Professional Documents
Culture Documents
Corporacin Mexicana de Investigacin en Materiales (COMIMSA), Calle Ciencia y Tecnologa, # 790, Frac. Saltillo 400, Saltillo, Coahuila, Mexico
Centro de Innovacin, Investigacin y Desarrollo en Ingeniera y Tecnologa (CIIDIT), Universidad Autnoma de Nuevo Len, Km. 10 de la nueva carretera al
Aeropuerto Internacional de Monterrey, CP 66600, PIIT Monterrey, Apodaca, Nuevo Len, Mexico
c
Facultad de Sistemas, Universidad Autnoma de Coahuila, Ciudad Universitaria, Carretera a Mxico Km 13, Arteaga, Coahuila, Mexico
b
art ic l e i nf o
a b s t r a c t
Article history:
Received 24 May 2012
Received in revised form
5 April 2013
Accepted 3 June 2013
Available online 3 July 2013
A Hybrid Learning Process method was tted into a RBF. The resulting redesigned RBF intends to show
how to test if the statistical assumptions are fullled and to apply statistical inference to the redesigned
RBFNN bearing in mind that it allows to determine the relationship between a response (to a process)
and one or more independent variables, testing how much each factor contributes to the total variation
of the response is also feasible. The results show that statistical methods such as inference, Residual
Analysis, and statistical metrics are all good alternatives and excellent methods for validation of the
effectiveness of the Neural Network models. The foremost conclusion is that the resulting redesigned
Radial Basis Function improved the accuracy of the model after using a Hybrid Learning Process;
moreover, the new model also validates the statistical assumptions for using statistical inference and
statistical analysis, satisfying the assumptions required for ANOVA to determine the statistical
signicance and the relationship between variables.
& 2013 Elsevier Ltd. All rights reserved.
Keywords:
Radial Basis Function
Statistical inference
ANOVA
Residual Analysis
Hybrid Learning Process
1. Introduction
Articial Intelligence or Intelligent Systems are derived from
different applications or techniques, such as: fuzzy logic, Neural
Networks, evolutionary algorithms and hybrid systems (Benyounis
and Olabi, 2008; Paliwal and Kumar, 2009). Intelligent systems
attempt to imitate human reasoning for problem-solving, performing tasks and decision-making (Arbib, 2003). There are different
types of intelligent systems for modeling and optimizing industrial
processes (Benyounis and Olabi, 2008), such as articial neural
network backpropagation, support vector machines, adaptive resonance theory networks and Radial Basis Function Neural Network
(RBFNN) (Martn del Bro and Sanz, 2007; Nelles, 2001). Neural
Networks are able to process large amounts of information using
mathematical models. A Neural Network (NN) is an interconnection
of simple processing elements that represents the function of a
single neuron. The interconnection is made by articial synapses
called weights. The adjustment of these weights modies the
performance of the neural network. The NN has the ability to
1882
2. Methods
2.1. Radial basis function neural network
The structure in a RBF, in its basic form, includes three totally
different layers: input, hidden and output layers. The output layer
neurons are linear. The hidden layer neurons calculate the difference between the vector of inputs and the centroids (Haykin,
1999). This difference applies a radial function with a Gaussian
shape mainly (Gregori and Lightbody, 2008), but it has the
advantage of being able to use other radial functions. The function
of transfer radial in Gaussian type adopts the following form:
Gjjxt i jj expjjxt i jj2 :
yx wGjjxt i jj b
i1
4
GxN t 1 GxN t 2
3
Gx1 t m
7
Gx2 t m 7
7:
7
5
GxN t m
Step
Step
Step
Step
Step
Step
Step
yywGy
2 :
yy ni 1 yi =n
1883
where F is the cumulative distribution function of Normal distribution. And the data Y i are ordered observations.
2.3.2. The residual should be Equal Variances
Variances of the populations must be equal (or close) across all
residual values for a given factor, that is, they must have Homogeneity throughout the variances. Proving this assumption, statistical test as the BreuschPagan test and the White test are
employed (among others). These tests are used as evidence of
greater statistical power, the BreuschPagan test (1979) in particular due to its detachment from the functional form, and the
White test which is neither dependent of the functional form nor
the type of data distribution (Greene, 2003).
