997) Distributed learning,
recognition, and prediction by ART and ARTMAP
neural networks, Neural Networks 10, 1473 1494.  Grossberg (1995) The attentive brain, American
Scientist 88, 438  449
Gail Carpenter and Stephen Grossberg Boston
University Cognitive neural systems,
mathematical psychology
102yj
output categories




lateral connections
F2 layer, N nodes template choosing
topdown wji bottomup vij
xi
F1 layer, M nodes template matching
ii
binary inputs
 Single element from an ART1 network lateral
connections here limited to output layer, where
the best wins, or the winner takes all.
Examples of a set of bottomup connections and a
set of topdown connections.  In some versions of ART1 also lateral
connections in F1
Duifhuis
103The ART2 network architecture, dynamics
output
xi
Layer F2
j
yj
distributed AGCs
wji
vij
Topdown path
ri
, vigilance
F1
Layer F1

pi
qi
ui
si
Bottomup path
ti
xi
Expanded node I in layer F1
inputs
Duifhuis
104(1994) CBD Signal Function
SIMULATIONS Average of 10 runs, 1 training
epoch Training points 1,000 (DIAG) or 10,000
(CIS), Testing points 10,000
CircleinSquare (CIS)
DIAGONAL
91.70 correct 15.2 coding nodes
94.41 correct 40.7 coding nodes
105NEW Graded Signal Function
DIAGONAL
CircleinSquare (CIS)
95.26 correct (h1) 17.8 coding nodes
96.73 correct (h0.5) 46.4 coding nodes
 Smoother boundaries
 Improved correct but slightly more nodes
106Point ARTMAP
98.72 correct 275 coding nodes
97.8 correct 59.2 coding nodes
107NEURAL NETWORKRadial Basis Function
RBF
108Radial Basis Function
A hidden layer of radial kernels The hidden layer
performs a nonlinear transformation of input
space The resulting hidden space is typically of
higher dimensionality than the input space An
output layer of linear neurons The output layer
performs linear regression to predict the desired
targets Dimension of hidden layer is much larger
than that of input layer Covers theorem on the
separability of patterns A complex
patternclassification problem cast in a
highdimensional space nonlinearly is more
likely to be linearly separable than in a
lowdimensional space Support Vector Machines.
RBFs are one of the kernel functions
most commonly used in SVMs!
109Radial Basis Function
RBFs have their origins in techniques for
performing exact function interpolation Bishop,
1995
110Ricardo GutierrezOsuna
111RBFs Training Haykin, 1999
Unsupervised selection of RBF centers RBF centers
are selected so as to match the distribution of
training examples in the input feature
space Supervised computation of output
vectors Hiddentooutput weight vectors are
determined so as to minimize the sum squared
error between the RBF outputs and the desired
targets Since the outputs are linear, the optimal
weights can be computed using fast, linear matrix
inversion
Once the RBF centers have been selected,
hiddentooutput weights are computed so as to
minimize the MSE error at the output
112RBFs Training Haykin, 1999
Once the RBF centers have been selected,
hiddentooutput weights are computed so as to
minimize the MSE error at the output
Now, since the hidden activation patterns are
fixed, the optimum weight vector W can be
obtained directly from the conventional
pseudoinverse solution
113Nikravesh and Nikravesh et al. (19942003) 
Functional NNet (19931994) Faster and more
Robust Training (19931994)  Better
Uncertainty Analysis (19942000) As A Basis
for Computing with Words and Perceptions (CWP)
(20002003)
114NEURAL NETWORK
PNN
115Probabilistic Neural Networks
GRNN and PNN are more computationally intense
than RBF
Generalized regression (GRNN) and probabilistic
(PNN) networks are variants of the radial basis
function (RBF) network. Unlike the standard RBF,
the weights of theses networks can be calculated
analytically. In this case, the number of cluster
centers is by definition equal to the number of
exemplars, and they are all set to the same
variance. Use this type of RBF only when the
number of exemplars is so small (lt100) or so
dispersed that clustering is illdefined.
