Soft Computing For Neural Networks

1. INTRODUCTION TO
x it
SOFT COMPUTING
i
D
inku
R
a & Sachin Lakra,
r Assistant Professor & Head,
Lak Department of IT/MCA
hi n Rinku Dixit,
c
Sa
Assistant Professor,
Department of IT/MCA
Manav Rachna College of Engineering
1

2. Contents
 Intelligent systems ix it
D
 Soft computing
inku
R
 Application areas of soft computing
&
ra
L ak
in
ch
Sa
2

3. Traditions in human sciences
ix it
D
Materialism Mathematics
& bivalent
inku
logic
R
a &
Natural sciences
r
Empiricism Rationalism
ak
Positivism
L
Hermeneutics
n
etc.
c hi
Human sciences Human sciences
Sa
(quantitative) (qualitative)
3

4. Intelligent systems (ISs)
Intelligence: System must
ix it
perform meaningful operations. D
interpret information. inku
R
&
comprehend the relations between phenomena or objects.
a
r to new conditions.
ak
apply the acquired information
L
in
ch
Sa
4

5. Short-Term Objectives of ISs
 Everyday routine tasks of human ix it beings:
vision, language processing, u D common sense
reasoning, learning, robotics. i nk
 Artificial routine tasks &
R
identified and developed
a
rgames, mathematics, logic,
by human beings:
programming. L
ak
in developed by human beings:
ch
 Expert tasks
Sa Physicists, Mechanical Engineers,
Doctors,
accountants, other specialisations.
5

6. Long Term Objectives of ISs
Objectives: To develop a system whichix it
D
ku
 can in essence be a replacement for human
in
R
beings in difficult situations
&
ra
 can be physically merged with human
Lak
beings to replace failed body parts or to
in
ch
create cyborgs
Sa
6

7. Cyborgs
Mostly Sci-fi
ix it
D
inku
R
a &
r
Lak
hi n
c
Sa
7

8. Traditional approaches
 Mathematical ix it
models:
Black boxes,D
u number
i nk
crunching.
R
&
 Rule-based systems
ra(crisp & bivalent):
Lak Large rule bases.
hi n
c
Sa
8

9. Soft computing (SC)
Objective: ix it
D reasoning
ku
Mimic human (linguistic)
in
R
&
ra
Main constituents:
 Lak systems
Fuzzy
n
hi Neural networks
c
Sa  Evolutionary computing
 Probabilistic reasoning
9

10. Soft Computing:Definition
Soft computing is a term applied ix it a field
to
D
within computer scienceu which is
nk
characterized by the use Ri inexact solutions
of
&
to computationally-hard tasks such as the
ra
ak
solution of NP-complete problems, for
L
which an hin solution cannot be derived
exact
ac
in polynomial time.
S
en.wikipedia.org/wiki/Soft_computing
10

11. Hard Computing vs Soft Computing
 Hard computing ix it
D
 Real-time constraints
n ku
 Need of accuracy and precisioniin calculations and
R
outcomes &
ra
 Useful in critical systems
 Soft computing L ak
in
ch
 Soft constraints
Sa
 Need of robustness rather than accuracy
 Useful for routine tasks that are not critical
11

12. Hard Computing vs Soft Computing
 Soft computing differs from conventional
ix it (hard)
D
computing in that it is tolerant of the following


Imprecision
Uncertainty inku
 Partial truth, and R
 Approximation.
a &
r
ak
 In effect, the role model for soft computing is the human
mind.
n L
hi
The guiding principle of soft computing is:

c
Sa
Exploit the tolerance for imprecision, uncertainty, partial
truth, and approximation to achieve tractability,
robustness and low solution cost.
12

13. Constituents of SC
 Fuzzy systems => imprecision ix it
D
 Neural networks => learning
in ku
R
 Probabilistic reasoning => uncertainty
&
ra
ak
 Evolutionary computing => optimization
L
in
a ch 24,000 publications as of today
SOver
13

