You are on page 1of 10

Engineering Geology 117 (2011) 29–38

Contents lists available at ScienceDirect

Engineering Geology
j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / e n g g e o

New approaches to determine the ultimate bearing capacity of shallow foundations


based on artificial neural networks and ant colony optimization
Adem Kalinli a,⁎, M. Cemal Acar a, Zeki Gündüz b
a
Kayseri Vocational College, University of Erciyes, Kayseri, Turkey
b
Department of Civil Engineering, University of Sakarya, Sakarya, Turkey

a r t i c l e i n f o a b s t r a c t

Article history: In this study, two different approaches are proposed to determine the ultimate bearing capacity of shallow
Received 20 November 2009 foundations on granular soil. Firstly, an artificial neural network (ANN) model is proposed to predict the
Received in revised form 29 September 2010 ultimate bearing capacity. The performance of the proposed neural model is compared with results of the
Accepted 2 October 2010
Adaptive Neuro Fuzzy Inference System, Fuzzy Inference System and ANN, which are taken in literature. It is
Available online 13 October 2010
clearly seen that the performance of the ANN model in our study is better than that of the other prediction
methods. Secondly, an improved Meyerhof formula is proposed for the computation of the ultimate bearing
Keywords:
Ultimate bearing capacity
capacity by using a parallel ant colony optimization algorithm. The results achieved from the proposed
Shallow foundations formula are compared with those obtained from the Meyerhof, Hansen and Vesic computation formulas.
Neural network Simulation results showed that the improved Meyerhof formula gave more accurate results than the other
Ant colony optimization theoretical computation formulas. In conclusion, the improved Meyerhof formula could be successfully used
for calculating the ultimate bearing capacity of shallow foundations.
Crown © 2010 Copyright Published by Elsevier B.V. All rights reserved.

1. Introduction researchers for many years (De Beer, 1965; Steenfelt, 1977; Tatsuoka
et al., 1991; Jacek and Ivo, 1999). The scale effect arising from the
A shallow foundation is a load carrying structure that transmits grain size of soil has a significant role if the foundation width to grain-
loads directly to the underlying soil. Shallow is a relative term. A size ratio is less than 50–100. Therefore, caution must be taken in
foundation with a depth to width ratio less than or equal to four (D/ applying the results of very small-scale model footing tests instead of
B ≤ 4) is simply called a shallow foundation (Das, 1999). A foundation full-scale behaviors. One simple solution to solve possible problems
must satisfy two fundamental requirements: ultimate bearing due to the scale effect is to use a larger footing, giving an acceptable
capacity and settlement of foundations. The bearing capacity of soil size ratio, B/D50 greater than 100 (Taylor, 1995). Although it is
can be defined as the foundations resistance when maximum pressure necessary to test the actual size footing to understand real soil-
is applied from the foundation to the soil without arising shear failure foundation behavior, to do so is an expensive, time consuming and
in the soil. The load per unit area of the foundation at which shear experimentally difficult process. For this reason, taking the scale effect
failure takes place is called the ultimate bearing capacity. By taking into consideration most researchers have only worked on small-scale
into account these two criteria, there are, and have always been, many footings of different sizes in the laboratory to obtain the ultimate
theories and many approaches in laboratory and in situ studies to bearing capacity. Researchers try to evaluate reliable methods based
determine the ultimate bearing capacity. Firstly, Prandtl (1921), and on the load test data of real sized foundations, and also smaller size
thereafter Reissner (1924) presented theories based on the concept of model footings to predict the ultimate bearing capacity.
plastic equilibrium. Later, the formulation was modified by Terzaghi Terzaghi (1943) formed a semi-empirical equation for computing
(1943), Meyerhof (1963), Hansen (1968), Vesic (1973) and others. the ultimate bearing capacity of a foundation. Later, Meyerhof (1963)
The ultimate bearing capacity depends on the size of the proposed a general bearing capacity equation similar to that of
foundation for both square and rectangular footings. Therefore, Terzaghi's which included different shape and depth factors. He took
small models of footings prepared in a laboratory are different from into account the shear strength of the soil above the base level of the
real size footings with regard to behavior and stress distribution. This footing. Thereafter, Hansen (1968) modified the study of Meyerhof.
is called the scale effect, and it has been studied by numerous Vesic (1973) used an equation very similar to that suggested by
Hansen (1968). However, there are some restrictions and assump-
tions in of all these classical formulations. Therefore, they do not
⁎ Corresponding author. Tel.: +90 352 4374901/40706. always give reasonable results compared to available experimental
E-mail address: kalinlia@erciyes.edu.tr (A. Kalinli). data. Because of the uncertain nature of soils and the difficulties of

0013-7952/$ – see front matter. Crown © 2010 Copyright Published by Elsevier B.V. All rights reserved.
doi:10.1016/j.enggeo.2010.10.002
30 A. Kalinli et al. / Engineering Geology 117 (2011) 29–38