2.3.2.1. BreuschPagan. In the work of Breusch and Pagan (1979) a
test is designed that does not depend on the number of missing
observations or the identication of the variable that generates
heteroscedasticity for the order, or the functional form (Kennedy,
2008). This test covers a broad spectrum contrast, is simple and is
based on the Ordinary Least Square (OLS) residuals. It is assumed
for the model
Y X U
10
11
13
s2 u^ i =Nk:
14
1884
Varb. From this point, White built a contrast test to get through
the coefcient of determination R2 of an auxiliary model representing the errors against the original explanatory variables and
their squares and cross products. Next a statistic is built distributed as a Chi square with p1 degrees of freedom NR2 2p1 . Where
p is the number of regressors in the auxiliary model excluding the
constant. The statistic provides a test of hypothesis that all slopes
of the auxiliary model are zero, i.e.
H 0 2 3 4 p 0:
15
et et1
t2
16
e2t
t1
18
Model
Model
Model
SSError y0 yni 1 yi 2 =n
19
wGyni 1 yi 2 =n
20
21
SSError
yywGy
:
Na
Na
23
:
F 0 RBF Model
SSError =Na
MSError
24
1885
overall quality of the RBF model. If the error terms, i , in the RBF
model are normally and independently distributed, are distributed
as t random variables with n2 degrees of freedom. This leads to
the following denition of 100 1% condence intervals on
the weights.
q
q
wj t =2;np s2 C jj wj wj t =2;np s2 C jj :
27
An important application of model is predicting new or future
observations Y corresponding to a specied level of the variable
x. yx m
i 1 wGjjxt i jj b is the point estimator of the new
or future value of the response yxj dj .
Now consider obtaining an interval estimate for this future
observation yxj dj . This new observation is independent of the
observations used to develop the RBF model. The prediction in the
RBF model is given by y^ Gw then a point estimate of the future
observation y^ 0 under g 0 is y^ 0 g 0 w. A 100 1% prediction
interval on a future observation at the value g 0 is given by
q
y^ 0 t =2;np s2 1 g0 GG1 g 0 Y 0 y^ 0
q
t =2;np s2 1 g0 GG1 g 0 :
28
There are methods for the analysis and validation of experimental data using statistical techniques such as Residual Analysis,
including statistical metrics (listed below). For example, a metric
used for viewing the performance of the model or validating it is a
R2 given by Eq. (7), as used in the model of RBF. Another metric for
analyzing the model, is R2Adj , with the coefcient of determination
adjusted (Montgomery et al., 2006). The metric adjusted R2Adj
penalizes the addition of terms that are not useful. Cross evaluating with different models is rather encouraged. It is written as
follows:
R2Adj 1
SSError =np
:
SSRBF Model n1
29
wj
wj
:
T 0 q
sewj
s2 C jj
MSError
26
Table 1
Example of ANOVA table.
Analysis of Variance
Source of
variation
Degrees of
freedom
Sum of
squares
Mean
square
Model
SSRBF
MSRBF
MSRBF Model
MSError
Residual error
Total
nk1
n1
SSError
SSTotal
Model
Model
MSError
1 n
SSError
yywGy
y y^ 2
:
np i 1 i i
Na
Na
1 n 2
e :
ni1 i
30
where ei are the residuals. The next metric that is used, it is the
PRESS (Prediction Error Sum Square), which is a measure of how
well a model works to predict new data and is dened in the given
equation.
2
n
ei
PRESS
31
i 1 1hii
where ei are the residuals and hii is the i-th diagonal element of
the hat matrix H, the hat matrix is given by H GGG1 G. The
other statistical information is the R2Prediction , which is the coefcient of determination of the prediction and is given by the
1886
following equation:
PRESS
R2Prediction 1
SSTotal
Table 2
Central composite design in the machining process.