14. SC: a user-friendly approach
ix it
D
inku
R
Soft computing
&
approach
r a
ak
Linguistic world
Soft data
n L
Mathematical world Interpretations
hi
Hard data Understanding
Quantitative methods Explanations
c
Sa
Bivalent reasoning Qualitative methods
Bivalent or multivalent
reasoning
Phenomenon under study
14

15. Advantages of SC
 Models based on human reasoning. ix it
D
 Closer to human thinking
inku
 Models can be R
&
 linguistic
ra
ak
 simple (no number crunching),
L
in
 comprehensible (no black boxes),
ch computing,
Sa
 fast when
 effective in practice.
15

16. SC today (Zadeh)
 Computing with words (CW) ix it
D
ku
 Theory of information granulation
in
(TFIG) R
&
ra
 Computational theory of perceptions
(CTP) Lak
in
ch
Sa
16

17. Possible SC data & operations
 Numeric data: ix it
D
5, about 5, 5 to 6, about 5 to ku6
 Linguistic data: Rin
& medium or bad
cheap, very big, notahigh,
r
 Functions & L ak
relations:
n
if(x), fairly similar, much greater
ch
f(x), about
Sa
17

18. Neural networks (NN, 1940's)

x it
Neural networks offer
imethod to
D
a powerful
ku
explore, classify, and
in patterns in
R
identify
& data.
ra 
Neurons
L ak Neuron: y=Σwixi
n
Inputs Outputs
hi
(1 layer)
c
Sa
Walter Pitts
Warren S.
McCulloch 18

19. Machine learning (supervised)
ix it
 Pattern recognition
Orange based u D
on training
i nk
data.
R
Instructor a & Classification

r
Lak supervised by
instructor.
hi n
c  Neural (crisp or
?
Sa Apple
fuzzy), neuro-fuzzy
and fuzzy models.
19

20. Machine learning (unsupervised)

ix it
Pattern recognition
Orange based u D
on training
nk
data.
i
Mango  R
Classification based
& on structure of data
ra (clustering).
Lak
Apple
in  No instructor
a ch  Neural (crisp or
S fuzzy), neuro-fuzzy
Labeling
and fuzzy models.
20

21. Fuzzy systems (Zadeh, 1960's)

(computer environments) ix it
Deal with imprecise entities in automated environments
D

inku
Based on fuzzy set theory and fuzzy logic.

R
Most applications in control and decision making
a &
r
Lak
hi n
c
Sa
Omron’s fuzzy processor
Lotfi A. Zadeh 21

22. SC applications: control
 Heavy industry
ix it
 Matsushita, Siemens
 robotic arms, humanoid robots

u D
Home appliances
 k
 Canon, Sony, Goldstar, Siemens
n refrigerators,
iAutomobiles cameras
washing machines, ACs,
R

a &  Nissan, Mitsubishi, Daimler-
r Chrysler, BMW, Volkswagen
Lak  Travel Speed Estimation, Sleep
Warning Systems, Driver-less cars
hi n  Spacecrafts
 NASA
c
Sa
 Manoeuvering of a Space
Shuttle(FL), Optimization of Fuel-
efficient Solutions for a
Manoeuvre(GA), Monitoring and
Diagnosis of Degradation of
Components and Subsystems(FL),
Virtual Sensors(ANN)
22

23. SC applications: business
 supplier evaluation for  hospital stay ix
it
prediction,
D
ku
sample testing,  TV commercial slot
 customer targeting,
 sequencing, R in matching,
evaluation,
 address
 scheduling, & fuzzy cluster analysis,
a 
r
 optimizing R&D
 projects, Lak  sales prognosis for mail order
house,
 knowledge-based
hi n (source: FuzzyTech)
c
prognosis,Sa
 fuzzy data analysis
23

