Professional Documents
Culture Documents
a r t i c l e i n f o abstract
Article history: An accurate estimation of pile response to loading is a challenging task due to the complexity of the
Received 30 August 2010 soil–pile interactions and uncertainties in the soil properties. Conventional methods of predicting pile
Received in revised form load–settlement relationship either oversimplify the problem or require the parameters that are
5 February 2011
difficult to determine in the laboratory. In this study, a high-order neural network (HON) is developed
Accepted 8 February 2011
to simulate the pile load–settlement curve using properties of the pile and SPT data along the depth of
Available online 7 April 2011
pile embedment as inputs. The results indicated a significant improvement in the quality of HON
Keywords: predictions over that of BPN, RBF and GRNN models. Based on the comparisons with the predictions of
Artificial neural network elastic and hyperbolic models, the proposed HON model provides better predictions than existing
High-order neural network
theoretical models.
Soil–pile interaction
& 2011 Elsevier Ltd. All rights reserved.
SPT
Load–settlement behaviour
0952-1976/$ - see front matter & 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.engappai.2011.02.008
814 A. Ismail, D.-S. Jeng / Engineering Applications of Artificial Intelligence 24 (2011) 813–821
successfully applied to a wide variety of geotechnical problems correction factor for pile compressibility; Rh the depth correction
such as geo-material modelling (Ellis et al., 1995; Penumadu and factor; Rv the correction for Poisson’s ratio; Rb the correction for
Jean-Lou, 1997; Basheer and Najjar, 1998), soil liquefaction end bearing; s is the immediate settlement; Es the elastic modulus
problems (Agrawal et al., 1997; Ural and Saka, 1998), prediction of soil along the pile shaft and Db stands for the diameter of the
of pile capacity (Goh, 1995, 1996; Chan et al., 1995; Teh et al., pile base.
1997; Lee and Lee, 1996; Nawari et al., 1999; Das and Basudhar, Other methods based on elastic solution were developed
2006). Shahin et al. (2002) developed a BPN model for predicting by Vesic (1977), Randolph and Wroth (1978) and Fleming et al.
shallow foundation settlement using SPT blow counts as repre- (1992). Elastic solutions are severely limited to circumstances
sentative of soil property in the input space. Recently, Nejad et al. where the simple linear soil behaviour can be considered
(2009) used a BPN network to correlate between pile head adequate.
settlement and SPT blow counts of soil around the pile.
The successes of the aforementioned investigations strengthen 2.2. Hyperbolic models
the position of neural networks as alternative numerical methods of
solving intrinsically non-linear geotechnical problems such as soil– In this approach, the pile head settlement is related to the pile
pile interaction. Among the various types of neural networks, back- head load using a hyperbolic model. It is an attempt to represent
propagation network (BPN) is the most widely used model for the non-linear behaviour of soil-shaft and soil-base interaction
geotechnical engineering problems. This is due to the simplicity of using relatively simple model. The method was first described
learning rule used in training the network and the stability of by Chin (1970) and later improved by Fleming (1992) to treat the
sigmoid transfer function, which is used as activation function. components of shaft and base resistances separately. According
Despite their popularity, BPN networks are not devoid of problems; to Fleming (1992), the pile load and the pile head settlement are
the sigmoid transfer function is a bi-asymptotic one and becomes related as follows:
insensitive to the variation of inputs as it approaches either 1 or 0.
This severely limits the ability of the network to make reasonable aDt bDt
Pt ¼ þ ð2Þ
extrapolations. One of the solutions to the said problem is to use c þ Dt d þ eDt
non-asymptotic processing elements such as high-order neurons. in which Pt and Dt represent the pile load and pile head settle-
High-order neural networks use polynomial functions to map inputs ments, respectively, while a ¼Us, b¼DBEBUB, c¼MsDB, d ¼0.6UB
into output and can be trained through error back-propagation and e¼ DBEB. Us, UB, DB, EB and Ms, are the ultimate shaft
algorithm. They have successfully been applied in pattern recogni- resistance, ultimate base resistance, pile base diameter, soil
tion (Foresti and Dolso, 2004; Artyomov and Yadid-Pecht, 2005) and deformation soil modulus at the pile base and flexibility factor,
function approximation (Rovithakis et al., 2004), but have not yet respectively. The method tends to yield better results than elastic
been applied in geotechnical modelling. methods. However, the assumption of hyperbolic stress–strain
In this study, HON-PILE, a high-order neural network model relationship is not applicable to all circumstances. Furthermore,
for the simulation of load–settlement behaviour is developed. The there is no established correlation between the hyperbolic para-
model predicts the relationship between the pile head settlement meters and in-situ (SPT) tests, and one has to rely on correlations
and the magnitude of axial loading exerted on the pile given the between SPT and soil parameters, which are unreliable.
