Modeling Pile Capacity Using Support Vector Machines

and Generalized Regression Neural Network

Mahesh Pal1 and Surinder Deswal2

Abstract: This note investigates the potential of support vector machines based regression approach to model the static pile capacity
Downloaded from [Link] by Tokyo Univ Seisan Gijutsu on 05/13/15. Copyright ASCE. For personal use only; all rights reserved.

from dynamic stress-wave data. A data set of 105 prestressed precast high strength concrete spun pipe piles is used. Radial basis function
and polynomial kernel based support vector machines were used to model the total pile capacity and results were compared with a
generalized regression neural network approach. A total of 81 data set were used to train, whereas the remaining 24 data sets were used
to test the created model. A correlation coefficient value of 0.977 was achieved by generalized regression neural network in comparison
to values of 0.967 and 0.964 achieved by radial basis function and polynomial kernel based support vector machines, respectively. Results
suggest an improved performance by generalized regression neural network based approach in comparison to support vector machines but
polynomial kernel based support vector machines provide a linear relationship to predict total pile capacity using stress-wave data.
DOI: 10.1061/共ASCE兲1090-0241共2008兲134:7共1021兲
CE Database subject headings: Piles; Neural networks; Vector analysis; Stress wave.

Introduction namic formula. Rausche et al. 共1985兲 proposed to use a stress-

wave matching technique to determine the static pile capacity
Piles are being used for many years as a type of foundation. from the dynamic test data. The main principle of stress-wave
Prediction of axial pile capacity is very important for the quality matching approach is that important information regarding the
control of pile foundations. There are a number of static and soil model is contained in the measured stress-wave data.
dynamic load test based approaches to determine the pile capac- The one-dimensional stress-wave analysis is also executed by
ity. Due to high cost and time involved in static pile load tests dedicated software on a personal computer, such as CAPWAP
dynamic load tests are employed to determine the pile capacity. 共Rausche et al. 1985兲 or the TNOWAVE program developed
Practical problems and the cost considerations forced the re- by TNO Building and Construction Research Organization,
Waddinxveen, The Netherlands. Both are based on the method
searchers to look into the alternate approaches to determine the
of characteristics to have an accurate solution of the wave
pile capacities.
equation. Each of these programs has its own pile–soil models
A number of empirical pile driving formulas were also pro-
and input parameters, but the application of signal matching
posed so as to reduce the time and cost involved in static ap-
techniques is similar. A recent work by Middendorp 共2004兲 sug-
proach 共Meyerhof 1976; Coyle and Castello 1981; Randolph
gests that the signal matching programs such as CAPWAP and
1985兲. Although these formulas have been widely used to predict
TNOWAVE produce similar results when using the same input
pile capacity, it is agreed that these formulas are inaccurate due to data. Further, Rausche et al. 共1985兲 suggested a comparable per-
their oversimplification of the modeling of the hammer, driving formance by stress-wave data and static methods in determining
system, pile, and soil 共Fragaszy et al. 1985兲. the pile capacity.
Within last few decades, because of advances in data acquisi- Soft computing techniques, like artificial neural network, is
tion during pile driving dynamic testing are being widely used for being used to predict the load bearing capacity of piles using both
pile capacity prediction. Dynamic measurements of force and ve- static and dynamic data sets 共Chan et al. 1995; Chow et al. 1995;
locity at the upper end of the pile during pile driving, followed by Goh 1996; Teh et al. 1997; Lok and Che 2004兲. Results obtained
a signal matching procedure, is the most common method for by neural network were compared with different equations pro-
dynamic determination of pile capacity as an alternative to dy- posed in literature and found to be working well in comparison to
the empirical relations. A neural network based modeling algo-
1
Assistant Professor, Dept. of Civil Engineering, National Institute of rithm requires setting up of different learning parameters 共like
Technology, Kurukshetra, 136119, Haryana, India 共corresponding learning rate, momentum兲, the optimal number of nodes in the
author兲. E-mail: mpce_pal@[Link] hidden layer and the number of hidden layers. A large number
2
Assistant Professor, Dept. of Civil Engineering, National Institute of of training iterations may force a neural network to over train,
Technology, Kurukshetra, 136119, Haryana, India. which may affect the predicting capabilities of the model. Further,
Note. Discussion open until December 1, 2008. Separate discussions presence of local minima is another problem with the use of a
must be submitted for individual papers. To extend the closing date by
backpropagation neural network.
one month, a written request must be filed with the ASCE Managing
Editor. The manuscript for this technical note was submitted for review Within last few years, several studies reported the use of
and possible publication on May 24, 2007; approved on October 8, 2007. support vector machines in civil engineering 共Dibike et al. 2001;
This technical note is part of the Journal of Geotechnical and Geoenvi- Pal and Mather 2003; Gill et al. 2006; Pal and Goel 2006; Pal
ronmental Engineering, Vol. 134, No. 7, July 1, 2008. ©ASCE, ISSN 2006兲 and found it to be working well in comparison to a neural
1090-0241/2008/7-1021–1024/$25.00. network approach. Similarly generalized regression neural net-

