Professional Documents
Culture Documents
PUK in Yarn Tenacity - Unlocked PDF
PUK in Yarn Tenacity - Unlocked PDF
A new kernel function of SVM based on the Pearson VII function has been applied and compared with the commonly
applied kernel functions, i.e. the polynomial and radial basis function (RBF), to predict yarn tenacity. It is found that the
SVM model based on Pearson VII kernel function (PUK) shows the same applicability, suitability, performance in
prediction of yarn tenacity as against SVM based RBF kernel. The comparison with the ANN model shows that the two
SVM models give a similar predictive performance than ANN model.
Keywords: Artificial neural network, Pearson VII kernel function (PUK) kernel, Radial basis function kernel, Support
vector machines, Yarn properties
f ( x ) = w. ( x ) + b (1)
2 i =1 j =1
( i* i )( *j i ) K ( xi , x j )
ABAKAR & YU: SVM BASED ON PUK KERNEL IN COMPARISON TO SVM BASED ON RBF KERNEL 57
used. In 10-fold cross validation, the initial dataset is C=160.84, = 0.02, = 85.0, and = 9.1, which give
randomly partitioned into 10 mutually exclusive the smallest value of RMSE.
subsets or folds D1, D2, , D10, each of The initial architecture of the ANN selected has
approximately equal size. The training and testing are used all the eight variables in input layer with four
performed 10 times. In iteration i, partition Di is neurons in the hidden layer, selected by the auto
reserved as the test set and the remaining partition are builtin function and one output neuron. The
collectively used to train the model. In this method, optimization of the BP multilayer perceptron model is
each data point is used for the same number of times done with 10-fold cross validation. After several
for training and once for testing. Then the error trials, the optimal learning rate of 0.01 and a
estimates are calculated and then averaged. The errors momentum coefficient of 0.3 are determined, and the
that were used as an indicator of the predictive number of epochs is selected as 300.
performance of the models were root mean-squared After the completion of model development or
error (RMSE), relative error (RE %), and correlation training, all the models based on SVM and ANN are
coefficient (R). For implementation and to carry out subjected to the unseen testing data set. Statistical
our experiments, the SVM and ANN models were parameters such as root mean-squared error (RMSE),
executed by using rapid miner software program. relative error (RE %), and correlation coefficient (R)
are used to compare the predictive power of the SVM-
3 Results and Discussion based and ANN based models. Results are
3.1 Models Performance Comparison summarized in Table1.
The goal of this part of research is to compare the It can be seen from Table 1 that the largest values
prediction results provided by SVM based on of RMSE and RE are provided by SVM based on
polynomial, Gaussian radial basis (RBF) and PUK polynomial kernel. This indicates that the
kernels function as well as the ANN model. The generalization performance of SVM based on
optimization of SVM parameters is performed by polynomial kernel is bad. Therefore, this model does
using grid search approach in the training set data not fit the data well. The RMSE of SVM based on
using 10-fold cross validation, and depending on RBK kernel is bound to be lower than that of SVM
smallest RMSR error we selected the optimal based on PUK kernel. The RE and R values provided
parameters of the model. The regularization constant by both SVM models based on RBK and PUK kernels
( the complexity parameter C ), parameters of the are the same.
- insensitive loss function, kernel parameters of Therefore, the RMSE values of both SVM models
the degree d of polynomial kernel , width of RBF based on RBK and PUK kernels are lower than that of
kernel , and PUK kernel parameters and were ANN model. The RE and R values are found to be
optimized. The optimal parameters of SVM based on close with a little difference. This comparison shows
polynomial kernel are C=1.0, = 0.002, and d = 3.0. that both SVM models based on RBK and PUK
The optimal parameters of SVM based on RBF kernel kernels and ANN model are pretty similar in terms of
are C=160.84, = 0.001, and = 0.04, and the the average errors. The comparison of performance of
optimal parameters of SVM based on PUK kernel are the nonlinear models demonstrates that the SVM
Table 1Comparison of the predictive performance of SVM and ANN model
Experimental data Predicted value using
SVM based SVM based SVM based ANN
polynomial kernel RBF kernel PUK kernel
11.86 11.266 11.155 11.275 11.559
13.52 12.247 12.582 12.616 13.564
18.28 30.042 20.582 18.748 16.705
11.77 6.092 11.278 10.796 9.484
10.82 11.362 10.736 10.794 11.691
Error
RMSE 5.369 0.911 0.646 1.280
RE, % 22.42 5.02 4.31 7.91
R 0.934 0.987 0.989 0.901
ABAKAR & YU: SVM BASED ON PUK KERNEL IN COMPARISON TO SVM BASED ON RBF KERNEL 59
Table 2Anova test model shows that the two SVM models have similar
Source Square DF Mean F Prob prediction performance as ANN model. The result of
sums squares this study will be helpful for further work of feature
Between 0.003 1 0.003 0.060 0.812 selection to know the importance of fibre properties
Residuals 0.821 18 0.046 on yarn tenacity by using SVM based on PUK kernel.