geometric parameters and elastic modulus of pile and the proper-
ties of foundation soil, represented by the standard penetration
2.3. Step integration methods
test blow counts (SPT-N). A comprehensive comparison between
the proposed HON model, BPN and the elastic approximation and
hyperbolic model will be performed. The relative performance of Step integration approach was developed by Coyle and Reese
HON will also be determined in comparison with other poly- (1966). In this approach, shear-deformation relation for the soil–
nomial regression, RBF and GRNN networks. pile interface is modelled using series of independent non-linear
curves known as t–z curves. The base pressure versus deformation
is represented by another non-linear model known as q–y curve.
2. Existing pile load–settlement models The soil–pile system is simulated in this method as series of
incompressible weights connected by weightless springs to
In this section, the currently used methods of predicting pile represent the pile and side springs attached to the weights to
settlement, including elastic hyperbolic, step integration and FEM represent soil resistance described by t–z curves. At the bottom of
methods will be outlined. the bottom most weight, another spring is attached to represent
base resistance described by q–y curve. The problem is then
solved either using finite difference or one-dimensional finite
2.1. Elastic models
element method. The major challenge in step integration method
is the determination of parameters defining t–z and q–y curves.
Elastic models are based on Mindlin equations for the dis-
Various attempts have been made to develop an empirical
placement due a point load on a linearly elastic, semi-infinite soil
correlation between the curve parameters and in-situ tests
layer. They are also based on the assumption that the pile is acted
(Chang and Broms, 1991, Tan et al., 1998; Balakrishnan et al,
upon by a system of uniform shear stress around its shaft, while
1999). However, high scatter undermines the reliability of the
the base is acted upon by a uniform pressure. Elastic solutions are
correlations. Their inability to develop a good correlation is not
severely limited to circumstances where the simple linear soil
unconnected to the data size, the noise associated with the data
behaviour can be considered adequate. The most popular elastic
and, possibly due to limitations of the regression models used.
solution was developed by Poulos and Davis (1980) and can be
expressed as follows:
2.4. Finite element method
P
s¼ I ð1Þ
Es Db
Finite element method is the most rigorous technique used in
where I¼IoRkRhRvRb: Io is the displacement influence factor for a the prediction of pile head settlement. Although more computa-
rigid pile in a deep layer, for a Poisson’s ratio of 0.5; Rk the tionally intensive, than other methods discussed earlier, it is more
A. Ismail, D.-S. Jeng / Engineering Applications of Artificial Intelligence 24 (2011) 813–821 815
Table 2 and that of the pile. The second and third hidden layers consist of
Input and output parameters for various models. high-order networks with three neurons in each layer, while the
output layer consists of single linear neuron. Alongside the type I
Model Input parameters Output
HON model, a BPN with 4-8-3-1 configuration was developed for
BPN Ep Ab Pile head settlement the sake of comparison. Type II HON model consists of 14 input
P, ,N s Ai ,Nb Ab
L nodes, six hidden nodes in the first hidden layer and one output
HON (Type I) Ep Ab Pile head settlement
P, ,N s Ai ,Nb Ab layer (14-6-1) at the initial stage. Second hidden layer is then
L
HON (Type II) Ep Ab h i Pile load introduced in the course of training and the number of nodes
s, , N s ,Nb Ab , % sand,
L thereof is gradually increased from 0 to 5 the final configuration
% silt, % clay, pile Type
at the end of training session is 14-6-5-1. RBF (with 15 Gaussian
nodes) and GRNN networks are trained together with HON type II
model for assessment of relative performance.