32
R2Prediction
Output variable
X1
X2
X1 X1
X2 X2
X1 X2
Diameter (mm)
2500
2750
3000
2750
2750
2750
2396.45
2750
2750
2500
3103.55
3000
2750
300
275
250
275
275
239.645
275
310.355
275
250
275
300
275
6,250,000
7,562,500
9,000,000
7,562,500
7,562,500
7,562,500
5,742,972.6
7,562,500
7,562,500
6,250,000
9,632,022.6
9,000,000
7,562,500
90000
75625
62500
75625
75625
57429.726
75625
96,320.226
75,625
62,500
75,625
90,000
75,625
750000
756250
750000
756250
756250
659023.75
659023.75
853,476.25
756,250
625,000
853,476.25
900,000
756,250
4.16
4.15
4.07
4.18
4.09
3.81
4.12
4.02
4.06
4.15
4.1
3.89
4.04
centers of the RBF and the distance weighting matrix associated with the hidden layer, are all unknown parameters that
have to be learned. However, the activation functions of the
hidden layer in the regularization RBFNN (Section 3.1.2) are
known, being dened by a set of basis functions centered at the
training data points; the linear weights of the output layer are
the only unknown parameters of the RBFNN.
Hence, in this application (machining process) the structure of
the RBF redesigned includes three layers: 5 inputs, 5 hidden and
1 output layers. Therefore the function used is given by Eq. (6),
therefore RBF does need to calculate and aggregate the widths
(Praga-Alejo et al., 2012).
Degrees of
freedom
Sum of
squares
Mean
square
Model
Residual error
Total
5
7
12
0.125033
0.01489
0.139923
0.025007
0.002127
11.76 0.003
Table 4
Standard error of weights and terms in the machining process.
Term
se (weights)
X1 SpindleSpeed
X2 ToolFeed
X1 X1
X2 X2
X1 X2
Bias
0.00000744
0.00055755
0.00000083
0.00000031
0.00000122
0.40136000
1887
The condence intervals (Eq. (27)) on the weights are mentioned in Table 6.
Parting from such analysis is possible to obtain condence
interval of these parameters wj and their signicant contribution.
Prediction condence interval at the value g 0 is given by (28):
The predictions in the RBF redesigned model are mentioned in
the Table 7:
Considering an interval estimate for future observations is of
critical relevance.
There are methods for the analysis and validation of experimental data using statistical techniques such as Residual Analysis,
including statistical metrics. For example, a metric used to view
the performance of the model or validate it is an R2 given by
Eq. (7), in the model of RBF R2 equal to 89.4%. The results show
that RBF redesigned with GA and Mahalanobis distance is over 89%
in R2 ; it refers to R2 as the amount of variability in the data
explained by the models, i.e. the RBF redesigned accounts for more
than 89% of the variability in the data. The metric is the quantity
used to express the proportion of total variability in the response
accounted for by the model, so that R2 indicates the proportion of
variability in y explained by the model. If the R2 value is very close
to 100% or above 70%, it means that the model will be a good
predictor (Montgomery et al., 2006).
The results resumed in the Table 8 indicates the statistics
metrics in the model that is adjusted by Radial Basis Function
redesigned, the metrics used are: the coefcient of determination
R2 , the coefcient of determination adjusted R2Adj , The Mean Square
Residuals MSRes , the Prediction Error Mean Square PREMS, the
Prediction Error Sum Square PRESS and the coefcient of determination of prediction R2Prediction .
The other statistical method used in the evaluation of the
model, is the PREMS, where the Mean Square Residuals (MSError )
should be greater than PREMS. If the PREMS is less than the MSError
the model is going to be a good predictor; in this case the
application complies with this premise, in which the model can
be used to predict and optimize. The results summarized in Table 8
Table 7
Condence intervals on the predictions in the machining process.
Table 5
Test hypotheses in the inputs in the machining process.
Term
To
To p-value
X1 SpindleSpeed
X2 ToolFeed
X1 X1
X2 X2
X1 X2
Bias
5.06080000
5.40530000
4.62970000
5.35970000
5.55520000
13.14700000
0.00146200
0.00100270
0.00239840
0.00105310
0.00085524
0.00000344
Table 6
Condence intervals on the weights in the machining process.
Lower limit
y^ 0
Upper limit
4.03
3.98
3.93
3.98
3.98
3.65
4.00
3.89
3.98
3.99
3.95
3.76
3.98
4.16
4.10
4.06
4.10
4.10
3.80
4.14
4.03
4.10
4.14
4.09
3.90
4.10
4.30
4.22
4.19
4.22
4.22
3.96
4.27
4.17
4.22
4.28
4.22
4.04
4.22
Table 8
Results in statistical metrics in the machining process.