116RBF Networks
117NEURAL NETWORKGeneralized Regression Neural
Network
GRNN
118KNearest Neighbors
KNN
119KNearest Neighbors
The K Nearest Neighbor Rule (kNNR) is a very
intuitive method that classifies unlabeled
examples based on their similarity with examples
in the training set For a given unlabeled
example xu.D, find the k closest labeled
examples in the training data set and assign xu
to the class that appears most frequently within
the ksubset The kNNR only requires An integer
k A set of labeled examples (training data) A
metric to measure closeness
Example In the example below we have three
classesand the goal is to find a class label for
the unknown example xu In this case we use the
Euclidean distance and a value of k5
neighbors Of the 5 closest neighbors, 4 belong to
1 and 1 belongs to 3, so xu is assigned to 1,
the predominant class
Ricardo GutierrezOsuna
120KNearest Neighbors
Data a 2dimensional 3 class problem, where the
classconditional densities are multimodal, and
nonlinearly separable, as illustrated in the
figure Use kNNR with k five Metric
Euclidean distance The resulting decision
boundaries and decision regions are shown below
Ricardo GutierrezOsuna
1211NNR versus kNNR The use of large values of k
has two main advantages Yields smoother decision
regions Provides probabilistic information The
ratio of examples for each class gives
information about the ambiguity of the
decision However, too large values of k are
detrimental It destroys the locality of the
estimation since farther examples are taken into
account In addition, it increases the
computational burden
Ricardo GutierrezOsuna
122NEURAL NETWORKBoltzmann Machine
BM
123 The Boltzmann Machine is related to the Hopfield
network. It is an extension in that it can have
input units, hidden units and output units.  supervised learning in general networks
 stochastic network (Hinton, Sejnowski, ..)
 symmetric weight, like Hopfield, but with hidden
units  special training to avoid local minima, and reach
global minima simulated annealing.  different uses
 associative memory
 hetero associative mapping
 solve optimization problems
 Partial, or noisy input patterns as inputs gt
complete patterns as outputs. Similar to previous
figure (called heteroassociative). This is called
autoassociative. No separate in / out, only
visible units.
124SOFT COMPUTINGFuzzy Logic
FL
125VARIABLES AND LINGUISTIC VARIABLES
 one of the most basic concepts in science is that
of a variable  variable numerical (X5 X(3, 2) )
 linguistic (X is small (X, Y) is much
larger)  a linguistic variable is a variable whose values
are words or sentences in a natural or synthetic
language (Zadeh 1973)  the concept of a linguistic variable plays a
central role in fuzzy logic and underlies most of
its applications
126Fuzzy Sets
Fuzzy Logic Element x belongs to set A with a
certain degree of membership (x)0,1
Classical Logic Element x belongs to set A or it
does not (x)0,1
A(x)
A(x)
Ayoung
Ayoung
1
1
0
0
x years
x years
127Membership Functions
Predicate Old
Predicate Old
Crisp Set
Fuzzy Set
128EXAMPLES OF FGRANULATION (LINGUISTIC VARIABLES)
color red, blue, green, yellow, age young,
middleaged, old, very old size small, big, very
big, distance near, far, very, not very far,
µ
young
old
middleaged
1
very old
0
Age
129LINGUISTIC VARIABLES AND FGRANULATION (Zadeh,
1973)
example Age primary terms young, middleaged,
old modifiers not, very, quite, rather,
linguistic values young, very young, not very
young and not very old,
µ
young
old
middleaged
1
very old
0
Age
130Fuzzy Logic
Linguistic Rule Knowledge Base
Defuzzifier Module
Fuzzy Inference Engine
Fuzzifier Module
Crisp Input
Crisp Output
Fuzzy Sets Fuzzy Numbers Fuzzification Fuzzy
Operators Fuzzy Rules Fuzzy Inference Defuzzificat
ion
131Observation
132A Discrete Fuzzy Set
Age Young0.25, MiddleAged0.75
Membership of Young to the set Age is 0.25
Membership of MiddleAged to the set Age is 0.75
133Fuzzy Rule Base
If Age is old then Roya is 70
If Age is milddleAged then Roya is 45
If Age is Young then Roya is 20
134Inferencing
Decision 200, 450.75, 700.25
Age
135Defuzzification
Output (200 450.75 700.25) (0 0.75
0.25)
Output 51.2 MiddleAged
136WHAT IS FUZZY LOGIC (FL)
fuzzy logic (FL) has four principal facets
logical (narrow sense FL)
FL/L
F
F.G
FL/E
FL/S
settheoretic
epistemic
G
FL/R
relational
F fuzziness/ fuzzification G granularity/
granulation F.G F and G
137Schema of a Fuzzy Decision
Inference
Fuzzification
Defuzzification
rulebase
if temp is cold then valve is open
cold
warm
hot
open
half
close
cold 0.7
0.7
0.7
if temp is warm then valve is half
0.2
0.2
warm 0.2
t
v
measured temperature
if temp is hot then valve is close
crisp output for valvesetting
hot 0.0
138SOFT COMPUTINGFuzzy Logic
FL
139Fuzzy Logic Genealogy
 Origins MVL for treatment of imprecision and
vagueness  1930s Post, Kleene, and Lukasiewicz attempted to
represent undetermined, unknown, and other
possible intermediate truthvalues.  1937 Max Black suggested the use of a
consistency profile to represent vague
(ambiguous) concepts  1965 Zadeh proposed a complete theory of fuzzy
sets (and its isomorphic fuzzy logic), to
represent and manipulate illdefined concepts
1401965 ZADEHS FUZZY LOGIC
 Stated informally, the essence of this
principle is that as the complexity of a
system increases, our ability to make precise and
yet significant statements about its behavior
diminishes until a threshold is reached beyond
which precision and significance (or
relevance) become almost mutually exclusive
characteristics.  ...in contrast to traditional hard computing,
soft computing exploits the tolerance for
imprecision, uncertainty, and partial truth to
achieve tractability, robustness, low
solutioncost, and better rapport with reality  1949 The concept of a timevarying transfer
function  1950 with J. R. Ragazzini generalization of
Wieners theory of prediction, generalization of
Wieners theory of prediction  1952 with J. R. Ragazzini pioneered in the
development of the ztransform  1953 design of nonlinear filters and constructed
a hierarchy of nonlinear systems based on the
VolterraWiener representation  1963 with Charles Desoer Classic text on the
statespace theory of linear systems
Lotfi A. Zadeh Electrical Engineering, Columbia
University An alumnus of the University of
Teheran, MIT, and Columbia University.