24. SC applications: finance
 Fuzzy scoring for mortgage applicants,
ix it
 creditworthiness assessment,
D
ku
 fuzzy-enhanced score card for lease risk assessment,
in
 risk profile analysis, R
 insurance fraud detection,
a &
r
ak
 cash supply optimization,
L
hi n
 foreign exchange trading,
c
Sa
 insider
 trading surveillance,
 investor classification etc.
Source: FuzzyTech 24

25. SC applications: robotics
ix it
D
inku
R
a &
r
Lak
hi n
c
Sa
25

26. SC applications: others
ix it
 Statistics D
 Social sciences
inku
R
 Behavioural sciences
a &
r
 Biology
Lak
 Medicine
hi n
c
Sa
26

27. (Neuro)-fuzzy system construction
it
ixExperts
Training Fuzzy rules D
data (SOM, c-means ku
etc.)
Rin
&
ra
Control ak
System
Levaluation
Tuning
(NN)
data
in
ch (errors)
Sa
New system
27

28. Model construction (mathematical)
 Mathematical models are functions. Deep knowledge on
mathematics.
ix it
 D
If non-linear (eg. NN), laborious calculations and computing.
 Linear models can be too simplified.
inku
 How can we find appropriate functions? R
1,2
a &
r
Lak 1
hi n 0,8
c
Y=1-1./(1 + EXP(-2*(X-5)))
Sa
0,6
Y
0,4
0,2
0
0 2 4 6 8 10 12
X 28

29. Model construction (trad. rules )
If 0<x<1, then y=1

ix it
Rule for each input. => Large rule bases.
 Only one rule is fired for each input.
If 1<x<2, then y=0.99 D
ku
:  Coarse models.
If 8<x<10, then y=0
1,2
Rin
1
a &
r
ak
0,8
If 0<x<1, then y=f(x)
If 1<x<2, then y=g(x)
n L 0,6
Y
:
c hi 0,4
Sa
If 8<x<10, then y=h(x)
0,2
0
0 2 4 6 8 10 12
X
29

30. Model construction (SC/fuzzy)
 Approximate values

=> Small rule bases. ix it
Rules only describe typical cases (no rule for each input).
D
ku
 A group of rules are partially fired simultaneously.
in
R
&
1,2
If x≈0, then y≈1
r a
ak
1
If x≈5, then y≈0.5
L
0,8
If x≈10, then y≈0
hi n 0,6
Y
c
Sa
0,4
0,2
0
0 2 4 6 8 10 12
X
30

31. SC and future
ix it be
SC and conventional methods should
D
used in combination.
inku
R
&
ra
Lak
in
ch
Sa
31

32. Sources of SC
 Books: ix it
D
ku
www.springer.de/cgi-bin/search_book.pl?series=2941,
www.elsevier.com/locate/fss, Rin
www.springer.de/cgi-bin/search_book.pl?series=4240,
a &
www.wkap.nl
r
 Others:
Lak
hi n
http://http.cs.berkeley.edu/projects/Bisc/bisc.memo.html
c
Sa
32

33. References
it
ix New York,
1. D
J. Bezdek & S. Pal, Fuzzy models for pattern recognition (IEEE Press,
2.
1992).
in
L. Zadeh, Fuzzy logic = Computing with words, IEEE ku Transactions on Fuzzy
L. Zadeh, From Computing with Numbers RComputing with Words -- From
Systems, vol. 2, pp. 103-111, 1996.
3.
& to
on Circuits and Systems, 45, 1999, ra
Manipulation of Measurements to Manipulation of Perceptions, IEEE Transactions
L. Zadeh, Toward a theory of k
105-119.
4.
a fuzzy information granulation (1997)its111-127. in
and centrality
L theory and its applications (Kluwer, Dordrecht, 1991).
human reasoning and fuzzy logic, Fuzzy Sets and Systems 90/2
in
ch
5. H.-J. Zimmermann, Fuzzy set
Sa
33