Weights
Lower limit
wj
Upper limit
w1
w2
w3
w4
w5
w0
0.00002006
0.00433211
0.00000188
0.00000239
0.00000388
4.32766037
0.00003765
0.00301372
0.00000383
0.00000166
0.00000676
5.27673172
0.00005524
0.00169533
0.00000579
0.00000093
0.00000964
6.22580306
Results Statistical
method
Results Statistical
method
R2
89.4
MSError
0.0021
PRESS
R2Adj
81.8
PREMS
0.0011
R2Prediction
Results
0.033
76.2
1888
Table 9
Experimental design in the GMAW process.
Input variables
Electrode
Extension
Heat
input
230
240
240
240
240
240
240
240
240
250
250
250
250
250
250
250
250
260
260
260
260
260
260
260
260
270
15.9
12.7
19.1
12.7
19.1
12.7
19.1
12.7
19.1
15.9
15.9
9.5
15.9
15.9
22.2
15.9
15.9
12.7
19.1
12.7
19.1
12.7
19.1
12.7
19.1
15.9
6670
7566.7
7566.7
5346.3
5346.3
8730.8
8730.8
6168.8
6168.8
6000
10500
7000
7000
7000
7000
5250
8000
8197.3
8197.3
5791.8
5791.8
9458.4
9458.4
6682.9
6682.9
7560
28
26
26
26
26
30
30
30
30
24
28
28
28
28
28
28
32
26
26
26
26
30
30
30
30
28
60
49.5
49.5
70
70
49.5
49.5
70
70
60
40
60
60
60
60
80
60
49.5
49.5
70
70
49.5
49.5
70
70
60
Output
variable
Penetration
1.7
2.8
2
2.1
2.1
2.4
2.2
2.1
2.5
1.4
2.2
3.1
2.4
2.3
2.5
2.7
3
3.1
2.5
1.7
2.6
3.1
2.5
3.3
2.5
2.2
On the other hand, there are methods for the analysis and
validation of experimental data using statistical techniques such as
1889
Considering only the information showed, it could be concluded that the RBFNN is a good option to realize predictions
about the process; nevertheless, the proposed statistical analysis
suggests the opposite (Fig. 2).
Table 12 shows the estimated weights as well as the p-values
for making the signicance proof.
Considering the It could be concluded from Table 12 that the
variables which correspond to w1 and w2 are not statistically
signicant and must not be included in the RBNN model. This
inference provides a great advantage when a RBNN is used for
modeling a real process, because it is possible to decide which
process variables should be monitored and controlled to keep the
process in control with a condence given (95% for this case).
Since the assumptions were assessed, the statistical inference
showed in Table 12 is reliable.
Hence, supported by the statistical analysis, it could be concluded that the tted RBNN is not suitable and it should not be
used to make predictions as well as optimize the process even if
the graphical representation suggest the opposite.
4. Discussion
Fig. 2. Predictions with RBFNN with the objective, 3 and 5 inputs.
Table 10
ANOVA in RBF redesigned in the GMAW process.
Analysis of Variance
Source of
variation
Degrees of
freedom
Sum of
squares
Mean
square
Model
Residual error
Total
5
20
25
3.3457
2.1604
5.5062
0.6691
0.1080
6.19 0.001
Table 11
Standard error of weights and terms in GMAW
process.
Term
se (weights)
X1 Amperage
X2 Voltage
X3 SpeedFeed
X4 ElectrodeExtension
X5 HeatInput
Bias
1.2309
0.0002
0.0037
0.0026
0.0009
0.0000
1890
Table 12
Test hypotheses in the inputs in GMAW process.
Table 16
Results in statistical metrics with 3 inputs.
Term
To
To p-value
X1 Amperage
X2 Voltage
X3 SpeedFeed
X4 ElectrodeExtension
X5 HeatInput
Bias
0.3911
0.6861
1.8619
3.9932
3.8309
2.6278
0.6999
0.5005
0.0774
0.0007
0.0010
0.0161
Statistical
method
Results Statistical
method
R2
65.2
MSError
0.0872
PRESS
R2Adj
60.4
PREMS
0.0738
R2Prediction
Table 13
Condence intervals on the weights in GMAW process.