experimental tests in the laboratory and in situ, there is an increasing (1963) suggested a classical general expression for the soil bearing
tendency to seek alternative bearing capacity prediction methods, capacity equation as
other than the traditional computing techniques, to obtain more
accurate results (Perloff and Baron, 1976; Al-shamsi, 1993). 1
qult = cNc Fcs Fcd Fci + γDNq Fqs Fqd Fqi + γBNγ Fγs Fγd Fγi ð1Þ
The great complexity and difficulties encountered in geotechnical 2
engineering such as slope stability, liquefaction, shallow foundation
where, qult is the ultimate bearing capacity of a shallow footing; C is
and pile capacity prediction have motivated researchers to use
the cohesion; Nc, Nq and Nγ are the cohesion, surcharge and density
powerful new optimization algorithms and methods. The most
bearing capacity factors, respectively. They are changed by the
popular of these new algorithms include genetic algorithms (GAs),
internal friction angles of cohesionless soil. Fcs, Fqs, Fγs are the footing
simulated annealing (SA), ant colony optimization (ACO), tabu search
shape factors; Fcd, Fqd, Fγd are the footing depth factors, Fci, Fqi, Fγi are
(TS), and artificial neural networks (ANNs) (Reeves, 1995; Corne
the footing inclination factors; γ is soil density; D is the depth of the
et al., 1999). These algorithms have been identified as potential
footing; B is the width of the footing. After extensive in situ and
solutions to many geotechnical engineering problems.
laboratory tests, Meyerhof (1963), Hansen (1968), Vesic (1973) and
One of the most popular prediction methods is the ANN which
several other researchers proposed different shape, depth and
simulates the biological neural network structure and the learning
inclination factors to use in this equation. For centric loading without
system. The ANN is a family of massively parallel structures that solve
foundation inclination on granular soil, the general bearing capacity of
various problems via the cooperation of highly interconnected but
the Meyerhof Equation becomes
simple computing elements called neurons. This allows the assess-
ment of non-linear relationships between any of the soil and 1
foundation parameters and also gives faster and better results qult = γDNq Fqs Fqd + γBNγ Fγs Fγd ð2Þ
2
compared to previous traditional methods. ANNs have been applied
to many geotechnical engineering problems (Goh, 1994; Ellis et al., Most geotechnical engineers accept Eq. (2) as the basic formula for
1995; Lee and Lee, 1996; Teh et al., 1997; Sivakugan et al., 1998; TRB, granular soils. The required parameters for the calculation of the
1999; Shahin et al., 2000; Rahman et al., 2001; Padmini et al., 2007; Meyerhof formula are given in Eqs. (3)–(7).
Shahin et al., 2009). Some of the above mentioned studies include the  
prediction of the ultimate bearing capacity of shallow foundations on ∅ 2
Kp = tan 45 + ð3Þ
cohesionless soils (Shahin et al., 2000; Padmini et al., 2007). 2
The ACO algorithm, another modern optimization method which  
simulates the behavior of real ant colonies, was proposed by Dorigo πtan∅ 2 ∅
Nq = e tan 45 + ð4Þ
et al. (1991). ACO algorithms have been applied to solve a number of 2
engineering problems in the literature. However, ACO applications in  
the field of geotechnical engineering are very limited. Changfu et al. Nγ = Nq −1 tanð1:4∅Þ ð5Þ
(2003), Gao (2005), Liang et al. (2008) and Kahatadeniye et al. (2009)
8
adopted the modified ant colony algorithm to locate critical slip < 1 + 0:2K B ;∅ N 10
p
surfaces. In these studies, their common reason for using the ACO Fqs = Fγs = L ð6Þ
:
algorithm was to find the slip surface with the minimum safety factor. 1;∅ = 0
Furthermore, to the best our knowledge no study related to determine
8 qffiffiffiffiffiffi
the ultimate bearing capacity by using the ACO algorithm has been < 1 + 0:1 K D ;∅ N 10
p
reported in the available literature. However, the ACO algorithm could Fqd = Fγd = B ð7Þ
:
be a useful method for developing an alternative ultimate bearing 1;∅ = 0
capacity computation formula.
In this study, two different approaches are proposed to determine The Meyerhof theory also differs from the basic Terzaghi solution.
the ultimate bearing capacity of shallow foundations on granular soil. In Meyerhof's theory, the influence of the shear strength of soil above
Firstly, a black box model based on ANN is proposed to predict the the base of the foundation was included. This means that the
ultimate bearing capacity. The performance of the proposed neural beneficial effect of foundation depth (e.g. surcharge) was included
model is compared to that of other prediction methods in the in the analysis. Failure plane analysis is slightly more complex when
literature. Secondly, by using the ACO algorithm, an improved the soil is still in plastic equilibrium, but has a log spiral failure surface
Meyerhof formula is proposed for the computation of the ultimate that includes shear above the base of the foundation.
bearing capacity. The results achieved from the proposed formula
were compared with those obtained from the Meyerhof, Hansen and 3. Prediction of the ultimate bearing capacity using ANN
Vesic formulas.
Section 2 briefly describes the theoretical background of the Artificial neural networks (ANNs) were developed in the form of
bearing capacity. The prediction of the ultimate bearing capacity using distributed network models based on the behavior and complex
ANN is presented in Section 3. The principles of ACO algorithms are functioning of the human brain. They are composed of several layers
presented in Section 4. In Section 5, the application of the ACO of many interconnected processing elements (neurons) operating in
algorithm to develop an improved Meyerhof formula is explained. The parallel. A simple neuron is shown in Fig. 1. Each neuron has an input
work is finally concluded in Section 6. (x), an output (y), a summation function and an activation function.
The multilayered perceptron (MLP) is one of the most popular
2. Theoretical background of bearing capacity ANN architectures. An MLP is very efficient for function approxima-
tion in high dimensional spaces and is composed of neurons and
The first extensive theory used to obtain the ultimate bearing layers connected to each other via weights (Du et al., 2002). The
capacity equation was expressed by Terzaghi (1943). However, in this performance of an MLP network depends mainly on the adjustment of
equation, the shearing resistance along the failure surface in the soil the weights (w) between the layers and neurons. The training process
above the bottom of the foundation was not considered. In addition, of an MLP network involves finding the connection weight values that
he did not consider the rectangular and inclined footings of shallow minimize the error function between the actual network output and
foundations. Taking into account all these missing factors, Meyerhof corresponding target values. This is achieved through a learning
A. Kalinli et al. / Engineering Geology 117 (2011) 29–38 31

Table 1
The data used for developing the neural model.

Source B (m) D (m) L/B γ (kN/m3) Φ (°) qu (kPa)