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JULY 2008 / 1021

J. Geotech. Geoenviron. Eng. 2008.134:1021-1024.

 Table 1. Value of User-Defined Parameters with Support Vector Data Set
Machines and Generalized Regression Neural Network
Data used in the present study are taken from an earlier study
RBF kernel Polynomial kernel
by Lok and Che 共2004兲. They used a backpropagation neural
Classifier used C ␥ C d network to associate the input stress-wave data with total pile
Support vector machines 1 0.1 0.15 1 capacity derived from the commercial computer software
Generalized regression TNOWAVE 共represents a group of wave equation application
neural network Spread= 0.25 programs and is a powerful simulation tool to predict pile behav-
ior兲. The data set consists of 105 prestressed precast high strength
concrete spun pipe piles, collected from more than 50 construc-
tion projects at Macau 共China兲. All piles used in this study are
work has successfully been used for several civil engineering circular shaped with hollow section. To determine the total pile
problems 共Abu Keifa 1998; Nawari et al. 1999; Juang et al. 2001; capacity 关i.e., ␴total 共MN兲兴, seven input variables, namely F1, F2,
Downloaded from [Link] by Tokyo Univ Seisan Gijutsu on 05/13/15. Copyright ASCE. For personal use only; all rights reserved.

Juang and Jiang 2003; Cigizoglu 2005; Kurup and Griffin 2006兲. Z␯1, Z␯2, ⌽, Th, and L was considered, where F and Z␯ = force
Both of these modeling approaches require setting of few users 共MN兲 and velocity 共MN兲 multiplied by impedance, respectively at
defined parameters and provides global minima. time t p and 共t p + 2L / c兲; ⌽ = diameter of hollow pipes piles 共m兲;
Keeping in view the better performance by support vector ma- Th = wall thickness 共m兲; and L = length of the pile 共m兲. For
chines, this study compares its performance with a generalized further details about the data set, as well sampling process used
regression neural network in predicting the total pile capacity to select training and test data, readers are referred to Lok and
using stress-wave data. This work uses “TNOWAVE predicted Che 共2004兲.
pile capacity” rather than the true bearing capacity of the pile.