Table 3 The database used for training and testing the models consists
Summary of training and testing results (BPN and HON (Type I) models). of results of static load tests carried out on 60 driven piles and 38
bored piles with over 900 data points. The pile data is drawn from
Network type R2-value RMSE (mm)
the sources given in Table 1. The detailed information of database
HON Type I (training) 0.91 6.7 are provided in a spread sheet that is available from the authors
HON Type I (testing) 0.88 7.1 and journal, upon request. All piles in the database are concrete
BPN (training) 0.87 6.5 cylindrical piles. The methods of installation include impact
BPN (testing) 0.78 8.9
driving in the case of driven piles and wet and dry excavations
in the case of bored piles. The range of pile diameters and lengths
are from 0.3 to 0.8 and 6 to 30 m, respectively. The elastic moduli
Table 4 of the piles range from 2.19 107 to 3.44 107 kN/m2. The soil
Summary of training and testing results (polynomial regression, RBF, GRNN and test results in the database consist of soil classification and
HON (Type II) models).
uncorrected SPT-N values determined at 1.5–3 m intervals and
Model R2-value RMSE (kN) soil classification. The rationale behind the use of uncorrected
values in this study is that they reflect the variation of over-
HON Type II (training) 0.91 554.4 burden stress, which in turn affects shaft and base resistance in
HON type II (testing) 0.89 607.6 coarse soils. The soil classification serves as additional informa-
Non-linear regression (training) 0.756 896.05
Non-linear regression (testing) 0.659 1090.04
tion for a better characterisation of the foundation soil. The soil
RBF network (training) 0.769 753.41 types in the database range from coarse soils (sand) to fine soils
RBF network (testing) 0.828 738.2 (silt and clay), with coarse soil being predominant. BPN and HON
GRNN model (training) 0.851 712.15 type I model were developed based on driven piles only, while
GRNN model (testing) 0.676 1124.28
type II HON was trained to predict the behaviour of both driven
Hyperbolic 0.562 1614.14
and bored piles.
180 120
160 R2 = 0.780
R2 = 0.870 100
140
120 80
100
60
80
60 40
40
20
20
0 0
0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120
Measured settlement (mm) Measured settlement (mm)
180 120
160
R2 = 0.87
100
R2 = 0.910
140
Predicted settlement (mm)
Predicted settlement (mm)
120 80
100
60
80
60 40
40
20
20
0 0
0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120
Measured settlement (mm) Measured settlement (mm)
Fig. 2. Comparison of measured versus predicted settlements: (a) BPN versus training data; (b) BPN versus testing data; (c) HON type I versus training data; and (d) HON
type I versus testing data.
a function of settlement, and soil and pile properties: Other factors considered are the soil type and method of
installation. The settlement function in Eq. (11) is approximated
P ¼ f 0 ðs,ks ,Ep ,D,LÞ ð6Þ
using BPN and HON type I as
0
The function f ( ) is the inverse of the settlement function. The
Ep Ab
soil stiffness (ks) in Eqs. (11) and (12) is a function of the soil sNN ¼ NN P,Ns As ,Nb Ab , ð9Þ
L
resistance around the pile, which can be expressed in terms of SPT
as follows: The NN approximated load function by HON type II is
expressed by
k ¼ f ðNs ,Nb Þ ð7Þ
h i
Ep Ab
in which Ns and Nb are the SPT-N along the shaft and at the base, PNN ¼ NN s, Ns ,Nb Ab , ,styp,ptyp ð10Þ
L
respectively.