Muhs et al. (1969) 0.6 0.3 2 9.85 34.9 270


0.6 0 2 10.2 37.7 200
0.6 0.3 2 10.2 37.7 570
0.6 0 2 10.85 44.8 860
0.6 0.3 2 10.85 44.8 1760
Weiß (1970) 0.5 0 1 10.2 37.7 154
0.5 0 1 10.2 37.7 165
Fig. 1. The simple artificial neuron. 0.5 0 2 10.2 37.7 203
0.5 0 2 10.2 37.7 195
0.5 0 3 10.2 37.7 214
0.52 0 3.85 10.2 37.7 186
0.5 0.3 1 10.2 37.7 681
algorithm used to train the neural network. ANNs can be trained using 0.5 0.3 2 10.2 37.7 542
many different learning algorithms, such as back propagation (BP) 0.5 0.3 2 10.2 37.7 530
and Levenberg–Marquardt (LM) algorithms. These algorithms, which 0.5 0.3 3 10.2 37.7 402
are used in training ANNs, employ some form of gradient descent. This 0.52 0.3 3.85 10.2 37.7 413
Muhs and Weiß (1971) 0.5 0 1 11.7 37 111
is done by simply taking the derivative of the error function with 0.5 0 1 11.7 37 132
respect to the network weights, and then changing those weights in a 0.5 0 2 11.7 37 143
gradient-related direction. The LM algorithm is also known to provide 0.5 0.013 1 11.7 37 137
faster and better convergence than the BP algorithm. 0.5 0.029 4 11.7 37 109
0.5 0.127 4 11.7 37 187
One of the distinct characteristics of the ANN is its ability to learn
0.5 0.3 1 11.7 37 406
from experience and examples and to generalize them. Typically there 0.5 0.3 1 11.7 37 446
are two steps in neural methods: training and recalling. Training of an 0.5 0.3 4 11.7 37 322
ANN usually requires a large training pattern set. However, after 0.5 0.5 2 11.7 37 565
training, it can be used directly to substitute complex system 0.5 0.5 4 11.7 37 425
0.5 0 1 12.41 44 782
dynamics (Picton, 1994). 0.5 0 4 12.41 44 797
In this study, an MLP model, as shown in Fig. 2, was proposed to 0.5 0.3 1 12.41 44 1940
predict ultimate bearing capacities. In the model, B, D, L/B, γ, and Φ 0.5 0.3 1 12.41 44 2266
variables are the inputs to the network, and quann is the output of the 0.5 0.5 2 12.41 44 2847
0.5 0.5 4 12.41 44 2033
network. Where, B, D, L/B, γ, Φ and quann are the width of the footing,
0.5 0.49 4 12.27 42 1492
depth of the footing, footing geometry, unit weight of sand, angle of 0.5 0 1 11.77 37 123
the shearing resistance and the ultimate bearing capacity, respective- 0.5 0 2 11.77 37 134
ly. The training and testing data set given in Table 1 was taken from 0.5 0.3 1 11.77 37 370
previous experimental studies in the literature (Weiß, 1970; Muhs 0.5 0.5 2 11.77 37 464
0.5 0 4 12 40 461
and Weiß, 1971; Padmini et al., 2007). The database contains results 0.5 0.5 4 12 40 1140
of rectangular, square and strip footings of different sizes with centric Muhs and Weiß (1973) 1 0.2 3 11.97 39 710
loading and without foundation inclinations on sand beds of various 1 0 3 11.93 40 630
densities. It consists of 97 data sets, 47 of which deal with load tests on Briaud and Gibbens (1999) 0.991 0.711 1 15.8 32 1773.7
3.004 0.762 1 15.8 32 1019.4
large-scale footings and 50 with smaller scale sized model footings. In
2.489 0.762 1 15.8 32 1158
the large-scale test, the authors calculated the angles of shearing 1.492 0.762 1 15.8 32 1540
resistance for axisymmetric and plain strain conditions. In smaller 3.016 0.889 1 15.8 32 1161.2
sized model tests the angles of shearing resistance were calculated Gandhi (2003) 0.0585 0.029 5.95 15.7 34 58.5
from direct shear tests. In the case of large-scale footings, the ultimate 0.0585 0.058 5.95 15.7 34 70.91
0.0585 0.029 5.95 16.1 37 82.5
load is defined as corresponding to the point where the slope of the 0.0585 0.058 5.95 16.1 37 98.93
load settlement curve reaches a minimum value; in smaller sized 0.0585 0.029 5.95 16.5 39.5 121.5
model footings it is defined as the breaking point of the load 0.0585 0.058 5.95 16.5 39.5 142.9
settlement curve in a log–log-plot. 0.0585 0.029 5.95 16.8 41.5 157.5
0.0585 0.058 5.95 16.8 41.5 184.9
0.0585 0.029 5.95 17.1 42.5 180.5
0.0585 0.058 5.95 17.1 42.5 211
0.094 0.047 6 15.7 34 74.7
0.094 0.094 6 15.7 34 91.5
0.094 0.047 6 16.1 37 104.8
0.094 0.094 6 16.1 37 127.5
0.094 0.047 6 16.5 39.5 155.8
0.094 0.094 6 16.5 39.5 185.6
0.094 0.047 6 16.8 41.5 206.8
0.094 0.094 6 16.8 41.5 244.6
0.094 0.047 6 17.1 42.5 235.6
0.094 0.094 6 17.1 42.5 279.6
0.152 0.075 5.95 15.7 34 98.2
0.152 0.15 5.95 15.7 34 122.3
0.152 0.075 5.95 16.1 37 143.3
0.152 0.15 5.95 16.1 37 176.4
0.152 0.075 5.95 16.5 39.5 211.2
0.152 0.15 5.95 16.5 39.5 254.5
0.152 0.075 5.95 16.8 41.5 285.3

(continued on next page)


Fig. 2. Neural model for ultimate bearing capacity prediction of shallow foundations.
32 A. Kalinli et al. / Engineering Geology 117 (2011) 29–38

Table 1 (continued) measured and computed ultimate bearing capacity were the same
Source B (m) D (m) L/B γ (kN/m3) Φ (°) qu (kPa) and good for the four methods. The RMSE and MBE values of the
Gandhi (2003) 0.152 0.15 5.95 16.8 41.5 342.5 proposed neural model and ANFIS were the same and the lowest in
0.152 0.075 5.95 17.1 42.5 335.3 calibration. However, in validation, the RMSE and MSE values were
0.152 0.15 5.95 17.1 42.5 400.6 lowest for the proposed model. In our ANN model, RMSE and MBE
0.094 0.047 1 15.7 34 67.7 values were 44.17% and 15% less than the ANFIS model, respectively.
0.094 0.094 1 15.7 34 90.5
0.094 0.047 1 16.1 37 98.8
The RMSE value for the validation data set was almost twice the RMSE
0.094 0.094 1 16.1 37 131.5 value of the calibration data set for ANFIS. However, the RMSE values
0.094 0.047 1 16.5 39.5 147.8 for the validation and calibration data sets were nearly the same for
0.094 0.094 1 16.5 39.5 191.6 the proposed model. The RMSE value of our model was 50% lower in
0.094 0.047 1 16.8 41.5 196.8
calibration and 62% lower in validation, than the ANN model
0.094 0.094 1 16.8 41.5 253.6
0.094 0.047 1 17.1 42.5 228.8 described by Padmini et al. (2007). It is possible to say that the
0.094 0.094 1 17.1 42.5 295.6 performance of the proposed ANN model was better than that of the
0.152 0.075 1 15.7 34 91.2 other three methods for the same data set.
0.152 0.15 1 15.7 34 124.4 In addition to the above mentioned performance evaluation
0.152 0.075 1 16.1 37 135.2
0.152 0.15 1 16.1 37 182.4
criteria, the percentage relative error (RE) in ultimate bearing
0.152 0.075 1 16.5 39.5 201.2 capacity was also considered for the validation data set in order to
0.152 0.15 1 16.5 39.5 264.5 evaluate the performance of the proposed model and that of the
0.152 0.075 1 16.8 41.5 276.3 others. The relative error criterion is expressed in Eq. (8).
0.152 0.15 1 16.8 41.5 361.5
0.152 0.075 1 17.1 42.5 325.3 qume −quann
0.152 0.15 1 17.1 42.5 423.6 RE = x100 ð8Þ
qu