Details of Support Vector Machines and GRNN

Support Vector Machines
In situations with nonlinear decision surfaces, support vector ma-
chines use the concept of the kernel 共Vapnik 1995兲. A number of
Vapnik 共1995兲 proposed ␧-support vector regression 共SVR兲 by
kernels are discussed in the literature, but studies suggest a better
introducing an alternative ␧-insensitive loss function. This loss
performance by polynomial and radial basis kernels for different
function allows the concept of margin to be used for regression
civil engineering problems 共Pal and Mather 2003; Pal and Goel
problems where margin is defined as the sum of the distances of
2006; Gill et al. 2006兲. Thus, the present study uses a polynomial
the hyper plane from the closest point of the two classes. The 2
purpose of the SVR is to find a function having at most ␧ devia- K共x , y兲 = 关共x · y兲兴d and a radial basis kernel 共e−␥兩x − y兩 兲, where d and
tion from the actual target vectors for all given training data and ␥ = kernel specific parameters.
have to be as flat as possible 共Smola and Schölkopf 1998兲. Vapnik The use of support vector machines requires setting of the
共1995兲 introduced the concept of kernel function for nonlinear appropriate user-defined parameters, as the accuracy of a SVMs
support vector regression. For further details of support vector model is largely dependent on the selection of these parameters.
regression, readers are referred to Smola and Schölkopf 共1998兲 The SVMs, in addition to the choice of a kernel requires the
and Vapnik 共1995兲. setting up of kernel specific parameters. The optimum values of
the regularization parameter C and the size of the error-insensitive
zone ␧ also need to be determined.
Generalized Regression Neural Network To select user-defined parameters 共i.e., C, ␥, and d兲, a manual
The generalized regression neural network 共GRNN兲 proposed by method, which involves carrying out a large number of trials by
Specht 共1991兲 does not require an iterative training procedure as using different combinations of C and d for polynomial kernel
in the backpropagation method. A GRNN configuration consists and C and ␥ for radial basis function kernel based support vector
of four layers. The input units are in the first layer, the second machines 共SVMs兲, was used. Similarly, a number of trials were
layer has the pattern units, the outputs of this layer are passed also carried out to find a suitable value of ␧ 共error-insensitive
on to the summation units in the third layer, and the final layer zone兲 with a fixed value of C and kernel specific parameters.
covers the output units. The first layer is fully connected to the For this study, a value of 0.0010 was found to be working well.
second, pattern layer, where each unit represents a training pattern Table 1 provides the optimal values of user-defined parameters
and its output is a measure of the distance of the input from the for this data set with polynomial and radial basis function kernel
stored patterns. The optimal value of spread parameter共s兲 is often based SVMs. In the present study, SVM_RBF and SVM_POLY
determined experimentally. For further details of GRNN readers represent radial basis function and polynomial kernel based
are referred to Specht 共1991兲. SVMs, respectively.

Table 2. Correlation Coefficient, Root-Mean-Square Error and Root-Mean-Square Percentage Error Using Support Vector Machines and Generalized
Regression Neural Network
RBF kernel Polynomial kernel
Correlation RMSE Correlation RMSE
Classifier used coefficient 共MN兲 RMSPE coefficient 共MN兲 RMSPE
Support vector machines 0.964 0.352 0.060 0.962 0.359 0.061
Generalized regression neural network 0.977 0.303 0.050

1022 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JULY 2008

J. Geotech. Geoenviron. Eng. 2008.134:1021-1024.

Downloaded from [Link] by Tokyo Univ Seisan Gijutsu on 05/13/15. Copyright ASCE. For personal use only; all rights reserved.

Fig. 1. Actual versus predicted total pile capacity with RBF kernel Fig. 3. Actual versus predicted total pile capacity with GRNN
based SVMs
Figs. 1–3 provide the graphs plotted between actual and pre-
For GRNN, only one parameter, i.e., spread of radial basis dicted value of total pile capacity obtained by using support vec-
functions, needs to be determined for a given data set. After a tor machines and generalized regression neural network with the
number of trials a value of 0.25 was found to be performing well test data set. Figs. 1–3 suggest that most of the predicted values
for this data set. are lying within ⫾10% error from the line of perfect agreement
with both SVMs and GRNN. A higher value of R2 共i.e., 0.954兲
with GRNN confirms that this approach works well in predicting
Results the total pile capacity in comparison to both the radial basis func-
tion and the polynomial kernel based SVMs. Further, a com-
To assess the usefulness of SVMs and GRNN in predicting the parison of results suggests that both polynomial and radial basis
total pile capacity, 81 data sets for training and 24 for testing were function kernel based SVMs work equally well for this data set.
used. Both training and test data sets are normalized so as to bring Fig. 4 represent the variation of actual and predicted total pile
all input values between zero and one 共Chow et al. 1995兲. Corre- capacity with the number of test data. It is evident from Fig. 4 that
lation coefficient, root-mean-square error 共RMSE兲 and root-mean- both GRNN and SVMs show almost the same trend in predicting
square percentage error 共RMSPE兲 共Chan et al. 1995兲 were used the total pile capacity.
to compare the performance of SVMs with the GRNN. Table 2 In spite of better performance by GRNN, the training data set
provides the value of correlation coefficient, RMSE and RMSPE used with polynomial kernel based SVMs 共with d = 1, i.e., poly-
obtained by radial basis function and polynomial kernel based nomial of degree one兲, provide a model in the form of a simple
SVMs as well as by GRNN. In comparison to the correlation linear relation to predict the total pile capacity. The following
coefficient value of 0.977 共RMPSE= 0.050兲 achieved by the gen- equation represents the linear regression function 共i.e., model兲
eralized regression neural network, the radial basis function and provided by polynomial kernel based SVMs:
polynomial kernel based SVMs provide correlation coefficient
values of 0.964 and 0.962 共RMPSE= 0.060 and 0.061兲, respec- 0.6225F1 − 0.0383F2 + 0.6864Z␯1 − 0.6782Z␯2 + 1.0959⌽
tively. Thus, suggesting a better performance by generalized + 0.6793Th + 0.6941L + 4.2005 共1兲
regression neural network in comparison to support vector
machines for this data set. The constant term in Eq. 共1兲 represents the bias term in the
support vector regression function. Eq. 共1兲 suggests that values

