See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/272251584

Prediction of ultimate bearing capacity of eccentrically inclined loaded strip

footing by ANN, part I

Article in International Journal of Geotechnical Engineering · January 2013

DOI: 10.1179/1938636212Z.00000000012

CITATIONS READS

13 121

4 authors, including:

Rabi Behera C. R. Patra

National Institute of Technology Rourkela National Institute of Technology Rourkela
13 PUBLICATIONS 55 CITATIONS 45 PUBLICATIONS 309 CITATIONS

SEE PROFILE SEE PROFILE

Nagaratnam Sivakugan
James Cook University
190 PUBLICATIONS 1,910 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Principles of Foundation Engineering 9th Edition with Braja M Das View project

Cemented Paste Backfill Modification Using Different Types of Binders and Insight into its Mechanical, Micro-Structure, and Flow Properties View project

All content following this page was uploaded by Nagaratnam Sivakugan on 24 November 2016.

The user has requested enhancement of the downloaded file.

Prediction of ultimate bearing capacity of
eccentrically inclined loaded strip footing by
ANN: Part II
R. N. Behera1, C. R. Patra*1, N. Sivakugan2 and B. M. Das3
Laboratory model tests were conducted on a strip footing resting over dry sand bed subjected to
eccentrically inclined load in reinforced condition to determine the ultimate bearing capacity. Eccentrically
inclined load on a strip footing can be referred to as partially compensated when the line of load application
at the base of the footing is inclined toward the centerline of the foundation or reinforced when the line of
load application is inclined away from the centerline. Based on the model load test results, a neural network
model was developed to predict the reduction factor that will be used in computing the ultimate bearing
capacity of an eccentrically inclined loaded strip footing. This reduction factor (RF) is the ratio of the ultimate
bearing capacity of the footing subjected to an eccentrically inclined load to the ultimate bearing capacity
of the footing subjected to a centric vertical load. A thorough sensitivity analysis was carried out to evaluate
the parameters affecting the reduction factor. Based on the weights of the developed neural network
model, a neural interpretation diagram is developed to find out whether the input parameters have direct or
inverse effect on the output. A prediction model equation is established using the trained weights of the
neural network model. The results were compared with the developed empirical equation for the reduction
factor (Patra et al., 2012b). The ANN model results were found to be more accurate than the regression
equation proposed by Patra et al. (2012b) based on the laboratory model test data and the predictability
was reasonably good.
Keywords: Eccentrically inclined load, Reinforced condition, Ultimate bearing capacity, Reduction factor, Sand, Neural network

Introduction results of laboratory model tests to estimate the reduction

factor. Reduction factor (RF) is the ratio of the ultimate
Shallow strip footings are at times subjected to eccentrically bearing capacity of the strip footing subjected to an
inclined loads. Figure 1 shows two possible modes of load eccentrically inclined load to the ultimate bearing capacity
application. In this figure, B is the width of the footing, e is of the strip footing subjected to a centric vertical load.
the load eccentricity, a is the load inclination and Qu is the Artificial neural network (ANN) is an artificial intelligence
ultimate load per unit length of the footing. In Fig. 1a, the system inspired by the behavior of human brain and
line of load application of the footing is inclined toward the nervous system. Backpropagation neural network is most
centerline and was referred to as partially compensated suitable for prediction problems and Levenberg–
(Perloff and Baron, 1976). It is also possible for the line of Marquadrt algorithm is adopted as it is efficient in
load application on the footing to be inclined away from the comparison to gradient descent backpropagation algo-
centerline as shown in Fig. 1b. Perloff and Baron (1976) rithm (Goh et al., 2005; Das and Basudhar, 2006). In the
called this type of loading as reinforced case. present study, the feedforward backpropagation neural
Laboratory model tests were conducted with the load network is trained with Levenberg–Marquadrt algorithm
application as per Fig. 1b. The present paper deals with which is known as Levenberg–Marquadrt neural network.
the development of a neural network model using the Different sensitivity analysis was carried out to study the
important parameters. Neural Interpretation Diagram
1
National Institute of Technology, Rourkela, India
(NID) is constructed based on the weights of the
2
James Cook University, Townsville, Qld, Australia developed neural network model, to find out direct or
3
California State University, Sacramento, CA, USA inverse effect of input parameters on the output. A
*Corresponding author, email [email protected] prediction model equation is also established using the