The percentage mean absolute relative error (MARE) computed


using RE is defined by Eq. (9) as
The training and testing data sets selected to compare the results
were the same as those used in (Padmini et al., 2007). Therefore, 78 of 1 n
the data sets were used for training and 19 for validation. The MARE = ∑ jREj ð9Þ
n i=1
proposed technique involved training the MLP model to determine
the ultimate bearing capacity of shallow foundations when input where n is the number of data in the validation set.
values were entered. The LM algorithm was used to train the neural The RE and MARE values for the proposed ANN model and those of
model. Training the MLP by the LM algorithm involved different input the other methods taken from the literature are presented in Table 4.
sets and corresponding measured values (qume). Differences between The MARE value of the proposed ANN model was lower than the other
the target output (qume) and the actual output of the MLP (quann) models which were used for comparison.
were used to adjust the weights of the neural model through the LM
learning algorithm. The adaptation was carried out after the 4. Ant colony and parallel ant colony optimization algorithms
presentation of each set until the calculation accuracy of the network
was deemed to be satisfactory, according to the root-mean-square Real ants are capable of finding the shortest path from their nest to
(RMS) error between qume and quann for the whole training set, or the the food source, back or around an object. They also have the ability to
maximum allowable number of iterations reached. adapt to changes in their environment. Another interesting point is
Ten hidden layers consisting of neurons were selected in the MLP that ants are almost blind. Studies on ants have shown that their
model. Tangent hyperbolic functions were used in the hidden layer ability to find the shortest path is the result of chemical communi-
and the output layer. The input and output data tuples were scaled cation among them. They use a chemical substance called pheromone
between −1.0 and +1.0 before training. The learning coefficient and to communicate with each other. This type of indirect interaction
momentum coefficient of the LM algorithm was 0.1, and the total through modification of the environment, which is called stimergy, is
number of epochs was 1000. the main principle of ACO algorithms.
The criteria used in this work for the evaluation of the model are While walking, ants deposit a certain amount of pheromone on
given in Table 2. In the table, 
qume and  quann are the mean values of their path and each ant probabilistically chooses a direction to follow.
the measured and computed ultimate bearing capacity respectively, The degree of probability of any one path being the chosen direction
and n is the number of samples. depends on the amount of pheromone deposited on that direction. If
The performance of the proposed neural model was also compared the pheromone amount of all directions is equal, then they have the
with the results of the basic Adaptive Neuro Fuzzy Inference System same probability of being preferred by ants. Since it is assumed that
(ANFIS), ANN and Fuzzy Inference system (FIS), which is taken from the speed of all ants is the same and, therefore, deposit the same
Padmini et al. (2007). The performance statistics of all models are amount of pheromone on their paths, shorter paths will receive more
presented in Table 3. The correlation coefficients (R) between pheromone per time unit. Consequently, large numbers of ants will
rapidly choose the shorter paths. This positive feedback strategy is
Table 2
also known as auto-catalytic process. Furthermore, the quantity of
The performance evaluation criteria. pheromone on each path decreases over time because of evaporation.
Therefore, longer paths lose their pheromone intensity and become
Evaluation criteria Definition
  
less attractive as time passes. This is known as pheromone-based
Coefficient of correlation (R)  
∑ni = qume − qume quann − quann negative feedback strategy.
R = rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1
  2 The ant colony optimisation algorithm is an artificial version of the
 
∑ni = 1
qume − qume ∑ni = 1
quann − quann
natural optimization process carried out by real ant colonies described
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2 above. The first ACO algorithm was proposed by Dorigo et al. (1991)
∑ni = 1 qume −quann
Root-mean-square error (RMSE) RMSE =
n
and was called ant system (AS) (Dorigo et al., 1991, 1996). Real ants
  communicate with each other by leaving a pheromone substance on
Mean bias error (MBE) MBE = 1n ∑ni = 1 quann −qume
the path they choose and this chemical substance leads other ants.
A. Kalinli et al. / Engineering Geology 117 (2011) 29–38 33

Table 3
Performance statistics of all models.

Performance index Calibration Validation

ANN ANFISa ANNa FISa ANN ANFISa ANNa FISa

R 0.9989 0.9986 0.9950 0.9901 0.9990 0.9968 0.9920 0.9899


RMSE (kPa) 26.4 26.4 52.9 71.1 29.2 52.3 77.2 98.0
MBE (kPa) 0 0 −1.78 0 9.78 11.50 −12.04 13.93
a
Padmini et al. (2007).