To account for variation of N-value along the shaft, the pile is where Ns is the average value of N over the shaft length, denoted
divided into five segments. The weighted average value of N over by
the length of each segment is determined by
Pn Pi ¼ 5
Nj l j Ns,i i¼1
Ns,i ¼ 1 ð8Þ Ns ¼ ð11Þ
li 5
The term EpAb/L stands for the pile stiffness. ½Ns is the vector
where n is the number of sub-layers along the shaft of pile segment. of the product of average N-values along the shaft segments and
lj is the thickness of sub-layer j. li is the thickness of pile segment. the corresponding shaft area, respectively. It is expressed as
818 A. Ismail, D.-S. Jeng / Engineering Applications of Artificial Intelligence 24 (2011) 813–821
12000 10000
10000 R2 = 0.910
8000
8000
6000
6000
4000
4000
2000
2000
0 0
0 2000 4000 6000 8000 10000 12000 0 2000 4000 6000 8000 10000
Actual pile load (kN) Actual pile load (kN)
Fig. 3. Comparison of actual versus predicted loads: (a) HON type II versus training data and (b) HON type II versus testing data.
15000 follows:
T h i
R2 = 0.562 Ns ¼ N s,1 As,1 N s,2 As,2 N s,3 As,3 Ns,4 As,4 Ns,5 As,5 ð12Þ
where As,1 to As,1 are the areas of shaft segments. The parameters
Predicted pile load (kN)
styp and ptyp stand for soil type and method of pile installation.
10000
Soil type is represented by the percentage particle size, which are
approximately deduced from soil classification. The approximate
values so generated will then serve as inputs to the network four
nodes are dedicated for the approximate values of sand, silt and
clay percentages. The fourth node is assigned to plasticity index.
5000
Method of installation is represented by two nodes. For a
particular type of pile the appropriate node (or nodes) will fire a
value of unity. For example if a pile type is driven then both nodes
are disabled, when it is bored then one node is switched on. The
second node is activated to fire a value of 1.0 when the bored
0
pile is installed by dry excavation. The input parameters to the
0 5000 10000 15000
network include the pile head settlement; s, pile stiffness; EpAb/L.
Actual pile load (kN)
Others include six parameters to represent soil resistance; ½Ns
Fig. 4. Comparison of measured versus settlements predicted by hyperbolic model. and NbAb, percentage of sand, silt and clay and three input
parameters to represent method of pile installation (driving,
wet excavation and dry excavation). The input parameters used
in different models (BPN and HON) is summarized in Table 2.
10000
while the remaining 21% of the data was used for testing the
6000
network prediction. The training data is, as much as possible,
5000 made to capture the widest variations in input and output
patterns in the database. This is to avoid having extreme data in
4000 the testing set, which could make it impossible to assess the true
Measured data
HON type II simulation
generalization capability of the model within the domain of the
3000 training data as in the case of completely randomized selection
(Shahin et al., 2004).
2000
The synaptic weights were initialized uniformly distributed
random numbers in both networks, while learning parameters of
1000
0 5 10 15 20 25 30 0.01 and 0.1 were used in the training of HON and BPN,
respectively. Offline training was used in the case of HON while
Settlement (mm)
the BPN was trained using online method. The exponents of HON
Fig. 5. Comparison of HON type II simulations and training data. model are restricted to a range of 3.0 to 3.0. This is to prevent
A. Ismail, D.-S. Jeng / Engineering Applications of Artificial Intelligence 24 (2011) 813–821 819
compared with the training and testing data in Fig. 2(c) and (d),
500 respectively. The high-order network, compared to the BPN,
correlate better with the training data (R2 ¼0.91 for HON and
400 0.88 for BPN). This implies the level of learning in the case of HON
type I model exceeds that of BPN. The R2 values of 0.87 and 0.78
300
Measured data for type I HON and BPN, respectively, point to the better predic-
HON type II simulation tion of HON compared to BPN. It is however noted that at
200
settlement levels of 40 mm or less, the correlation is rather poor
100 in both HON type I and BPN.
The predictions of type II HON model are compared with
0 training and testing data in Fig. 3(a) and (b), respectively. The
0 1 2 3 4 5 6 7 strong correlation between HON predictions and duo of training
Settlement (mm) and testing data is indicative of the high level of learning achieved
by the network and the high quality of prediction the network is
Fig. 6. Comparison of HON type II simulations and testing data.
capable of doing. Also, by comparing the scattergrams of HON
type II (Fig. 3(b)) and Fleming’s hyperbolic model (Fig. 4), it could
be rightly said that the former remarkably out performs the
latter. Results of HON type II model (R2 and RMSE) are compared
Table 5
with four different models in Table 4. The models used in the
Summary of sensitivity analysis.
comparison include Fleming’s hyperbolic model, multivariate
Parameter Network input RMSE (kN) polynomial regression model, RBF network and GRNN model.