Weights
Lower limit
wj
Upper limit
w1
w2
w3
w4
w5
w0
2.0862
0.0002
0.0145
0.0050
0.0055
0.0000
0.4814
0.0001
0.0068
0.0106
0.0036
0.00007
3.0490
0.0004
0.0008
0.0161
0.0016
0.0001
Table 14
Condence intervals on the predictions in GMAW process.
Lower limit
y^ 0
Upper limit
1.1815
1.7893
1.1919
0.9812
1.349
2.1642
1.7742
1.7433
1.7597
0.9406
1.3011
2.3754
1.407
1.407
1.4979
1.7954
2.101
1.9733
1.5564
1.5223
1.736
1.9757
1.8558
2.003
2.0065
1.8762
1.916
2.5983
1.9413
1.7907
2.083
2.9044
2.474
2.4522
2.4957
1.7166
2.2368
3.1797
2.1518
2.1518
2.2525
2.6144
2.9051
2.7017
2.2642
2.2457
2.4735
2.748
2.577
2.7243
2.7613
2.6399
2.6505
3.4073
2.6907
2.6002
2.8171
3.6446
3.1738
3.1612
3.2317
2.4925
3.1726
3.984
2.8966
2.8966
3.007
3.4335
3.7093
3.4302
2.972
2.969
3.2109
3.5203
3.2982
3.4455
3.5161
3.4037
Table 15
Results in statistical metrics in the GMAW process.
Radial Basis Function redesigned
Statistical
method
Results Statistical
method
Results Statistical
method
R2
60.8
MSError
0.1080
PRESS
R2Adj
51.0
PREMS
0.0831
R2Prediction
Results
3.3563
39.04
Results Statistical
method
Results
2.7381
50.27
Then with this adjusted the model is better, the R2 and R2Adj
should not be greater than 5 perceptual units and this case
whether complies the assumption. Graphically the three models
seem equals with the objective, the RBFNN with 3 inputs and
5 inputs. However, the statistic R2Prediction 50:27 represents about
11 units better than the model with 5 inputs (R2Prediction 39:04).
Then the new RBFNN prediction with 3 inputs is better than
5 inputs. Furthermore If the PREMS is less than the MSError the
model is going to be a good predictor; in this case the application
complies with this premise, in which the model can be used to
predict and optimize.
In addition to this, these statistics are better than the values
that the model provides us with 5 inputs; it means that there is a
better model with more variables statistically signicant, which in
this case were: X3 SpeedFeed, X4 ElectrodeExtension and
X5 HeatInput. The relationship between inputs and outputs
from RBFNN is given by yx m
i 1 wGjjxt i jj b and is written
in matrix form as Gw yx, then the new RBFNN model with
more variables statistically signicant to predict and optimize is
yx G3 0:000072 G4 0:0001 G5 0:000031 2:920717:
33
5. Conclusions
The method of Hybrid Learning Process presented in this work
successfully applies a Genetic Algorithm to calculate the matrix of
centers and the coefcient of determination R2 becomes the
statistical evaluation function of GA improving the accuracy of
the prediction and optimization of the model. With these modications, the original contribution is that our redesigned RBF
model fullls the assumptions to realize statistical inference.
A very important contribution, it is given to the application of
ANOVA to determine the statistical signicance and the relationship between variables. For our case studies, machining and
welding process, some terms are deemed signicant; have a
contribution in the process, and are important to t the model.
A statistical inference is therefore made to apply ANOVA, test
hypothesis, condence intervals in weights wj and y^ i , to apply
statistical metrics as R2 , R2Adj , PREMS, PRESS, R2Prediction , among
others. This method might identify which variable is most important and has the largest effects on the process, e.g. in the GMAW
application it could be concluded from Table 12 that the variables
which correspond to w1 and w2 are not statistically signicant and
must not be included in the RBNN model, in this case we conclude
that variables contributes to the total variation of the response and
there are the more important are: Speed Feed, Electrode Extension
and Heat Input.
Moreover, it shows that statistical methods are a good alternative for validating the efciency of Neural Networks models.
Hence, it can be concluded that the RBF model is a good tool for
prediction and optimization when complies all the statistical
assumptions.
1891