Thus, stimergy is provided and swarm intelligence emerges in the The PACO algorithm is a hybrid algorithm model which aims to
colony behavior. The main features of the algorithm are distributed avoid the premature convergence behavior of ant algorithms, and to
computation, positive feedback and constructive greedy search. Since benefit from advantages of parallel structure. The proposed PACO
1991, several studies have been carried out on new models of the ACO algorithm is based on the data structure of the TACO and the crossover
algorithm and their application to difficult optimization problems. operator of GAs. The aim of the PACO algorithm is to combine the
Some of these algorithms include the AS with elitist strategy (ASelit), convergence capability of the ant colony metaphor and the global
the rank based version of AS (ASrank), MAX-MIN AS and the ant search capability of the GA. In the TACO algorithm, a solution is a
colony system (ACS) (Gambardella and Dorigo, 1996; Stützle and vector of design parameters, which are coded as a binary bit string.
Hoos, 1997; Bullnheimer et al., 1999). In most application areas, these Therefore, artificial ants search for the value of each bit in the string.
algorithms are mainly used for optimization in discrete space The concept of the TACO algorithm is shown in Fig. 3.
(Gambardella and Dorigo, 1996; Stützle and Hoos, 1997; Caro and When deciding on the value of a bit, ants only use the pheromone
Dorigo, 1998; Stützle and Dorigo, 1999; Gambardella et al., 1999). In information. Once an ant completes the decision process for the
addition, different kinds of ant algorithms such as continuous ACO values of all bits in the string, it means that it has produced a solution
(CACO), API, continuous interacting ant colony (CIAC) and touring to the problem. This solution is evaluated in the problem, and a
ACO (TACO) have been introduced for optimization in the continuous numeric value showing its quality is assigned to the solution using a
field (Bilchev and Parmee, 1995; Monmarché et al., 2000; Dreo and function, often called the fitness function. With respect to this value,
Siarry, 2004). an artificial pheromone amount is attached to the links, forming the
It's known that there is premature convergence (stagnation) artificial way, between the chosen bits. An ant on the bth bit position
problem in the nature of ant algorithms (Dorigo et al., 1996). chooses the value of 0 or 1 for the bit on the (b + 1)th position
Therefore, as the problem size grows, the ability of the algorithm to depending on the probability defined by the following equation:
discover the optimum solution gets weaker. However when the
hiα
problem size and the number of parameters increase, parallel τij
implementation of the algorithm can give more successful results pij ðt Þ = ð10Þ
2 h iα
(Stützle, 1998; Bullnheimer et al., 1998). Some parallel implementa- ∑ τij
tions of ant algorithms exist in the literature (Bolondi and Bondanza, j=1

1993; Krüger et al., 1998; Michel and Middendorf, 1998; Stützle,


1998; Talbi et al., 1999; Middendorf et al., 2000; Delisle et al., 2001). where pij(t) is the probability associated with the link between bit i
One of the newest approaches is the parallel ACO (PACO) algorithm and j, τij(t) is the artificial pheromone of the link and α is the weight
which was proposed by Kalinli and Sarikoc (2009). parameter. Artificial pheromone is computed by the following
formula:
8
Table 4 >
Comparison of error in model predicted bearing capacity values for validation data set.
< Q if the ant k passes the link ði;jÞ
k
Δτij ðt; t + 1Þ = Fk ð11Þ
a
VD b
MUBC RE in the predicted ultimate bearing capacity
>
:
0 otherwise
ANN ANFISc ANNc FISc

1 137 −0.027 13.043 −17.99 26.56 where Δτkij is the pheromone quantity attached to the link (i,j) by the
2 322 −0.005 4.287 29.62 53.33 artificial ant k. This quantity shows the influence of positive feedback
3 2033 0.002 0.448 −0.698 −6.45
strategy. Q is a positive constant and Fk is the objective function value
4 464 0.002 −15.529 −2.385 −41.60
5 214 −0.005 11.342 −3.324 8.82
calculated using the solution found by the ant k.
6 681 0.003 −13.633 14.871 4.36 After M ants complete the search process and produce their paths,
7 630 0.007 −12.817 18.707 14.12 the pheromone amount to be attached to the sub-path (0→1)
8 1140 0.001 6.384 6.590 3.77 between the time t and (t + 1) is computed by
9 461 0.007 16.854 24.511 40.54
10 1540 0.000 −7.234 −5.59 −9.35
M
11 1760 0.002 −4.635 0.395 −7.30 k
Δτij ðt; t + 1Þ = ∑ Δτij ðt; t + 1Þ ð12Þ
12 244.6 0.515 0.762 4.58 5.82 k=1
13 143.3 −0.818 2.162 24.34 −5.90
14 135.2 5.052 2.934 4.71 −32.1
15 264.5 32.232 −2.943 24.87 −31.3
16 131.5 −0.047 2.216 0.675 −19.02
17 253.6 37.053 −1.018 10.84 −8.44
18 180.5 −0.833 0.311 −49.16 −33.76
19 91.5 1.188 2.705 −9.12 −17.04
MARE 4.09 6.38 16.64 20.95
a
Validation data.
b
Measured ultimate bearing capacity (kPa).
c
Padmini et al., 2007. Fig. 3. An artificial path (solution) found by an ant.
34 A. Kalinli et al. / Engineering Geology 117 (2011) 29–38

The pheromone amount on the sub-path (i,j) at the time (t + 1) is


calculated by using the following equation:

τij ðt + 1Þ = ρ τij ðt Þ + Δτij ðt; t + 1Þ ð13Þ Fig. 5. Representation of the coefficients of improved formula.

where ρ is a coefficient called the evaporation parameter.


In the PACO algorithm, each solution is represented by a binary
vector of design parameters, and artificial ants search for the value of
each bit in the string as in the algorithm TACO. For this reason, the
proposed algorithm can search in a sampled finite subset of
continuous space.
The flowchart of the PACO algorithm is given in Fig. 4. In the model,
different independent ant colonies are executed in parallel. Each
colony has a copy of the same search space with the same initial
pheromone quantities. However, it is possible to use different control
parameters values for each colony. As the algorithm runs, the
Fig. 6. System identification scheme using PACO algorithm.
pheromone quantities of each copy may be different. A colony does
not change the pheromone quantities of another colony. However,
they have the ability to exchange information implicitly. The
information exchange process between the ant colonies is based on
out at predetermined moments. Therefore, the parallelism used in this
the crossover operation.
work is synchronous. Details of the PACO algorithm can be found in
Execution of the colonies is stopped after a given number of
Kalinli and Sarikoc (2009).
iterations. There is no certain rule to determine this number, it may be
defined experimentally. The number of iterations is normally chosen
5. An improved Meyerhof formula obtained by PACO for the
to be sufficiently large to allow the search to complete the local
computation of ultimate bearing capacity
searching. When all the ants in a colony complete their paths, the
quality of the path produced by each ant is evaluated, and then the
In order to derive an improved Meyerhof formula, four additional
best is reserved as the local best of the colony. In every fixed number
coefficients are used in the basic Meyerhof formula given in Eq. (2).
of iterations, the local best solutions of each colony are added to the
Therefore, the improved formula, in which the set of coefficient values
solution population. Later, this population is altered by a crossover
will be optimized, can be written as
procedure to produce a new population. This new population is
formed by implementing the crossover operation among the solutions x x
belonging to the previous solution population. After the crossover, the qult = x1 γD 2 Nq Fqs Fqd + x3 γB 4 Nγ Fγs Fγd ð14Þ
best part of the population survives and the solutions of this part are
used to update the pheromone quantities of the best paths in each The PACO algorithm used in this study is the one described in
colony. Thus, one epoch of the algorithm is completed. In successive Kalinli and Sarikoc (2009). Initial solutions are produced by a random
epochs, the search continues depending on the pheromone quantities number generator. Using these solutions, which represent the
updated in the previous epoch. This process is repeated until the possible coefficient set of the improved formula, the PACO algorithm
predefined number of epochs is completed. The number of epochs searches the best coefficient set by means of certain strategies. A
may be changed by the problem, so the value of this parameter is solution is a string comprising of n elements, where n is the number of
experimentally defined. the coefficients to be determined (Fig. 5). In order to avoid the
Different independent ant colonies are sequentially executed in a complexity of searching a large solution space, the coefficient values
single processor and for this reason, implementation of the algorithm of the proposed formula are assumed to be in the range of [−3.0, 3.0].
is virtually parallel. Communication between the colonies is carried The use of the PACO algorithm to obtain a coefficient set is
illustrated in Fig. 6, where, qu_m[k] and qu_im[k] are the outputs of the