From the summary of results, the HON type II has the best
All Ep Ab h i 554.40 performance (R2 ¼ 0.890), followed by RBF network (R2 ¼0.828).
s, , N s ,Nb Ab , % sand, % silt, % clay, pile Type
L
Settlement Ep Ab h i 2961.09 The polynomial regression model and hyperbolic model are at the
, N s ,Nb Ab , % sand, % silt, % clay, pile Type
hL i
bottom of the performance list with the later having the least
Ep Ab s, N s ,Nb Ab , % sand, % silt, % clay, pile Type 717.00 prediction quality among the other models based on the testing
h L i data (R2 ¼0.562). To assess the quality of load–settlement curves
Ns Ep Ab 1275.91
s, ,Nb Ab , % sand, % silt, % clay, pile Type
L predicted by HON type II, the simulations are compared with a
NbAb Ep Ab h i 767.47 selected pile loading test from training data in Fig. 5 and with a
s, , N s ,Nb Ab , % sand, % silt, % clay, pile type
L
Pile type Ep Ab h i 684.73 loading test from testing data in Fig. 6. Both Figs. 5 and 6
s, , N s ,Nb Ab , % sand, % silt, % clay, pile Type
L indicated that HON simulations correlate well with field mea-
Soil type Ep Ab h i 778.30 surements. A further comparison of type II HON, elastic and
s, , N s ,Nb Ab , pile Type
L
hyperbolic models were also made against the measured values
(Fig. 7). It is clearly observed from the figure that HON gives the
the network from diverging during training. The training was closest estimate of settlement and therefore most accurate.
terminated when no further improvement in the quality of
prediction in comparison with the test data is observed. Termina-
tion of training at this juncture is necessary in order to prevent 10000
over-learning and loss of generalization. After termination of the
training, the network simulations were compared with the test 9000
data, which is not part of the training. The assessment parameters 8000
in both training and testing are the roots mean square error
(RMSE) and the coefficient of determination (R2). 7000
The root mean square error is defined as follows:
6000
Load (kN)
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Pn 2
i ðyp ym Þ 5000
RMSE ¼ ð13Þ
n
4000
The coefficient of determination R is a measure of scatter or
lack of it between two sets of data. It is given by 3000
Measured data
P P P 2 2000 HON type II prediction
n ym yp ð ym Þð yp Þ
R2 ¼ P P
P P
ð14Þ Elastic solution (Poulos and Davis, 1980)
nð y2m Þð ym Þ2 nð y2p Þð yp Þ2 1000 Hyperbolic model (Fleming, 1992)
0
in which ym and yp are measured and predicted parameters, 0 5 10 15 20 25 30 35 40
respectively.
Settlement (mm)
Tables 3 and 4 summarize the training and testing results for
settlement and load functions, respectively. Fig. 7. Performance comparison of various settlement prediction methods.
820
Settlement per unit load (mm/kN) A. Ismail, D.-S. Jeng / Engineering Applications of Artificial Intelligence 24 (2011) 813–821
1.2 0.6
0.2 0.52
0 0.5
0 7 14 21 28 35 10 20 30 40 50
Pile length (m) Slenderness ratio, L/D ( mm)
0.5 1.2
0.45 1
0.8
0.4
0.6
0.35
0.4
0.3 0.2
0.25 0
0 0.05 0.1 0.15 0.2 0.25 0 20 40 60 80
Pile stiffness (GN/m) SPT Number of blows (shaft)
0.7
Settlement per unit load (mm/kN)
0.6
0.5
0.4
0.3
0.2
0.1
0
0 30 60 90 120 150
SPT Number of blows (base)
Fig. 8. Sensitivity of HON type II model to: (a) pile embedded length, (b) pile stiffness, (c) average SPT blow counts along the pile shaft, and (d) SPT blow counts at the
pile base.