ß 2013 W. S. Maney & Son Ltd

Received 3 February 2012; accepted 26 April 2012 International Journal of
DOI 10.1179/1938636213Z.00000000019 Geotechnical Engineering 2013 VOL 7 NO 2 165
Behera et al. Ultimate bearing capacity of eccentrically inclined loaded strip footing by ANN: Part II

a partially compensated case; b reinforced case

1 Eccentrically inclined load on strip foundation

trained weights of the neural network model. Finally, the are presented in Table 2. They are computed from the
predicted RF values were compared with those derived database. The schematic diagram of ANN architecture is
empirically by Patra et al. (2012b). shown in Fig. 2. The number of hidden layer neurons is
varied with mean square error (MSE). The minimum
Database and preprocessing MSE is found to be 0?001 for four neurons in the hidden
layer (Fig. 3). Therefore, the final ANN architecture used
The database available in Patra et al. (2012b) has been in this study will be 3–4–1, i.e. 3 (input)–4 (hidden layer
used in the present study. Load tests were carried out on neuron)–1 (output), as shown in Fig. 4. MSE is defined as
model strip footings subjected to eccentrically inclined
n 
P 2
loads in the manner as shown in Fig. 1b that were RFi {RFp
increased to failure. The details of the tests and the
MSE~ i~1 (2)
procedure have been described in detail by Patra et al. n
(2012a, b). The data consists of parameters like load Coefficient of efficiency, R2 is expressed as
eccentricity (e), load inclination (a), embedment ratio (Df /
B), friction angle (w), and ultimate bearing capacity (qu). E1 {E2
R2 ~ (3)
Seventy-eight laboratory model tests were conducted. The E1
input parameters for the ANN model are e/B, a/w and Df /
where
B and the output is the RF. The RF is given by
quðDf =B,e=B,a=wÞ X
n  2
RF ~ (1) E1 ~ RFi {RF (3a)
quðDf =B,e=B~0,a=w~0Þ i~1

where quðDf =B,e=B,a=wÞ is the ultimate bearing capacity with and

eccentricity ratio e/B and inclination ratio a/w at an X
n  2
embedment ratio Df /B and quðDf =B,e=B~0,a=w~0Þ is the E2 ~ RFi {RFp (3b)
ultimate bearing capacity with centric vertical loading (e/ i~1
B50 and a/w50) at the same embedment ratio Df /B.
Out of 78 test records shown in Table 1, 59 tests are where RFi, RF and RFp are the experimental, average
considered for training and the remaining 19 are reserved experimental, and predicted RF values respectively, and n
for testing. Each record represents a complete model test is the number of training data
where an eccentrically inclined loaded strip footing was The coefficient of efficiency (R2) for the training and
subjected to failure. All the variables (i.e. inputs and testing data are found to be 0?994 and 0?988 respectively
output) are normalized in the range [21, 1] before (Figs. 5 and 6). Data used in this analysis have been
training. A feedforward backpropagation neural network obtained from laboratory model tests carried out in
is used with hyperbolic tangent sigmoid function and a duplicate, in a calibration chamber, the details of which
linear function as the transfer function. The network is are given in Patra et al. (2012a, b). All the data used in the
trained (learning) with Levenberg–Marquardt algorithm training and the testing are from the same source and are
as it is efficient in comparison to gradient descent of same nature. Probably, this may be one of the causes
backpropagation algorithm (Goh et al. 2005; Das and for better fitting in both testing and training phase as well.
Basudhar 2006). The ANN has been implemented using The weights and biases of the network are presented in
MATLAB V 7?11?0 (R2010b). Table 3. The weights and biases can be utilized for
interpretation of the relationships between the inputs
and output, to carry out a sensitivity analysis, and for
Results and discussion framing an ANN model in the form of an equation that
The maximum, minimum, average, and standard devia- can be used for predicting RF. The residual analysis
tion values of the three input and one output parameters was carried out by calculating the residuals from the

166 International Journal of Geotechnical Engineering 2013 VOL 7 NO 2

 Behera et al. Ultimate bearing capacity of eccentrically inclined loaded strip footing by ANN: Part II

Table 1 Dataset used for training and testing of ANN model (Patra et al., 2012b)

Calculated
Data Expt no. e/B Df/B a/w Experimental qu Experimental RF (equation
type (1) (2) (3) (4) (5) (kN/m2) (6) RF (equation (1)) (7) (19)) (8)

