Professional Documents
Culture Documents
Dr. A. K. Santra
Senior Professor
School of Information Technology
VIT University, Vellore
1
INTRODUCTION
2
INTRODUCTION Contd
3
INTRODUCTION Contd
4
NEURON
5
SINGLE BIOLOGICAL NEURON AND ITS MODEL
6
MULTIPLE NEURONS AND THEIR MODEL REPRESENTATION
7
NEURAL NETWORKS
NEURONS ARE CONNECTED IN LAYERS
EACH LAYER RECEIVES INPUT FROM THE PRECEEDING
LAYER FEED FORWARD
- INPUT LAYER
- MIDDLE / HIDDEN LAYER (S)
- OUTPUT LAYER
OUTPUT LAYER
MIDDLE LAYERS
INPUT LAYER 8
UNSUPERVISED LEARNING METHODS
CLASSICAL STATISTICS
CLUSTERING
k - MEANS
FEATURE MAPS
INFOMAX PRINCIPLE
MAXIMUM ENTROPY
COMPETITIVE LEARNING PARADIGMS NEURAL NET
9
SUPERVISED LEARNING METHODS
BACK PROPAGATION
PARZEN WINDOWS
NEAREST NEIGHBOUR NORMS
PROJECTION PURSUIT
GAUSSIAN MIXTURE
GROUP METHOD OF DATA HANDLING ( GMDH )
MULTIVARIATE ADAPTIVE REGRESSION SPLINE
HYPER SPHERE CLASSIFIERS
CLASSIFICATION AND REGRESSION TREES
10
BENEFITS
11
APPLICATIONS OF ARTIFICIAL NEURAL NETWORKS
IRON - MAKING
STEEL - MAKING
DATA MINING AND CLUSTERING
IMAGE PROCESSING (2D / 3D)
- FACE RECOGNITION
- VEHICLE CLASSIFICATION
- SATELLITE OBJECT RECOGNITION
- TARGET (ENEMY) DETECTION AND RECOGNITION
WEATHER FORECASTING
CROP YIELD PREDICTION
.
.
.
FACILITY LOCATION PROBLEMS
12
Applicatin of ANN in Steel Industry
Optimum Stove Heating and Changing, Optimum Driving
Rate of Blast Furnace
Adjusting End Point Composition Based on In-Blow
Sampling in Oxygen Furnace
Optimum Scheduling and Sequencing in Continuous
Casting
Optimization of Annealing Parameters based on Material
and Equipment Characteristics
Desulphurization of Hot Metal in Torpedo
Optimization of Refractory Lining Life in Basic Oxygen
Furnace
13
Case Study
Optimization of Refractory Lining Life in Basic Oxygen
Furnace (BOF)
Various Factors that affect the Campaign Life of
Refractory Lining in BOF :
Vessel Design, Refractory Quality, Operating
Parameters, etc.
Solution by ANN, Regression Tech. & GMDH
A Multilayer Feedforward Architectures of Neural Nets
used
8 Parameters and 10 Campaign Data used
For Training: 8 Data Sets & For Prediction: 2 Data Sets
14
Models Output:
Method Actual Life Predicted % Error
(Hrs.) Life (Hrs.)
Regression 595 558 6.1
(Linear) 539 623 15.5
Regression 595 677 13.4
(Exponential) 539 732 35.5
GMDH 595 660 10.92
539 558 3.5
ANN 595 583 2.02
539 540 0.18
15
Discussion
Prediction Error is very low in case of ANN.
Iron and Steel making process is very complicated.
Practically impossible to describe the process
mathematically.
Where the classical mathematical model has got its
inherent limitations, ANN technique provides a
better solution to Prediction / Process Optimization
problem.
ANN techniques are becoming very popular in the
recent years for their process optimizing capability.
16
FACILITY LOCATION PROBLEMS
What is a Facility Location Problem ?