Table 5
The additional data set used for developing the improved Meyerhof formula by using
the ACO algorithm.

Source B (m) D (m) L/B γ (kN/m3) Φ (°) qu (kPa)

Golder (1941) 0.08 0 1 17.2 42.8 133


0.15 0 1 17.2 42.8 246
Eastwood (1951) 0.05 0 1 17.2 42.8 109
0.08 0 1 17.1 42.8 130
0.10 0 1 17.1 42.8 152
0.15 0 1 17.1 42.8 214
0.20 0 1 17.1 42.8 266
0.25 0 1 17.1 42.8 333
0.30 0 1 17.1 42.8 404
Subrahmanyam (1967) 0.03 0 1 15.89 42 52
0.04 0 1 15.89 42 92
0.05 0 1 15.89 42 95
Cerato (2005) 0.06 0 1 13.2 32 14
0.06 0 1 14.8 42 72
0.06 0 1 15.4 42 106
Fig. 4. Flowchart of the PACO algorithm.
A. Kalinli et al. / Engineering Geology 117 (2011) 29–38 35

Table 6
The experimental and the theoretical results for ultimate bearing capacity values.

Source qu Meyerhoff Improved Meyerhoff Vesic Hansen


(kPa) (kPa) (kPa) (kPa) (kPa)

Muhs et al. (1969) 270 265.96 269.331 259.5722 219.8095


200 223.90 182.578 181.7531 130.6388
570 438.16 436.97 404.5569 340.4771
860 1054.58 859.958 680.9346 502.9433
1760 1795.12 1719.27 1372.563 1127.926
Weiß (1970) 154 218.63 177.148 113.5957 81.64923
165 218.63 177.148 113.5957 81.64923
203 186.58 151.178 151.461 108.8656
195 186.58 151.178 151.461 108.8656
214 175.89 142.521 164.0827 117.9378
186 178.03 144.445 176.4422 126.8214
681 473.85 480.209 404.4536 346.0345
542 404.39 409.809 378.9193 323.0877
530 404.39 409.809 378.9193 323.0877
402 381.23 386.343 370.4079 315.4387
413 377.10 380.828 372.5056 316.0412
Muhs and Weiß (1971) 111 218.50 177.038 116.1671 83.15783
132 218.50 177.038 116.1671 83.15783
143 187.16 151.644 154.8894 110.8771
137 228.84 182.113 127.6856 93.67963
109 189.70 150.02 191.7966 141.7227
187 253.99 215.03 254.6397 202.5597
406 481.47 489.822 418.2128 359.0669
446 481.47 489.822 418.2128 359.0669
322 377.88 384.434 378.9479 322.8998
565 586.77 707.135 583.175 515.5613
425 537.64 647.927 543.9326 482.618
782 1019.92 826.383 418.1571 308.2249
797 746.90 605.172 627.2356 462.3374
1940 1925.99 1881.53 1353.181 1114.327
2266 1925.99 1881.53 1353.181 1114.327
2847 2164.67 2476.65 1809.541 1548.547
2033 1929.56 2207.66 1675.396 1453.289
1492 1292.44 1491.94 1179.738 1029.516
123 219.81 178.097 116.8621 83.65536
134 188.28 152.551 155.8161 111.5405
370 484.35 492.753 420.715 361.2152
464 590.28 711.366 586.6641 518.6459
461 313.39 253.921 295.4085 214.7597
1140 902.15 1064.73 861.1616 757.5614
Muhs and Weiß (1973) 710 712.93 632.129 656.1764 515.5376
630 644.54 535.095 565.6161 411.1984
Briaud and Gibbens (1999) 1773.7 647.87 846.834 648.8068 574.8993
1019.4 1111.20 1266.93 915.389 752.7894
1158 995.95 1166.83 848.2038 708.2355
1540 779.69 986.669 731.0316 634.1224
1161.2 1186.92 1424.37 1003.972 835.7602
Gandhi (2003) 58.5 32.09 21.5033 34.45245 28.89468
70.91 51.68 38.043 55.19354 49.27472
82.5 52.97 35.8459 54.32758 45.49397
98.93 83.41 61.474 84.92803 75.40177
121.5 82.88 56.5484 81.37907 67.95868
142.9 128.08 94.4956 124.6163 109.9944
157.5 120.88 83.0158 114.0684 94.96339
184.9 183.96 135.848 171.8074 150.8261
180.5 148.44 102.277 137.0456 113.8859
211 224.16 165.601 204.7193 179.1868
74.7 51.79 37.4748 55.61393 46.67785
91.5 83.58 68.535 89.28149 79.76374
104.8 85.46 62.1148 87.67243 73.46926
127.5 134.85 110.156 137.3318 122.0129
155.8 133.68 97.5313 131.2975 109.7194
185.6 207.02 168.537 201.4485 177.9353
206.8 194.92 142.648 184.0056 153.2876
244.6 297.29 241.336 277.6675 243.9279
235.6 239.33 175.416 221.0512 183.8137
279.6 362.20 293.591 330.8177 289.7595
98.2 83.16 65.1575 89.29039 74.85365
122.3 133.76 122.195 142.8612 127.4919
143.3 137.28 107.49 140.8215 117.8769
176.4 215.93 195.571 219.8641 195.1303
211.2 214.84 168.12 210.9669 176.1102
254.5 331.61 298.1 322.6599 284.6992