Training 1 0.05 0 0.123 113.80 0.682 0.740

2 0.1 0 0.123 107.91 0.647 0.658
3 0.05 0 0.245 88.29 0.529 0.590
4 0.1 0 0.245 85.35 0.512 0.525
5 0.15 0 0.245 81.42 0.488 0.459
6 0.05 0 0.368 68.67 0.412 0.453
7 0.15 0 0.368 64.75 0.388 0.352
8 0.1 0 0.490 51.99 0.312 0.291
9 0.15 0 0.490 49.05 0.294 0.255
10 0 0.5 0.000 264.87 1.000 1.000
11 0.05 0.5 0.123 196.20 0.741 0.774
12 0.1 0.5 0.123 173.64 0.656 0.688
13 0.15 0.5 0.123 152.06 0.574 0.602
14 0.05 0.5 0.245 166.77 0.630 0.651
15 0.1 0.5 0.245 151.07 0.570 0.579
16 0.05 0.5 0.368 137.34 0.519 0.531
17 0.15 0.5 0.368 112.82 0.426 0.413
18 0.1 0.5 0.490 105.95 0.400 0.369
19 0.15 0.5 0.490 95.16 0.359 0.323
20 0 1 0.000 353.16 1.000 1.000
21 0.05 1 0.123 284.49 0.806 0.811
22 0.1 1 0.123 251.14 0.711 0.721
23 0.05 1 0.245 249.17 0.706 0.719
24 0.15 1 0.245 203.07 0.575 0.559
25 0.05 1 0.368 217.78 0.617 0.624
26 0.1 1 0.368 193.26 0.547 0.554
27 0.15 1 0.368 171.68 0.486 0.485
28 0.1 1 0.490 156.96 0.444 0.467
29 0.15 1 0.490 143.23 0.406 0.408
30 0 0 0.000 101.04 1.000 1.000
31 0.1 0 0.133 62.78 0.621 0.645
32 0.15 0 0.133 52.97 0.524 0.565
33 0.05 0 0.267 56.90 0.563 0.565
34 0.1 0 0.267 51.99 0.515 0.502
35 0.05 0 0.400 42.58 0.421 0.418
36 0.15 0 0.400 38.65 0.383 0.325
37 0.05 0 0.533 31.39 0.311 0.287
38 0.1 0 0.533 30.41 0.301 0.255
39 0.15 0 0.533 29.43 0.291 0.223
40 0 0.5 0.000 143.23 1.000 1.000
41 0.05 0.5 0.133 105.95 0.740 0.763
42 0.1 0.5 0.133 94.18 0.658 0.679
43 0.15 0.5 0.133 77.50 0.541 0.594
44 0.1 0.5 0.267 77.50 0.541 0.560
45 0.15 0.5 0.267 67.69 0.473 0.490
46 0.05 0.5 0.400 73.58 0.514 0.500
47 0.1 0.5 0.400 63.77 0.445 0.445
48 0.05 0.5 0.533 58.86 0.411 0.375
49 0.15 0.5 0.533 48.07 0.336 0.291
50 0 1 0.000 208.95 1.000 1.000
51 0.1 1 0.133 156.96 0.751 0.713
52 0.15 1 0.133 144.21 0.690 0.624
53 0.05 1 0.267 148.13 0.709 0.702
54 0.1 1 0.267 135.38 0.648 0.624
55 0.05 1 0.400 124.59 0.596 0.598
56 0.1 1 0.400 114.78 0.549 0.532
57 0.15 1 0.400 103.01 0.493 0.465
58 0.05 1 0.533 99.08 0.474 0.489
59 0.15 1 0.533 86.33 0.413 0.380

International Journal of Geotechnical Engineering 2013 VOL 7 NO 2 167

Behera et al. Ultimate bearing capacity of eccentrically inclined loaded strip footing by ANN: Part II

Table 1 Continued

Calculated
Data Expt no. e/B Df/B a/w Experimental qu Experimental RF (equation
type (1) (2) (3) (4) (5) (kN/m2) (6) RF (equation (1)) (7) (19)) (8)