Work of Cabot, Francis and Stray[1], Pritsker and Ghare [2], Francis
[3], White [4], Eyster and White [5], Santra and Nasira [6,7], Hartstein
[8], Mortin Moller [9], Trentin and Gori [10], Sarkar and Tassiulas [11],
Katz [12], and Kuhn [13]. Multifacility Location Problem (MLP) has
been solved without considering Area Constraint, except by Santra
and Nasira [6,7], by assuming that the entire space is available as
solution space.
17
OBJECTIVE
18
MATHEMATICAL FORMULATION AND ANALYTICAL
SOLUTION OF THE PROBLEM:
m n 1/2
2 2
w [ ( x x ) ( y y i ) ] (2.1)
ji j i j
i 1j 1
2 2 2
Subject to:
(xj x ) ( yj y) R 0 (2.2)
x j 0, y j 0 (j 1,2, ., n)
where n = number of new facilities,
m = number of existing facilities,
(xj, yj) = co-ordinates of the jth new facility,
( x , y ) = co-ordinates of the ith existing facility,
i i
( x , y )= co-ordinates of the centre of gravity of the
existing facilities and
Wji = cost per unit distance between new
facility j and existing facility I,
R = radius of the circular area.
19
MATHEMATICAL FORMULATION AND ANALYTICAL
SOLUTION OF THE PROBLEM: (Contd.)
Without any loss of generality we can always suitably choose our origin
2 2 2
such that the circular area given by (x x) (y y) R
will always be in the first quadrant of the xy-plane. Then this circle will
divide the plane into three sets of points, viz.,
We use Kuhn -Tucker conditions to get the solution of the problem for which we
construct the auxiliary function as :
2 2 2 2
h(x, y) f(x, y) [(x x ) (y y ) R ] (2.3)
n 2 2 2 2
j [(x j x ) (y j y) R j ] (2.4)
j1
where j2 are the artificial variables and are given by
2 2 2 2
j [(x j x ) (y j y ) R ].
21
SOLUTION PROCEDURE : (contd.)
h f
2 (y j y) j 0 (2.6)
y j y j
2 2 2
j[( x j x ) ( y j y ) R ] 0 (2.7)
2 2 2 (2.8)
( xj x ) ( yj y) R 0
j 0 ( j = 1, 2, , n).
22
SOLUTION PROCEDURE : (contd.)
^
f m w ji (y j y i )
(2.10)
y j i 1 E ji
^ ^
2 2 1/2
E ji [(x j x i ) (y j y i ) ] (2.11)
There are only two possible cases to be examined, viz., (i) when j = 0 and
(ii) when j 0 . First of all we consider the case when j = 0.
23
SOLUTION PROCEDURE : (contd.)
Case - I : j = 0 (j = 1,2,..,n)
f f
As j = 0 the equations (2.5) and (2.6) reduce to 0 and 0
x j y j
respectively which in view of (2.9) and (2.10) lead to
^
m w ji (x j x i )
0 (2.12)
i1 E ji
^
m w ji (y j y i )
0 (2.13)
i 1 E ji
24
SOLUTION PROCEDURE : (contd.)
m ^
w ji x i
x j i 1 (2.14)
m w ji
i 1 E ji
m ^
w ji y i
y j i 1 (2.15)
m w ji
i1 E ji
m ^
w ji x i
(N 1)
xj i 1 (2.16)
m w ji
i 1 E (N)
ji
m ^
w ji y i
(N 1)
yj i 1 (2.17)
m w ji
i 1 E (N)
ji
(N) (N) ^ 2 (N) ^ 2 1/2
E ji [(x j x i ) (y j y i ) ] (2.18)
m ^
w ji x i
(0) i 1
xj m (2.19)
w ji
i 1
m ^
w ji y i
(0) i 1
yj m (2.20)
w ji
i 1
From the structure of the equations (2.16) and (2.17) it is clear that they represent
n independent iterative schemes from each of which we get the optimum
location of a single new facility.
27
SOLUTION PROCEDURE : (contd.)