3.4. Parametric study and sensitivity analysis indicative of the good agreement between the HON model and
actual pile behaviour, where the capacity of pile increases with the
The results of parametric study carried out to assess the shaft area (a function of length and perimeter) and the resistance
generalization ability of HON type II model are presented in to compression is proportional to stiffness. The settlement is
Fig. 8(a)–(e). The approach used in the analysis is similar to the observed (in Fig. 8(c)) to vary proportionally with slenderness
one used by Shahin et al. (2004), where all input variables, except ratio (with diameter varied while length is kept constant). This is in
the subject of the analysis are re-fixed to the mean of training agreement with the study carried out by Poulos (1977). With
values while the subject parameter is allowed to vary between regard to the behaviour of foundation soil, the average number of
the minimum and maximum values used for model training. The blows along the shaft varies inversely with settlement; although
response of the model to the variation of the single parameter is the curve is not very smooth (Fig. 8(d)). Inverse relationship is also
examined in the light of the current knowledge about the observed in the case SPT-N for the pile base (Fig. 8(e)).
phenomenon. The analysis is then repeated with different con- Sensitivity analysis has also been carried out to see how
trolling parameters. strongly various input parameters affect the HON output. To test
Based on the results obtained, it is noticeable that there is a a particular parameter, its values are taken away from the input
good agreement with the HON type II response and expected vector and the performance of the model is evaluated. The results
behaviour of piles when subjected to axial loading. The consistent of the analysis are summarized in Table 5. It can be seen that all
decrease in the predicted pile head settlement (per unit load) with input parameters considered in this model significantly affect the
increment in pile length and stiffness in Fig. 8(a) and (b) is quality of HON prediction.
A. Ismail, D.-S. Jeng / Engineering Applications of Artificial Intelligence 24 (2011) 813–821 821
4. Conclusions Fleming, W., 1992. A new method of single pile settlement and analysis.
Geotechnique 42 (3), 411–425.
Fleming, W.G.K., Weltman, A.J., Randolf, M.F., Elson, W.K., 1992. A New Method of
Due to the complexity of soil–pile interaction and the soil Single Piling Engineering. John Wiley & Sons, New York.
disturbance during sampling and pile installation, the conven- Foresti, G.L., Dolso, T., 2004. An adaptive high-order neural tree for pattern
tional methods of predicting pile load–settlement relationship recognition. IEEE Transactions on Systems, Man, and Cybernetics, Part B:
Cybernetics 34 (2), 988–996.
suffer from lack of reliability. In this paper, a high-order network Goh, A.T.C., 1995. Modeling soil correlations using neural networks. Journal of
to simulate the pile load–settlement curve using simple input Computing in Civil Engineering, ASCE 9 (4), 275–278.
data is proposed. The input data consist of the average value of Goh, A.T.C., 1996. Pile driving records reanalyzed using neural networks. Journal of
Geotechnical Engineering, ASCE 122 (6), 492–495.
SPT along the pile shaft, the SPT value at the pile base, the pile Haykin, S., 1999. Neural networks. A comprehensive foundation, 2nd ed. Prentice-
stiffness, the shaft and base area, and the pile load. Other Hall, Upper Saddle River.
parameters include soil type and installation method. Based on Ismail, A., Engelbrecht, A.P., 2000. Global optimization algorithms for training
product-unit neural networks. In: Proceedings of the IEEE-INNS-ENNS Inter-
the coefficient of determination and root mean squared error,
national Joint Conference on Neural Networks (IJCNN 2000), Como Italy, vol. 1,
as well as the quality of load–settlement curves, a significant pp. 132–137.
improvement was observed from the comparison of HON results Ivakhnenko, A.G., 1971. Polynomial theory of complex systems. IEEE Transactions
with BPN, elastic and hyperbolic models. The proposed model was on Systems, Man, and Cybernetics SMC-1, 364–378.