(continued on next page)


36 A. Kalinli et al. / Engineering Geology 117 (2011) 29–38

Table 6 (continued)
Source qu Meyerhoff Improved Meyerhoff Vesic Hansen
(kPa) (kPa) (kPa) (kPa) (kPa)

Gandhi (2003) 285.3 313.37 245.12 295.7389 246.1195


342.5 476.36 425.512 444.9039 390.4373
335.3 384.84 300.952 355.3277 295.1785
400.6 580.46 516.786 530.1643 463.8851
67.7 66.21 47.9059 59.32463 51.05018
90.5 106.84 87.6118 110.004 98.23947
98.8 112.30 81.6293 93.8079 79.75149
131.5 177.22 144.763 171.172 150.4211
147.8 180.27 131.524 140.7389 118.2052
191.6 279.17 227.276 253.2589 219.114
196.8 268.89 196.783 197.4377 164.0071
253.6 410.11 332.923 351.2799 299.7198
228.8 334.18 244.934 237.2785 195.9339
295.6 505.74 409.942 419.7587 355.4899
91.2 106.26 83.2552 94.95367 81.6409
124.4 170.92 156.135 175.5246 156.6633
135.2 180.32 141.185 150.1900 127.5902
182.4 283.61 256.877 273.2147 239.9693
201.2 289.54 226.581 225.3832 189.1723
264.5 446.92 401.761 404.3535 349.6742
276.3 432.02 337.927 316.2442 262.5417
361.5 656.72 586.619 560.9877 478.4434
325.3 537.00 419.941 380.096 313.6917
423.6 809.96 721.11 670.4282 567.5506
Golder (1941) 133 172.13 130.777 74.2276 54.5239
246 322.74 250.677 139.1767 102.2323
Eastwood (1951) 109 106.95 79.9308 46.1225 33.8793
130 171.13 130.017 73.7960 54.2069
152 213.91 163.799 92.2450 67.7586
214 320.86 249.22 138.3675 101.6379
266 427.82 335.665 184.4900 135.5172
333 534.77 422.881 230.6125 169.3966
404 641.72 510.715 276.7351 203.2759
Subrahmanyam (1967) 52 49.96 36.6719 22.2441 16.2967
92 66.61 49.3921 29.6588 21.7290
95 83.26 62.2256 37.0735 27.1613
Cerato (2005) 14 11.56 8.69408 7.1790 4.9388
72 93.06 69.995 41.4364 30.3577
106 96.83 72.8327 43.1163 31.5884
Total absolute error (TAE) 10765.16 8905.91 12188.50 16388.35
R 0.938919 0.959224 0.944431 0.940836

Meyerhof formula and the improved Meyerhof formula at time k, In addition to the data set given in Table 1, the data set given in
respectively. U in this figure denotes the sequence of input variables Table 5 was also used in the simulations. Table 5 presents the results
given as of the square footing tests found in the literature (Golder, 1941;
Eastwood, 1951; Subrahmanyam, 1967; Cerato, 2005). The friction
T
U = ½B; D; L=B;γ;Φ ð15Þ angle was determined by both direct shear and triaxial testing, and
the footing load tests were performed on a number of different sand
Determination of the coefficients for the improved formula can be types and varying densities. Different methods have been used by
considered as an optimization problem of the cost function J(x) stated respective authors to determine ultimate bearing capacities. As seen
as the following: in these tables, the total data number is M = 112. First of all in the
optimization process, a sequence of input data u(k), (k = 1,2,…,112) is
min J ðxÞ ð16Þ fed to both the original formula and the improved formula, which is
x∈X
designed with coefficients obtained by the PACO algorithm. After that,
where, X = [x1, x2, x3,..., xn]T is the coefficient vector of the improved the total absolute error (TAE) value between the original and
formula. The aim is to minimize the cost function J(x) by adjusting x. improved formula responses is computed using Eq. (17). Then,
In this case, the coefficients of the improved formula are successively according to the TAE values computed for the candidate solutions
adjusted by the PACO algorithm until the error between the outputs of produced by each ant, the pheromone amount on the artificial path is
the original and the improved formula is minimized. In this study, the updated. In this study, the parameters of the PACO algorithm were the
cost function is considered as the total absolute error (TAE) function number of ant colonies running in parallel, which was 4, and the
defined by number of ants was 30. Each colony at any epoch was run for 20
iterations. This process was stopped when the error was not changed
M
any more.
JðxÞ = TAE = ∑ jqu−m ðkÞ−qu−im ðkÞj ð17Þ
k=1 The coefficients of the improved formula obtained by using the
PACO algorithm were x1 = 1.8102, x2 = 1.3199, x3 = 0.4151 and
Where, M is the number of samples used for the calculation of the x4 = 1.0351. The ultimate bearing capacity values of the experimental,
cost function. original and improved formulas are shown in Table 6. It can be seen
A. Kalinli et al. / Engineering Geology 117 (2011) 29–38 37