For example, we may have the iterative scheme for any one say, (x1, y1)
of the n new facilities as follows:
m ^
w1i x i
(N 1)
x1 i 1 (2.21)
m w1i
i 1 E (N)
1i
m ^
w1i y i
(N1)
y1 i 1 (2.22)
m w1i
i 1 E (N)
1i
with starting initial solution as (x (0) , y ( 0) ) given by (2.19) and (2.20) for j = 1,
(N) 1 1
where E is given by
(N) ^ 2 (N) ^ 2
1i
(N) 1/2
E1i [(x1 x i ) (y1 y i ) ] (2.23)
28
SOLUTION PROCEDURE : (contd.)
The solution (xj, yj) ( j = 1, 2, , n ) obtained by the above iterative procedure has
to be tested whether it satisfies the area constraint (2.2).
If it satisfies (2.2), the problem is solved and (xj, yj) ( j = 1, 2, , n ) give the
optimum locations for the new facilities sought.
29
SOLUTION PROCEDURE : (contd.)
f f
( y j y) ( xj x) (2.25)
x j y j
which in view of (2.9) and (2.10) takes the form
^ ^
m w ji (x j x i ) m w ji (y j y i )
(y j y ) (x j x ) (2.26)
i 1 E ji i 1 E ji
30
SOLUTION PROCEDURE : (contd.)
Further simplification of (2.26) by using (2.11) and the value of yj from (2.24)
finally leads to:
^
m w ji (x j x i )
R 2 (x x ) 2
j i 1 Tji
x j x (2.27)
^
2 2
w
m ji [ y y i R (x j x ) ]
i 1 Tji
^ 2
2 2 2 1/2
Tji [(x j x ) { y yi R (x j x ) } ]
31
SOLUTION PROCEDURE : (contd.)
^
w (y y i )
2 2 m ji j
R (y j y) (2.27A)
i 1 K ji
y j y
^
2 2
w
m ji [ x x i R (y j y ) ]
i 1 K ji
^ 2
2 2 2 1/2
K ji [( y j y ) { x x i R (y j y) } ] .
32
SOLUTION PROCEDURE : (contd.)
To solve the set of non-linear equations represented by (2.27) or (2.27A)
we use the following iterative scheme:
(N)
^
(N) m w ji (x j xi )
2 2
R (x j x ) (N)
i 1 Tji
(N 1)
xj x (2.28)
^ (N)
2 2
w
m ji [ y y i R (x j x ) ]
(N)
i1 Tji
(N) (N)
^ (N) 2
2 2 2 1/2
Tji [(x j x ) {y yi R (x j x ) } ]
and the superscripts denote the iteration number. We take a feasible initial
solution for rapid convergence of the iterative scheme.
For example, we may have the iterative scheme for the x co-ordinate of any one
of n new facilities, say, x1 of the new facility (x1, y1) as follows :
(N)
^
2 (N) m w1i (x1
2 xi )
R (x1 x ) (N)
(N 1)
i 1 T1i
x1 x
^ (N) (2.29)
2 2
w
m 1i [ y y i R (x1 x ) ]
(N)
i 1 T1i
(N) (N)
^ (N) 2
2 2 2 1/2
T1i [(x1 x ) {y y i R (x1 x ) } ]
It may be pointed out that after finding the values of all xj ( j = 1, 2, , n ) one
need not calculate the values of yj ( j = 1, 2, , n ) by using iterative schemes as
those can be found directly from the equation (2.24) with the help of the
determined values of xj.
We may mention that the problem where the new facilities have to be located on
the circumference of the circle can be solved by utilizing the above algorithm.