Jaksa, M.B., Maier, H.R., Shahin, M.A., 2008. Future challenges for artificial neural
also found to be more accurate in prediction than RBF, GRNN and network modelling. In: The 12th International Conference of International
polynomial regression. Also, based on the parametric study Association for Computer Methods and Advances in Geomechanics. pp.
carried out, the HON model was found to respond reasonably 1710–1719.
Lee, I.M., Lee, J.H., 1996. Prediction of pile bearing capacity using artificial neural
well to various input parameters in a manner consistent with the networks. Computers and Geotechnics 18 (3), 189–200.
anticipated behaviour of an axially loaded pile. Lee, I.M., Salgado, R., 1999. Determination of pile base resistance n sands. Journal
of Geotechnical and Geoevironmental Engineering, ASCE 125 (8), 122–131.
McCulloch, W.S., Pitts, W., 1943. A logical Calculus of the ideas Immanent in
nervous activity. Bulletin of Mathematical Biophysics 5, 115–133.
Appendix A. Supporting information Nawari, N.O., Liang, R., Nusairat, J., 1999. Artificial intelligence techniques for the
design and analysis of deep foundations. Electronic Journal of Geotechnical
Supplementary data associated with this article can be found Engineering /http://geotech.civeng.okstate.edu/ejge/ppr9909/index.htmlS.
Nejad, F.P., Jaksa, M.B., Kakhi, M., McCabe, B.A., 2009. Prediction of pile settlement
in the online version at doi:10.1016/j.engappai.2011.02.008. using artificial neural networks based on standard penetration test data.
Computers and Geotechnics 36 (2009), 1125–1133.
Penumadu, D., Jean-Lou, C., 1997. Geomaterial modeling using artificial neural
networks. Artificial Neural Networks for Civil Engineers: Fundamentals and
References Applications, ASCE, 160–184.
Prevost, J.H., Popescu, R., 1996. Constitutive relations for soil materials. Electronic
Journal of Geotechnical Engineering /http://www.ejge.com/1996/Ppr9609/
Abdelbar, A.G., Tagliarini, G., 1996. HONEST: a new high order feed forward neural Abs9609.htmS.
network. In: Proceedings of the IEEE International Conference on Neural Poskitt, T.J., 1993. A new method for single pile settlement prediction and
Networks, vol. 2, pp. 1257–1262. analysis—discussion. Geotechnique 43 (4), 615–619.
Agrawal, G., Chameau, J.A., Bourdeau, P.L., 1997. Assessing the liquefaction Poulos, H.G., Davis, E.H., 1980. Pile Foundation Analysis and Design. Wiley.
susceptibility at a site based on information from penetration testing. In: Poulos, H.G., 1977. Settlement of Pile Foundations. Numerical Methods in
Kartam, N., Flood, I., Garrett, J.H. (Eds.), Artificial Neural Networks for Civil Geotechnical Engineerng. McGraw-Hill, New York, pp. 326–363.
Engineers: Fundamentals and Applications, New York, USA, pp. 185–214. Randolph, M.F., Wroth, C.P., 1978. Analysis of deformation of vertically loaded
Artyomov, E., Yadid-Pecht, O., 2005. Modified high-order neural network for piles. Journal of Geotechnical Engineering Division 104, 1465–1488.
invariant pattern recognition. Pattern Recognition Letters 26 (6), 843–851. Rovithakis, G.A., Chalkiadakis, I., Zervakis, M.E., 2004. High-order neural network
Bahi, J.M., Contassot-Vivier, S., Sauget, M., 2009. An incremental learning algo- structure selection for function approximation applications using genetic
rithm for function approximation. Advances in Engineering Software 40, algorithms. IEEE Transactions on Systems, Man, and Cybernetics, Part B:
725–730. Cybernetics 34 (1), 150–158.
Balakrishnan, E.G., Balasubramaniam, A.S., Phien-wej, N., 1999. Load deformation Rumelhart, D.E., Hinton, G.E., Williams, R.J., 1986. Learning internal representation
analysis of bored piles in residual weathered formation. Journal of Geotechni- by error propagation. In: Rumelhart, D.E., McClelland, J.L. (Eds.), Parallel
cal & Geoenvironmental Engineering, ASCE 125 (2), 122–131. Distributed Processing, vol. 1. MIT Press, Cambridge, MA (Chapter 8).