Fig. 2. Actual versus predicted total pile capacity with polynomial Fig. 4. Variation in predicted values by both SVMs and GRNN in
kernel based SVMs comparison to the actual values

JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JULY 2008 / 1023

J. Geotech. Geoenviron. Eng. 2008.134:1021-1024.

 Table 3. Root-Mean-Square Error and Correlation Coefficient Using of pile capacity from stress-wave measurements: A neural network
Generalized Regression Neural Network with the Data Set Using F1, F2, approach.” Int. J. Numer. Analyt. Meth. Geomech., 19共2兲, 107–126.
Z␯1, Z␯2, and L as Input Parameters Cigizoglu, H. K. 共2005兲. “Application of generalized regression neural
networks to intermittent flow forecasting and estimation.” J. Hydrol.
Correlation RMSE
Eng., 10共4兲, 336–341.
Classifier used coefficient 共MN兲 RMSPE
Coyle, H. M., and Castello, R. R. 共1981兲. “New design correlations for
Generalized regression 0.957 0.378 0.069 piles in sand.” J. Geotech. Engrg. Div., 107共7兲, 965–986.
neural network with spread= 0.2 Dibike, Y. B., Velickov, S., Solomatine, D. P., and Abbott, M. B. 共2001兲.
“Model induction with support vector machines: Introduction and ap-
plications.” J. Comput. Civ. Eng., 15共3兲, 208–216.
of F and Z␯ at time t p have more influence on total pile capacity Fragaszy, R. J., Higgins, J. D., and Lawton, E. C. 共1985兲. “Development
in comparison to the values at time 共t p + 2L / c兲. Further, Eq. 共1兲 of guidelines for construction control of pile driving and estimation of
also suggests that in comparison to the thickness and length of pile capacity 共Phase I兲.” Rep. No. WA-RD-68.1, Washington State
the pile, diameter of the hollow pile have more weight in predict- Dept. of Transportation, Olympia, Wash.
Downloaded from [Link] by Tokyo Univ Seisan Gijutsu on 05/13/15. Copyright ASCE. For personal use only; all rights reserved.