Testing 1 0 0 0.000 166.77 1.000 1.000

2 0.15 0 0.123 92.21 0.553 0.575
3 0.1 0 0.368 66.71 0.400 0.402
4 0.05 0 0.490 53.96 0.324 0.328
5 0.15 0.5 0.245 132.44 0.500 0.507
6 0.1 0.5 0.368 129.49 0.489 0.472
7 0.05 0.5 0.490 113.80 0.430 0.415
8 0.15 1 0.123 228.57 0.647 0.630
9 0.1 1 0.245 225.63 0.639 0.639
10 0.05 1 0.490 179.52 0.508 0.525
11 0.05 0 0.133 71.61 0.709 0.726
12 0.15 0 0.267 49.05 0.485 0.440
13 0.1 0 0.400 41.20 0.408 0.372
14 0.05 0.5 0.267 88.29 0.616 0.630
15 0.15 0.5 0.400 56.90 0.397 0.389
16 0.1 0.5 0.533 53.96 0.377 0.333
17 0.05 1 0.133 170.69 0.817 0.803
18 0.15 1 0.267 120.66 0.577 0.546
19 0.1 1 0.533 92.21 0.441 0.435

experimental reduction factor and predicted reduction sums the products across all the hidden neurons, which is
factor for training data set. Residual (er) can be defined as defined as Si. The relative inputs are corresponding to
the difference between the experimental and predicted RF absolute Si values, with the most important input
value and is given by corresponding to the highest Si value. The details of
connection weight approach are presented by Olden et al.
er ~RFi {RFp (4) (2004).
The residuals are plotted with the experiment number as The sensitivity analysis based on Pearson correlation
shown in Fig. 7. It is observed that the residuals are coefficient is presented in Table 4 which shows the cross
distributed evenly along the horizontal axis of the plot. correlation of inputs with the reduction factor. It is seen
Therefore, it can be said that the network is well trained that the parameters (e/B) and (a/w) are inter-related with a
and can be used for prediction with reasonable accuracy. cross-correlation value of 0?3. This is possibly due to the
reinforcing effect. From the table, it is observed that RF is
Sensitivity analysis highly correlated to a/w with a cross-correlation values of
0?87, followed by e/B (50?59) and Df /B (50?28). The
Sensitivity analysis is of utmost concern for selection of
sensitivity analysis for the model as per Garson’s
important input variables. Different approaches have been
algorithm is presented in Table 5. The a/w is found to be
suggested to select the important input variables. The
the most important input parameter with the relative
Pearson correlation coefficient is defined as one of the
importance value being 37?3% followed by 34?1% for Df /B
variable ranking criteria in selecting proper inputs for the
and 28?6% for e/B. The relative importance of the input
ANN (Guyon and Elisseeff, 2003; Wilby et al., 2003). Goh
variables as calculated following connection weight
(1994) and Shahin et al. (2002) have used Garson’s approach (Olden et al., 2004) is also presented in
algorithm (Garson, 1991) in which the input-hidden and Table 5. As per connection weight approach method, a/w
hidden-output weights of the trained ANN model are is found to be the most important input parameter (Si
partitioned and the absolute values of the weights are used value: 22?13) followed by e/B (Si value: 21?69) and Df /B
in selecting the important input variables, the details of (Si value: 0?41). The Si values being negative imply that
which with example have been presented by Goh (1994). It both a/w and e/B are indirectly and Df /B is directly related
does not provide information on the effect of input to RF values.
variables in terms of direct or inverse relation to the
output. Olden et al. (2004) proposed a connection weights Neural Interpretation Diagram (NID)
approach based on the NID, in which the actual values of Ozesmi and Ozesmi (1999) proposed neural interpretation
input-hidden and hidden-output weights are considered. It diagram for visual interpretation of the connection weight

Table 2 Statistical values of the parameters

Parameter Maximum value Minimum value Average value Standard Deviation

e/B 0.15 0 0.092 0.047

Df /B 1 0 0.5 0.408
a/w 0.533 0 0.295 0.162
RF 1 0.291 0.56 0.179

168 International Journal of Geotechnical Engineering 2013 VOL 7 NO 2

 Behera et al. Ultimate bearing capacity of eccentrically inclined loaded strip footing by ANN: Part II

2 ANN architecture

neurons represent the weights. The positive weights are

represented by solid lines and negative weights by dashed
lines and the thickness of the lines is proportional to their
magnitude. The relationship between the input and
output is determined in two steps. Direct proportionality
of the input variables is depicted by positive input-hidden
and positive hidden-output weights, or negative input-
hidden and negative hidden-output weights. The positive
input-hidden and negative hidden-output; negative input-
hidden and positive hidden-output weights indicate the
inverse proportionality of the input variables. Therefore,
the multiplication of actual weights of input-hidden and
hidden-output rather than multiplication of absolute
weights indicates the effect of that input variable on the
output. The input directly related to the output is
represented with a grey circle and that having inverse
effect with blank circle.
It can be seen from Fig. 4 and fourth column of Table 5
3 Variation of hidden layer neuron with mean square error
that e/B and a/w are inversely related to RF values,
(MSE)
whereas Df /B is directly related to RF. It can be concluded
that RF value decreases with increase in e/B and a/w, but
among the neurons. For the present study with the weights increases with increasing Df /B. In other words, the
obtained as shown in Table 3, an NID is presented in Fig. 4. developed ANN model is not a ‘‘black box’’ and could
The lines joining the input-hidden and hidden-output explain the physical effect of inputs on the output.