34
SOLUTION OF MULTIFACILITY LOCATION PROBLEM UNDER
CIRCULAR AREA CONSTRAINT USING ANN TECHNIQUES:
35
NEURAL NETWORK STRUCTURE
x1
y1
x2
z1
x3
x4
o1
o2
o3
o4
x8 z25
y50
x9
We have solved a number of problems analytically which have been used for
training the nets. Each of these problems will have the following 9 basic input
parameters with their usual meanings as mentioned in (2.1) and (2.2) :
^ ^ ^ ^
(i) x1 (ii) x2 (iii) y1 (iv) y2 (v) w11
37
NUMERICAL EXAMPLE (contd.)
By using these input parameters we have found the locations of 2 new facilities
viz. (i) (x1, y1) and (ii) (x2, y2) as the solution.
The same set of examples has been considered for training the net. Once the
training is complete, we have tested the prediction capabilities of the trained net
by using a new set of data patterns.
The predicted output values obtained from the net have been compared with
those obtained through the Analytical method and it is found that they compare
well within the acceptable limit.
38
Table 1 : Training Data Set
^ ^ ^ ^
S.No. w11 w12 w21 w22 R
x1 x2 Y1 y2
1 2 1 1 2 5 3 7 8 0.5
2 1 1 1 1 2 4 3 1 0.5
3 2 1 1 2 5 3 7 10 0.5
4 2 1 1 1 8 2 4 5 0.5
5 2 1 1 2 7 3 9 1 0.5
6 1 1 2 1 3 3 9 4 0.5
7 2 1 1 2 10 4 8 3 0.5
8 2 1 1 2 4 6 2 9 1
9 1 1 1 1 1 1 1 1 1
10 3 1 1 2 7 3 7 4 1
11 3 3 1 2 5 3 7 8 1
12 2 1 1 2 7 2 7 9 1
13 2 1 1 2 7 7 6 7 1
14 2 1 1 2 3 4 1 2 1
15 2 1 1 2 5 3 7 8 1.5
39
Table 1 : Training Data Set (Contd.)
^ ^ ^ ^
S.No. w11 w12 w21 w22 R
x1 x2 Y1 y2
16 1 1 2 2 5 2 8 6 1.5
17 2 3 1 1 6 5 8 9 1.5
18 2 1 1 2 5 3 8 8 1.5
19 2 1 1 1 7 5 6 9 1.5
20 1 1 1 1 5 3 8 3 1.5
21 3 1 2 2 5 6 7 8 1.5
22 1 2 3 4 5 6 7 6 2
23 1 1 1 1 4 7 2 9 2
24 1 1 2 1 8 3 2 1 2
25 2 1 1 2 7 2 7 9 2
26 2 1 1 2 5 3 7 10 2
27 1 1 2 1 3 3 9 4 2
28 1 1 1 1 5 3 8 8 2
29 2 1 1 2 5 3 7 8 2.5
30 1 1 1 1 2 4 3 1 2.5
40
Table 1 : Training Data Set (Contd.)
^ ^ ^ ^
S.No. w11 w12 w21 w22 R
x1 x2 Y1 y2
31 2 1 1 1 3 4 1 2 2.5
32 1 1 2 1 3 3 9 4 2.5
33 1 1 1 2 5 6 2 10 2.5
34 2 1 1 2 7 7 6 7 2.5
35 3 3 1 2 5 3 7 8 2.5
36 2 1 1 2 2 5 6 8 3
37 1 1 1 1 2 4 3 1 3
38 2 1 1 1 3 4 1 2 3
39 1 1 2 1 3 3 9 4 3
40 1 1 1 2 5 6 2 10 3
41 2 1 1 2 7 7 6 7 3
42 2 1 1 2 4 6 2 9 3
43 2 1 1 2 2 5 6 8 3.5
44 1 1 1 1 2 4 3 1 3.5
45 1 1 1 1 3 4 1 2 3.5
41
Table 1 : Training Data Set (Contd.)