Total 0.824 19
Acknowledgement
based on PUK kernel has the similar ability in
The authors gratefully acknowledge Shanghai
predicting yarn tenacity with SVM based on RBF
Municipal Education Commission and Shanghai
kernel and ANN model.
Education Development Foundation for providing
3.2 Compare SVM Models by Significance Test
Shu Guang project (No. 10GG17).
The two SVM models with RBF and PUK kernel
have been compared to know which of them is more References
1 Majumdar P K & Majumdar A, Text Res J, 74(7) (2004) 652.
accurate and suitable in prediction of yarn tenacity
2 Chattopadhyay R & Guha A, Text Prog, 35(1) (2004) 1.
property, Hence, we applied two cross validation on 3 reyen M E & Grkan P, Fibers Polym, 9(1) (2008) 87.
the training set data using SVM model based on RBF 4 Majumdar A & Ghosh A, J Eng Fibers Fabrics, 3(4) (2008) 61.
kernel and on PUK kernel. The significance test 5 Admuthe L S &. Apte S, Text Res J, 80(9) (2010) 841.
(T test and Anova) is performed on the resulting 6 M. Dayik, Text Res J, 79(11), 963 (2009).
7 Yang J G, Lv Z J & Xiang Q, Yarn properties prediction
performance value to decide whether the difference in using support vector machines: an intelligent reasoning
performance is significant. Table 2 shows that the method. Proceedings, Fourth International Conference on
probability for random values with the same result is Fuzzy Systems and Knowledge Discovery, Haikou, 2007, 1.
0.812, these values are bigger than alpha= 0.050, 8 Ghosh A & Chatterjee P, Fibers Polym, 11(1) (2010) 84.
9 Nurwaha D & Wang X, Fibers Polym, 12(4) (2011) 546.
which indicates a probably difference between the
10 stn B, Melssen W J & Buydens L M C, Chemometrics
actual mean values was not significant, since Intelligent Laboratory Systems, 81(1) (2006) 29.
0.812 > alpha = 0.050. This means that the SVM 11 Cheng Y S J & Cheng K P S, Text Res J, 74(8) (2004) 718.
model based on PUK kernel has the same ability to 12 Vapnik V N, The Nature of Statistical Learning Theory
that of SVM model based on the RBF kernel to map (Springer-Verlag New York), 1995.
13 Vapnik V N, IEEE Transactions on Neural Networks,10(5)
the nonlinear relation between input and output data (1999) 988.
for predicting the yarn tenacity property. Hence, it is 14 Smola A J & Schlkopf B, Statistics Computing, 14(3)
possible to predict the yarn properties by using SVM (2004) 199.
model based on PUK kernel. 15 Suykens J A K, Vandewalle J & Moor B D, Neural
Networks, 14(1) (2001) 23.
16 Mukherjee S, Osuna E & Girosi F, Nonlinear prediction of
4 Conclusion chaotic time series using support vector machines,
It is observed that like SVM based on RBF, the Proceedings, IEEE Workshop on Neural Networks for Signal
SVM model based on PUK kernel shows the same Processing VII (Amelia Island Florida) 1997. 511.
applicability, suitability and performance to map the 17 Jeng J T, Chuang C C & Su S F, Fuzzy Sets Systems, 138(2)
nonlinear relation between input and output data for (2003) 283.
18 Tay F E H & Cao L, Omega, 29(4) (2001) 309.
predicting the yarn tenacity. A comparison of SVM 19 Chuang C C & Su S F, IEEE Transactions on Neural
models based on RBF and PUK kernels with ANN Networks, 13(6) (2002) 1322.