Bateni, S.M., Borghei, S.M., Jeng, D.-S., 2007. Neural network and neuro-fuzzy Scott, R.F. (1988). Constitutive relations for soil: present and future. In: Saada and
assessments for scour depth around bridge piers. Engineering Applications of Bianchini (Eds.), Constitutive Equations for Granular Non-cohesive Soils,
Artificial Intelligence 20, 401–414. Rotterdam.
Basheer, Najjar, Y.M., 1998. Modeling cyclic constitutive behavior by neural Shahin, M.A., Jaksa, M.B., Maier, H.R., 2002. Predicting settlement of shallow
networks: theoritical and real data. In: Proceedings of the 12th Engineering foundations using neural networks. Journal of Geotechnical and Geoenviron-
Mechanics Conference, La Jolla, California, 17–20 May, pp. 952–955. mental Engineering 128 (9), 785–793.
Chan, W.T., Chow, Y.K., Liu, L.F., 1995. Neural network: an alternative to pile Shahin, M.A., Jaksa, M.B., Maier, H.R., 2003. Settlement prediction of shallow
driving formulas. Computers and Geotechnics 17, 135–156. foundations on granular soils using B-spline neurofuzzy models. Computers
Chang, M.F., Broms, B.B., 1991. Design of bored piles in residual soils based on and Geotechnics 30 (8), 637–647.
field-performance data. Canadian Geotechnical Journal 28, 200–209. Shahin, M.A., Maier, H.R., Jaksa, M.B., 2004. Data division for developing neural
Chin, F.K., 1970. Estimation of the ultimate load of piles from tests not carried to networks applied to geotechnical engineering. Journal of Computing in Civil
failure. In: Proceedings of the Second SE Asian Conference on Soil Engineering, Engineering 18 (2), 105–114.
Singapore, pp. 81–92. Tan, Y.C., Chen, C.S., Liew, S.S., 1998. Load transfer behaviour of cast-inplace bored
Coyle, H.M., Reese, L.C., 1966. Load transfer for axially loaded piles in clay. Journal piles in tropical residuals soils. In: Proceedings of the 13th Southeast Asian
of Soil Mechanics and Foundations Division, ASCE 92 (SM2), 1–26. Geotechnical Conferences, Taipei, pp. 563–571.
Das, S.K., Basudhar, P.K., 2006. Undrained lateral load capacity of piles in clay using Teh, C.I., Wong, K.S., Goh, A.T.C., Jaritngam, S., 1997. Prediction of pile capacity
artificial neural network. Computers and Geotechnics 33 (8), 454–459. using neural networks. Journal of Computing in Civil Engineering, ASCE 11 (2),
Desai, C.S., 1977. Deep Foundation. Numerical Methods in Geotechnical Engi- 129–138.
neerng. McGraw-Hill, New York, pp. 235–271. Tomlinson, M.J., 1995. Pile Design and Construction Practice, 6th edition Longman.
Desai, C.S., 1974. Numerical design – analysis for piles in sands. Journal of Ural, D., Saka, H., 1998. Liquefaction prediction by neural networks. Electronic
Geotechnical Engineering Division, ASCE 100 (GT6), 613–635. Journal of Geotechnical Engineering /http://geotech.civen.okstate.edu/ejge/
Durbin, R., Rumelhart, R., 1989. Product units: a computationally powerful and ppr9803/index.htmlS.
biologically plausible extension to backpropagation networks. Neural Com- Vesic, A.S., 1977. Design of pile foundations, National Cooperative Highway
puting 1, 133–142. Research Program, Synthesis of Practice No. 42. Transportation Research
Ellis, G.W., Yao, C., Zhao, R., Penumadu, D., 1995. Stress–strain modelling of sands Board, Washington, DC.
using artificial neural networks. Journal of Geotechnical Engineering, ASCE Vijayvergiya, V.N., 1977. Load movement characteristics of piles. In: Proceedings
121 (5), 429–435. of the Ports’77 Conference, Long Beach, California.