^ ^ ^ ^
S.No. w11 w12 w21 w22 R
x1 x2 Y1 y2
46 1 1 2 1 3 3 9 4 3.5
47 2 2 1 2 6 6 6 6 3.5
48 3 3 1 2 5 3 7 8 3.5
49 2 1 1 2 4 6 2 9 3.5
50 2 1 1 2 7 7 3 7 4
51 2 2 1 2 5 3 8 8 4
52 1 1 1 1 2 4 3 1 4
53 1 1 1 1 3 4 1 2 4
54 3 1 1 2 5 6 7 8 4
55 2 1 1 2 5 6 2 10 4
56 3 3 1 2 5 3 7 8 4
57 2 1 1 2 7 7 3 7 4.5
58 1 1 1 1 2 4 3 1 4.5
59 2 3 1 1 6 5 8 9 4.5
60 1 1 1 1 3 4 1 2 4.5
42
Table 1 : Training Data Set (Contd.)
^ ^ ^ ^
S.No. w11 w12 w21 w22 R
x1 x2 Y1 y2
61 2 1 1 2 4 6 2 9 4.5
62 2 2 1 2 6 6 6 6 4.5
63 3 3 1 2 5 3 7 8 4.5
64 2 1 1 2 7 7 3 7 5
65 3 1 1 2 5 6 7 8 5
66 1 1 1 1 2 4 3 1 5
67 2 1 1 3 4 3 7 10 5
68 1 1 1 1 3 4 1 2 5
69 2 1 1 2 5 6 2 10 5
70 3 3 1 2 5 3 7 8 5
43
Table 2 : Target Data Set (Analytical output)
S.No. x1 x2 y1 y2
1 1.57 1.71 1.99 1.95
2 1.08 1.25 1.49 1.43
3 1.57 1.93 1.99 1.74
4 1.55 1.97 1.49 1.16
5 1.56 1.14 1.99 1.85
6 1.58 1.82 1.49 1.37
7 1.54 1.68 1.99 1.96
8 1.59 1.81 2.49 2.44
9 1.47 1.25 1.88 1.96
10 2.09 2.77 2.49 2.13
11 2.14 2.13 3.48 2.72
12 1.61 1.98 2.49 2.37
13 1.56 2.40 2.49 1.92
14 1.63 2.29 2.49 2.11
15 1.68 2.79 2.98 2.25 44
Table 2 : Target Data Set (Analytical output) (Contd.)
S.No. x1 x2 y1 y2
16 1.71 0.17 2.98 2.20
17 1.64 1.62 3.49 2.14
18 1.70 2.79 2.98 2.26
19 1.62 2.95 2.49 1.38
20 1.18 1.25 2.48 2.47
21 2.63 1.07 2.99 1.97
22 2.19 0.47 4.99 4.29
23 1.18 1.25 2.99 2.98
24 1.67 3.41 2.99 1.57
25 1.72 3.19 3.48 2.56
26 1.74 3.18 3.48 2.57
27 1.82 3.36 2.97 1.72
28 1.24 1.25 2.98 2.98
29 1.80 3.91 3.98 2.14
30 1.40 1.25 3.46 3.48 45
Table 2 : Target Data Set (Analytical output) (Contd.)
S. No. x1 x2 y1 y2
31 1.84 3.91 3.47 1.64
32 1.90 3.94 3.46 1.53
33 1.22 3.14 3.98 2.78
34 1.67 3.81 3.99 2.44
35 2.32 2.13 4.97 2.72
36 1.89 4.48 4.47 1.81
37 1.47 1.25 3.96 3.98
38 1.91 4.41 3.97 1.70
39 1.98 4.32 3.96 2.00
40 1.26 3.45 4.48 3.21
41 1.71 4.36 4.49 2.38
42 1.79 3.79 4.48 3.42
43 1.96 4.79 4.96 2.69
44 1.55 1.25 4.45 4.49
45 1.48 1.25 4.46 4.49 46
Table 2 : Target Data Set (Analytical output) (Contd.)