Fuzzy Set 1965 Fuzzy Logic 1973 Soft
Decision 1981 BISC 1990 HumanMachine
Perception 2000 
141WHAT IS FUZZY LOGIC
 fuzzy logic has been and still is, though to a
lesser degree, an object of controversy  for the most part, the controversies are rooted
in misperceptions, especially a misperception of
the relation between fuzzy logic and probability
theory  a source of confusion is that the label fuzzy
logic is used in two different senses  (a) narrow sense fuzzy logic is a logical system
 (b) wide sense fuzzy logic is coextensive with
fuzzy set theory  today, the label fuzzy logic (FL) is used for
the most part in its wide sense
142EVOLUTION OF LOGIC
generality
twovalued
multivalued
fuzzy
LAZ 102600
143EVOLUTION OF LOGIC
 twovalued (Aristotelian) nothing is a matter of
degree  multivalued truth is a matter of degree
 fuzzy everything is a matter of degree
 principle of the excluded middle every
proposition is either true or false
144STATISTICS
Count of papers containing the word fuzzy in
the title, as cited in INSPEC and MATH.SCI.NET
databases. (data for 2001 are not
complete) Compiled by Camille Wanat, Head,
Engineering Library, UC Berkeley, June 21, 2002
INSPEC/fuzzy
Math.Sci.Net/fuzzy
19701979 570 19801989 2,383 19901999 23,121
2000present 5,940 1970present 32,014
441 2,463 5,459 1,670 10,033
145(No Transcript)
146PRINCIPAL APPLICATIONS OF FUZZY LOGIC
FL
 control
 consumer products
 industrial systems
 automotive
 decision analysis
 medicine
 geology
 pattern recognition
 robotics
CFR
CFR calculus of fuzzy rules
147EMERGING APPLICATIONS OF FUZZY LOGIC
 computational theory of perceptions
 natural language processing
 financial engineering
 biomedicine
 legal reasoning
 forecasting
148NEW Tools
149Dividing the Input Space
Grid based
Individual membership functions
150Mamdani Inference System
Output Z
Input MF
A1
B1
C1
X
Y
Z1
A2
B2
C2
Z (centroid of area)
X
Y
Z2
x
y
Output MF
Input (x,y)
151FirstOrder Takagi Sugeno FIS
 Fuzzy Rule base
 If X is A1 and Y is B1 then Z p1x q1y r1
 If X is A2 and Y is B2 then Z p2x q2y r2
152SOFT COMPUTINGNeuroFuzzy Computing
NFC
153NeuroFuzzy Modeling
Hybrid Model
154Adaptive NeuroFuzzy Inference System (ANFIS )
 Takagi Sugeno FIS
 Input partitioning
 LSE gradient descent training
nonlinear parameters
linear parameters
w1
A1
P
w1z1
x
A2
P
S
Swizi
B1
P
/
z
y
B2
P
w4z4
Swi
w4
S
Forward pass
Backward pass
fixed
steepest descent
MF parameter (nonlinear)
leastsquares
fixed
Coefficient parameter (linear)
155NEFCON (NEuro Fuzzy CONtroller)
 Mamdani FIS
 Weights represent fuzzy
 sets.
 Intelligent neurons
 Shared weights to preserve
 semantical characteristics.
 Nodes R1, R2 represent
 the rules.