4 Neural Interpretation Diagram (NID) showing lines representing connection weights and effects of inputs on reduction fac-
tor (RF)

International Journal of Geotechnical Engineering 2013 VOL 7 NO 2 169

Behera et al. Ultimate bearing capacity of eccentrically inclined loaded strip footing by ANN: Part II

5 Correlation between predicted reduction factor with 6 Correlation between predicted reduction factor with
experimental reduction factor for training data experimental reduction factor for testing data

Step 1
The input parameters were normalized in the range [21, 1]
ANN model equation for RF based on trained by the following expressions
neural network  
A model equation is developed using the weights obtained X1 {Xmin
Xn ~2 (6)
from trained neural network model (Goh et al., 2005). The Xmax {Xmin
mathematical equation relating input parameters (e/B, Df / where Xn is the normalized value of input parameter X1,
B, a/w) to output (RF) can be given by and Xmax and Xmin are maximum and minimum values of
( " !#)
X h X
m input parameter X1 in the data set.
RFn ~fn b0 z wk fn bhk z wik Xi (5)
k~1 i~1 Step 2
where RFn is the normalized value of RF in the range [21, Calculate the normalized value of reduction factor (RFn)
1], fn is the transfer function, h is the no. of neurons in the using the following expressions
hidden layer, Xi is the normalized value of inputs in the e    
: : Df : a
range [21, 1], m is the no. of input variables, wik is the A1 ~{0 4188 z0 3928 {0 8511 {0:005
B n B n w n
connection weight between ith layer of input and kth (7)
neuron of hidden layer, wk is the connection weight
e    
between kth neuron of hidden layer and single output Df a
neuron, bhk is the bias at the kth neuron of hidden layer, A2 ~0:17 z5:2462 z3:1071 z3:874 (8)
B n B n w n
and bo is the bias at the output layer.
The model equation for the RF of a shallow strip e    
Df a
footing subjected to eccentrically inclined load as shown in A3 ~4:0168 z8:1273 {4:8781 z7:3396
B n B n w n
Fig. 1b was formulated using the values of the weights and (9)
biases shown in Table 3 as per the following steps.
Table 3 Values of connection weights and biases

Weight

wik wk Bias

Neuron e/B Df /B a/w RF bhk B0

Hidden neuron 1 (k51) 20.4188 0.3928 20.8511 0.7337 20.005 0.0972

Hidden neuron 2 (k52) 0.17 5.2462 3.1071 20.1013 3.874
Hidden neuron 3 (k53) 4.0168 8.1273 24.8781 0.0693 7.3396
Hidden neuron 4 (k54) 6.2788 20.3513 3.247 20.2621 4.8671

170 International Journal of Geotechnical Engineering 2013 VOL 7 NO 2

 Behera et al. Ultimate bearing capacity of eccentrically inclined loaded strip footing by ANN: Part II

7 Residual distribution of training data

e    
Df a
A4 ~6:2788 {0:3513 z3:247 z4:8671 8 Comparison of ANN results with experimental RF and
B n B n w n Patra et al. (2012b) for training data
(10)
 A1 
e {e{A1
B1 ~0:7337 (11)
eA1 ze{A1
 A2  RFn ~C1 (16)
e {e{A2
B2 ~{0:1013 A (12)
e 2 ze{A2
 A3  Step – 3
: e {e{A3 Denormalize the RFn value obtained from equation (16)
B3 ~0 0693 A (13)
e 3 ze{A3 to actual RF as
 A3  RF ~0:5(RFn z1)(RFmax {RFmin )zRFmin (17)
: e {e{A3
B4 ~{0 2621 A (14)
e 3 ze{A3
RF ~0:5(RFn z1)(1{0:291)z0:291 (18)
C1 ~0:0972zB1 zB2 zB3 zB4 (15)

Comparison
Table 4 Cross-correlation of input and output for reduction Patra et al. (2012b) proposed an empirical equation for
factor reduction factor (RF) and can be expressed as
Parameters e/B Df /B a/w RF quðDf =B,e=B,a=wÞ h  e i  
a ð1:5{0:7Df =BÞ
RF ~ ~ 1{2 1{
e/B 1 0 0.3 20.59 quðDf =B,e=B~0,a=w~0Þ B w
Df /B 1 0 0.28 (19)
a/w 1 20.87 Figures 8 and 9 show the comparison of RF obtained from
RF 1
equation 17 with equations (19) and (1). It can be seen