S. No. x1 x2 y1 y2
46 2.06 4.90 4.45 1.81
47 1.78 1.25 5.48 5.49
48 2.45 2.13 5.97 2.72
49 1.83 4.28 4.98 3.61
50 1.78 5.35 5.49 2.57
51 1.99 1.25 5.96 5.99
52 1.63 1.25 4.95 4.99
53 1.55 1.25 4.96 4.99
54 2.35 2.32 5.48 1.84
55 1.85 5.00 5.48 3.43
56 2.51 2.13 6.46 2.72
57 1.81 5.49 5.98 3.57
58 1.70 1.25 5.44 5.49
59 1.91 1.62 6.48 2.14
60 1.61 1.25 5.45 5.49 47
Table 2 : Target Data Set (Analytical output) (Contd.)
S. No. x1 x2 y1 y2
61 1.93 5.27 5.97 3.95
62 1.87 1.25 6.48 6.49
63 2.57 2.13 6.96 2.72
64 1.85 6.27 6.48 2.96
65 2.43 2.32 6.48 1.84
66 1.77 1.25 5.93 5.99
67 2.18 1.46 6.95 2.24
68 1.68 1.25 5.95 5.99
69 1.94 6.08 6.48 3.50
70 2.63 2.13 7.45 2.72
48
Table 3 : Testing Data Set (Resilient Back Propogation)
^ ^ ^ ^
S.No. x w11 w12 w21 w22 R
1 x2 Y1 y2
1 2 3 1 1 9 6 7 8 1.5
2 2 1 1 2 5 3 7 8 2
3 2 1 1 2 5 3 7 10 3
4 2 1 1 2 5 3 7 10 3.5
5 2 1 1 2 7 7 3 7 3.5
49
Table 4 : Analytical and ANN (Resilient) Model Output
y1
x1 x2 y2
S.No.
Anal Ann Anal Ann Anal Ann Anal Ann
50
Table 5 : Error Percentage (Resilient)
x1 x2 y1 y2
S. No.
%Error %Error %Error %Error
51
Graph 1 : Analytical and ANN (Resilient) Model Output (X1)
52
Graph 2 : Analytical and ANN (Resilient) Model Output (X2)
53
Graph 3 : Analytical and ANN (Resilient) Model Output (Y1)
54
Graph 4 : Analytical and ANN (Resilient) Model Output (Y2)
55
Table 6 : Output Locations (Resilient)
56
Table 7: Testing Data Set (Fletcher - Reeves Algorithm)
^ ^ ^ ^
S.No. x w11 w12 w21 w22 R
1 x2 Y1 y2
1 1 1 1 1 6 3 2 6 0.5
2 2 1 1 2 5 3 7 8 2
3 2 1 1 2 3 4 1 2 2.5
4 2 1 1 2 5 3 7 10 3
5 2 1 1 2 7 2 7 9 3
57
Table 8 : Analytical and ANN (Fletcher-Reeves) Model Output
y1
x1 x2 y2
S.No.
Anal Ann Anal Ann Anal Ann Anal Ann
58
Table 9 : Error Percentage (Fletcher-Reeves)
x1 x2 y1 y2
S. No.
%Error %Error %Error %Error
59
Graph 5 : Analytical and ANN (Fletcher-Reeves) Model Output (X1)
60
Graph 6 : Analytical and ANN (Fletcher-Reeves) Model Output (X2)
61
Graph 7 : Analytical and ANN (Fletcher-Reeves) Model Output (Y1)
62
Graph 8 : Analytical and ANN (Fletcher-Reeves) Model Output (Y2)
63
Table 10 : Output Locations (Fletcher-Reeves)
64
CONCLUSION
The present work deals with the finding of solutions analytically and
through ANN technique for any given multifacility location problem
having Interactions between Sources and Destinations with Circular
area constraint.
65
REFERENCES
[1] Cabot A.V., R. L . Francis and M .A. Stary, A Network Flow Solution to a
Rectilinear Distance Facility Location Problem, AIIE Transactions, Vol.
2, No. 2, 1970, pp. 132-141.
68
69
70