ing the pile capacity. This suggests that the more pile diameter Gill, M. K., Asefa, T., Kemblowski, M. W., and Makee, M. 共2006兲. “Soil
for a fixed value of other parameters, the higher the pile capacity moisture prediction using support vector machines.” Environ. Urban-
will be. ization, 42共4兲, 1033–1046.
Keeping in view the better performance of GRNN for this Goh, A. T. C. 共1996兲. “Pile driving records reanalyzed using neural net-
works.” J. Geotech. Engrg., 122共6兲, 492–495.
data set, another data set having F1, F2, Z␯1, Z␯2, and L as input
Juang, C. H., and Jiang, T. 共2003兲. “A new approach to site characteriza-
parameters was used so as to compare its performance with
tion using generalized regression neural network.” Geotech. Spec.
the backpropagation neural network 共Lok and Che 2004兲 共see Publ., 121, 79–99.
Table 3兲. GRNN provides a root-mean-square percentage error of Juang, C. H., Jiang, T., and Christopher, R. 共2001兲. “Three-dimensional
0.069 in comparison to a value of 0.088 provided by a backpropa- site characterization—Neural network approach.” Geotechnique,
gation neural network as reported by Lok and Che 共2004兲, thus 51共9兲, 799–809.
suggesting a better performance by GRNN in comparison to Kurup, P. U., and Griffin, E. P. 共2006兲. “Prediction of soil composition
backpropagation neural network also. from cpt data using general regression neural network.” J. Comput.
Civ. Eng., 20共4兲, 281–289.
Lok, T. M. H., and Che, W. F. 共2004兲. “Axial capacity prediction for
Conclusions driven piles using ann: Model comparison.” GeoTrans 2004, Geotech-
nical Engineering for Transportation Projects 共GSP No. 126兲, M. K.
This paper investigates the potential of SVMs and GRNN in pre- Yegian and E. Kavazanjian, eds.
dicting the total pile capacity using stress-wave data. A major Meyrehof, G. G. 共1976兲. “Bearing capacity and settlement of pile foun-
dations.” J. Geotech. Engrg. Div., 102共3兲, 196–228.
conclusion from this study is that the GRNN approach works well
Middendorp, P. 共2004兲. “Thirty years of experience with the wave equa-
in comparison to SVMs for this data set but both of these ap-
tion solution based on the method of characteristics.” 7th Int. Conf. on
proaches yields an approximation to the capacity predicted by
the Application of Stress Wave Theory to Piles.
TNOWAVE signal matching methods only. Even the performance Nawari, N. O., Liang, R., and Nusairat, J. 共1999兲. “Artificial intelligence
of polynomial kernel based SVMs is slightly inferior to GRNN, techniques for the design and analysis of deep foundations.” Electron.
an important finding of this study is a simple linear relation pro- J. Geotech. Eng., 4, 具[Link]
vided by this approach, which can easily be used by field engi- [Link]典.
neers to determine the total pile capacity by using stress-wave Pal, M. 共2006兲. “Support vector machines-based modelling of seismic
data with F1, F2, Z␯1, Z␯2, ⌽, Th, and L as input parameters. liquefaction potential.” Int. J. Numer. Analyt. Meth. Geomech.,
30共10兲, 983–996.
Pal, M., and Goel, A. 共2006兲. “Prediction of the end depth ratio and
Acknowledgments discharge in semicircular and circular shaped channels using support
vector machines.” Flow Meas. Instrum., 17共1兲, 50–57.
Pal, M., and Mather, P. M. 共2003兲. “Support vector classifiers for land
The stress-wave data set used in this study was provided by Dr. T.
cover classification.” Map India 2003, 具[Link]/
M. H. Lok, Faculty of Science and Technology, University of
technology/rs/pdf/[Link]典 共March 25, 2007兲.
Macau, Av. Padre Tomás Pereira S.J., Taipa, Macao, China.
Randolph, M. F. 共1985兲. “Capacity of piles driven into dense sand.”
The writers would like to thank three anonymous reviewers for
Rep. Soils TR 171, Engineering Dept., Cambridge Univ., Cambridge,
their constructive comments that led to the improvement of this U.K.
manuscript. Rausche, F., Goble, G. G., and Likins, G. E. 共1985兲. “Dynamic determi-
nation of pile capacity.” J. Geotech. Engrg., 111共3兲, 367–383.
Smola, A. J., and Schölkopf, B. 共1998兲. “A tutorial on support vector
References regression.” NeuroCOLT Technical Rep. No. NC-TR-98-030, Royal
Holloway College, Univ. of London, London.
Abu Keifa, M. A. 共1998兲. “General regression neural networks for driven Specht, D. F. 共1991兲. “A general regression neural network.” IEEE Trans.
piles in cohesionless soil.” J. Geotech. Geoenviron. Eng., 124共12兲, Neural Netw., 2共6兲, 568–576.
1177–1185. Teh, C. I., Wong, K. S., Goh, A. T. C., and Jaritngam, S. 共1997兲.
Chan, W. T., Chow, Y. K., and Liu, I. F. 共1995兲. “Neural networks: “Prediction of pile capacity using neural networks.” J. Comput. Civ.
An alternative to the pile driving formulas.” Comput. Geotech., 17共2兲, Eng., 11共2兲, 129–138.
135–156. Vapnik, V. N. 共1995兲. The nature of statistical learning theory, Springer,
Chow, Y. K., Chan, W. T., Liu, I. F., and Lee, S. L. 共1995兲. “Predication New York.

1024 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING © ASCE / JULY 2008

J. Geotech. Geoenviron. Eng. 2008.134:1021-1024.