Table 5 Relative importance of different inputs as per Garson’s algorithm and connection weight approach

Parameters Garson’s algorithm Connection weight approach

Ranking of Si values as Ranking of inputs

Relative inputs as per per connection as per relative
(1) Importance (%) (2) relative importance (3) weight approach (4) importance (5)

e/B 28.6 3 21.6918 2

Df /B 34.1 2 0.412 3
a/w 37.3 1 22.128 1

International Journal of Geotechnical Engineering 2013 VOL 7 NO 2 171

Behera et al. Ultimate bearing capacity of eccentrically inclined loaded strip footing by ANN: Part II

connection weight approach is able to explore the input–

output relationship using trained weights.
5. The developed ANN model could explain the
physical effect of inputs on the output, as depicted in
NID. It was observed that e/B and a/w were inversely
related to RF values, whereas Df /B was directly related to
RF.
6. A model equation is developed based on the trained
weights of the ANN.
7. The predictability of ANN models are found to be
better than the empirical equation developed by Patra
et al. (2012b).

References
Das, S. K. and Basudhar, P. K. 2006. Undrained lateral load capacity of
piles in clay using artificial neural network, Comput. Geotech., 33,
(8), 454–459.
Garson, G. D. 1991. Interpreting neural-network connection weights,
Artif. Intell. Exp., 6, (7), 47–51.
Goh, A. T. C. 1994. Seismic liquefaction potential assessed by neural
network, J. Geotech. Eng., ASCE, 120, (90), 1467–1480.
Guyon, I. and Elisseeff, A. 2003. An introduction to variable and feature
9 Comparison of ANN results with experimental RF and
selection, J. Mach. Learn. Res., 3, 1157–1182.
Patra et al. (2012b) for testing data Goh, A. T. C., Kulhawy, F. H. and Chua, C. G. 2005. Bayesian neural
network analysis of undrained side resistance of drilled shafts,
that ANN results are much closer to line of equality than J. Geotech. Geoenv. Eng., ASCE, 131, (1), 84–93.
the empirical ones from Patra et al. (2012b). Ozesmi, S. L. and Ozesmi, U. 1999. An artificial neural network approach
to spatial modeling with inter specific interactions, Ecol. Model.,
116, 15–31.
Conclusion Olden, J. D., Joy, M. K. and Death, R. G. 2004. An accurate comparison
Based on the neural network model developed herein, the of methods for quantifying variable importance in artificial neural
networks using simulated data, Ecol. Model., 178, (3), 389–397.
following conclusions may be drawn. Patra, C. R., Behera, R. N., Sivakugan, N. and Das, B. M. 2012a.
1. As per residual analysis, the errors are distributed Ultimate bearing capacity of shallow strip foundation under
evenly along the horizontal axis (Fig. 7). It can be eccentrically inclined load: part I, Int. J. Geotech. Eng., 6, (3),
concluded that the network was well trained and can 343–352.
predict the result with reasonable accuracy. Patra, C. R., Behera, R. N., Sivakugan, N. and Das, B. M. 2012b.
Ultimate bearing capacity of shallow strip foundation under
2. Using correlation coefficient it was observed that a/w
eccentrically inclined load: part II, Int. J. Geotech. Eng., 6, (4),
is the most important input parameter followed by e/B and 507–514.
Df /B. Perloff, W. H. and Barron, W. 1976. Soil mechanics: principles and
3. Similarly, using Garson’s algorithm, a/w is found to applications; New York, Ronald Press.
be the most important input parameter followed by Df /B Shahin, M. A., Maier, H. R. and Jaksa, M. B. 2002. Predicting settlement
of shallow foundations using neural network, J. Geotech. Geoenv.
and e/B.
Eng., ASCE, 128, (9), 785–793.
4. Based on connection weight approach method, and Wilby, R. L., Abrahart, R. J. and Dawson, C. W. 2003. Detection of
Pearson correlation coefficient gave similar results. Hence, conceptual model rainfall-runoff processes inside an artificial neural
it may be concluded that sensitivity analysis using network, Hydrol. Sci., 48, (2), 163–181.

172 International Journal of Geotechnical Engineering 2013 VOL 7 NO 2

View publication stats