Output node
 Learning two stages
 Learning fuzzy rules
 Learning fuzzy sets
Input Nodes
156NEFCON Learning
 Learning fuzzy sets
 Fuzzy error backpropagation (reinforcement ) 
FEBP  1. Fuzzy goodness measure
 2. Extended rule based fuzzy error
 3. Determine the contribution of each
rule  4. Modify membership functions
 Learning fuzzy Rules
 Incremental learning
 Decremental learning
NEFPROX ( Function approximation supervised
learning) NEFCLASS ( Classification problems
winner takes all interpretation)
157Fuzzy Adaptive learning Control Network (FALCON)
 Linguistic nodes for each
 output variable.
 One is for training data
 (desired output) and the
 other is for the actual output
 Hybridlearning algorithm
 comprising of unsupervised
 learning to locate initial
 membership functions/ rule
 base and a gradient descent
 learning to optimally adjust the
 parameters of the MF to
 produce the desired outputs
158Generalized Approximate Reasoning based
Intelligent Control (GARIC)
 Implements a neuro
 fuzzy controller by using
 two neural network
 modules, the ASN (Action
 Selection Network) and the
 AEN (Action State
 Evaluation Network).
 The AEN is an adaptive critic that evaluates the
 actions of the ASN.
 GARIC uses a mixture of gradient descent and
 reinforcement learning to finetune the node
 parameters.
 Not easily interpretable
159Self Constructing Neural Fuzzy Inference Network
(SONFIN)
 Takagi Sugeno inference system
 Input space partitioned by
 clustering algorithm related to
 required accuracy
 Projection based correlation
 measure using GramSchmidt
 Orthogonalization algorithm for
 rule consequent parts
 Learning
 Consequent parameters
 Least mean squares algorithm
 Precondition parameters
 Backpropagation
160Evolving Fuzzy Neural Network EFuNN
 Five layered Architecture
 Mamdani type FIS
 Nodes are created during
 learning.New neurons are added
 if MF of input variable lt
 threshold
 Rule base layer learning of
 temporal relationships
 (supervised / unsupervised 
 hybrid learning)
 Performance depends on the
 selection of network parameters
 (error threshold, sensitivity
 threshold etc)
161Fuzzy Inference Environment Software with Tuning
 FINEST
 Parameterization of the inference procedure
 Tuning of fuzzy predicates, combination
implication  functions
 Using backpropagation algorithm
162Fuzzy Net  FUN
 Network initialized with a
 rule base and MFs
 Each layer has different
 activation functions
 Triangular MFs
 Learning
 Rules pure stochastic
 search
 MF Gradient descent or
 reinforcement stochastic
 search
 NeuroFuzzy system
163Evolutionary Design of NeuroFuzzy Systems
164Performance of Some NeuroFuzzy systems
165(No Transcript)
166(No Transcript)
167(No Transcript)
168Compactification Algorithm InterpretationA
Simple Algorithm for Qualitative AnalysisRule
Extraction and Building Decision TreeDr.
Nikravesh and Prof. Zadeh
169Compactification Algorithm Interpretation
A1 A2 o An F1
a11 a21 o am1 a12 a22 o am2 O O O a1n a2n o amn a1 a2 o am
Test Attribute Set
a1 a2 o an b
170Table 1 (intermediate results)
A1 A2 A3 F1
a11 a11 a21 a31 a31 a11 a21 a31 a12 a22 a22 a22 a12 a22 a22 a22 a13 a13 a13 a13 a23 a23 a23 a23 a1 a1 a1 a1 a1 a1 a1 a1
a22 a22 a13 a23 a1 a1
a31 a23 a1
a11 a21 a31 a22 a22 a22 a22 a1 a1 a1 a1
Group 1(initial)
Pass (1)
Pass (2)
Pass (3)
171MAXIMALLY COMPACT REPRESENTATION
172THE CONCEPT OF PSEUDOX
n
PseudoNumber
(non precisiable granule)
n
f
PseudoFunction
Y
(nonprecisiable)
X
 if f is a function for reals to reals, f is a
function from reals to pseudonumbers
173PERCEPTIONBASED VS. MEASUREMENTBASED
INFORMATION
Y
Y
f
f
Y2 X X2
large
0
0
medium
X
X
f if X is small then Y is small if X is
medium then Y is large if X is large then Y
is small
(X,Y) is (small x small medium x large large
x small)
LAZ 11600
174PERCEPTION OF A FUNCTION
Y
f
0
Y
medium x large
f (fuzzy graph)
perception
f f
if X is small then Y is small if X is
medium then Y is large if X is large then Y
is small
0
X
LAZ 72202
175INTERPOLATION
Y is B1 if X is A1 Y is B2 if X is A2 .. Y is
Bn if X is An Y is B if X is A
AA1, , An
Conjuctive approach (Zadeh 1973) Disjunctive
approach (Zadeh 1971, Zadeh 1973,
Mamdani 1974)
176DEFINITION OF p ABOUT 2025
1
cdefinition
v
0
20
25
1
fde