that the calculated TAE values are 10765.16 and 8905.91 when using Dreo, J., Siarry, P., 2004. Continuous ant colony algorithm based on dense heterarchy.
Future Generation Computer Systems 20 (5), 841–856.
the original and improved Meyerhof formula, respectively. In terms of Du, K.L., Lai, A.K.Y., Cheng, K.K.M., Swamy, M.N.S., 2002. Neural methods for antenna
the TAE values, the proposed formula improved the performance of array signal processing: a review”. Signal Processing 82, 547–561.
the original Meyerhof formula by about 17.27%. The RMSE value of the Eastwood, W., 1951. A comparison of the bearing power of footings on dry and
inundated sand. Structural Engineering 29 (1), 332.
improved Meyerhof formula is 19.91% less than the original Meyerhof Ellis, G.W., Yao, C., Zhao, R., Penumadu, D., 1995. Stress–strain modelling of sands using
formula. The RMSE value of the obtained formula is also 35.30% and ANN. Journal of Geotechnical Engineering, ASCE 121 (5), 429–435.
48.41% less than the Vesic and Hansen formulas respectively. As can Gambardella, L.M., Dorigo, M., 1996. Solving symmetric and asymmetric TSPs by ant
colonies. Proceedings of IEEE International Conference on Evolutionary Computa-
be seen, a more accurate formula was obtained by using the PACO tion, IEEE-EC 96, May 20–22. IEEE Press, Nagoya, Japan, pp. 622–627.
algorithm, as compared to the original formula. Gambardella, L.M., Taillard, E., Agazzi, G., 1999. MACS-VRPTW: A Multiple Ant Colony
System for Vehicle Routing Problems with Time Windows, Technical Report, IDSIA-06,
Switzerland.
6. Conclusions Gandhi GN., 2003. Study of bearing capacity factors developed from laboratory
experiments on shallow footings on cohesionless soils. PhD thesis, Shri G.S.
In this study, two different approaches were proposed to Institute of Tech and Science, Indore (MP).
Gao, W., 2005. Method for searching critical slip surface of soil slope base on ant colony
determine the ultimate bearing capacity of shallow foundations on algorithm. Journal of Hydraulic Engineering 36 (9), 1100–1104 (in Chinese).
granular soil. Firstly, an ANN model was proposed to predict the Goh, A.T.C., 1994. Nonlinear modelling in geotechnical engineering using neural
ultimate bearing capacity. The performance of the proposed neural networks. Australian Civil Engineering Transactions CE36 (4), 293–297.
Golder, H.Q., 1941. The ultimate bearing pressure of rectangular footings. J. of the
model was compared with the results of ANFIS, ANN and FIS, which Institution of Civil Engineers 17, 161–174.
were taken in the literature. It was clearly seen that the performance Hansen, J.B., 1968. A revised extended formula for bearing capacity. Danish
of our proposed ANN model was better than that of the other Geotechnical Institute Bulletin, No. 28.
Jacek, T., Ivo, H., 1999. A ‘class A’ prediction of the bearing capacity of plane strain
prediction methods. footings on sand. Soils and Foundations 39 (5), 47–60.
Secondly, using the PACO algorithm, an improved Meyerhof formula Kahatadeniye, K.S., Nanakorn, P., Neaupane, K.M., 2009. Determination of the critical
was proposed for the computation of the ultimate bearing capacity. The failure surface for slope stability analysis using ant colony optimization.
Engineering Geology 108, 133–141.
results achieved from the proposed formula were compared with those
Kalinli, A., Sarikoc, F., 2009. A parallel ant colony optimization algorithm based on
obtained from the Meyerhof, Hansen and Vesic formulas. Simulation crossover operation. In: Michalewicz, Z., Siarry, P. (Eds.), Advances in Metaheur-
results showed that the responses produced by the improved Meyerhof istics for Hard Optimization. Springer Verlag, pp. 87–110.
formula were in good agreement with the experimental results. It is Krüger, F., Merkle, D., Middendorf, M., 1998. Studies on a parallel ant system for the BSP
model, unpublished manuscript, (Downloadable from http://citeseer.ist.psu.edu/
expected that the proposed Meyerhof formula obtained by using the 239263.html.).
PACO algorithm will find wide application in the calculation of the Lee, I.M., Lee, J.H., 1996. Prediction of pile bearing capacity using artificial neural
ultimate bearing capacity of foundations. networks. Computers and Geotechnics 18 (3), 189–200.
Liang LI, Shichun CHI, Yungming CHENG, Gao LIN, 2008. Improved genetic algorithm
and its application to determination of critical slip surface with arbitrary shape in
Acknowledgement soil slope. Higher Education Press and Springer-Verlag. materials and the strength
of cutting edges, Math. Mech., 1(1), 15–20.
Meyerhof, G.G., 1963. Some recent research on the bearing capacity of foundations.
This work was supported by Erciyes University's, Research Fund Canadian Geotechnical Journal 1 (1), 16–26.
(project number: FBD-10-3035). Michel, R., Middendorf, M., 1998. An island model based ant system with lookahead
for the shortest supersequence problem. In: Eiben, A.E., Back, T., Schoenauer, H.,
Schwefel, P. (Eds.), Parallel Problem Solving from the Nature-PPSN V. Lecture Notes
References in Computer Science, 1498. Springer-Verlag, Heidelberg, Germany, pp. 692–701.
Middendorf, M., Reischle, F., Schmeck, H., 2000. Information exchange in multicolony
Al-Shamsi, K.S.M., 1993. Bearing Capacity of Shallow Foundations. Internal Report. algorithms. In: Rolim, J., Chiola, G., Conte, G., Mansini, L.-V., Ibarra, O.-H., Nakano,
Dept. of Civil Engineering, Sultan Qaboos University, Oman, p. 60. H. (Eds.), Parallel and distributed processing: 15 IPDPSP workshops Mexico.
Bilchev, G., Parmee, I.C., 1995. The ant colony metaphor for searching continuous design Lecture Notes in Computer Science, 1800. Springer-Verlag, Heidelberg, Germany,
spaces. Lecture Notes in Computer Science. LNCS, 993. Springer-Verlag, pp. 25–39. pp. 645–652. May.
Bolondi, M. and Bondanza, M., 1993. Parallelizzazione di un algoritmo per la risoluzione Monmarché, N., Venturini, G., Slimane, M., 2000. On how Pachycondyla apicalis ants suggest
del problema del commesso viaggiatore. Master's thesis, Dipartimento di is new search algorithm. Future Generation Systems Computer 16 (8), 937–946.
Elettronica e Informazione, Politecnico di Milano, Italy. Muhs, H., Weiß, K., 1971. Untersuchung von Grenztragfähigkeit und Setzungsverhalten
Briaud, J.L., Gibbens, R., 1999. Behaviour of five large spread footings in sand. Journal of flachgegründeter Einzelfundamente im ungleichförmigennichtbindigen Boden.
Geotechnical and Geoenvironmental Engineering 125 (9), 787–796. Deutsche Forschungsgesellschaft für Bodenmechanik (DEGEBO), Berlin. HEFT 69.
Bullnheimer, B., Kotsis, Strauss, C., 1998. Parallelization strategies for the ant system. In: Muhs, H., Weiß, K., 1973. Inclined load tests on shallow strip footings. Proceedings of
De Leone, R., Murli, A., Pardalos, P., Toraldo, G. (Eds.), High Performance Algorithms the 8th international conference on soil mechanism and foundation engineering,
and Software in Nonlinear Optimization, Kluwer Series of Applied Optimization, 24. Vol. II, pp. 173–179.
Kluwer Academic Publishers, Dordrecht, The Netherlands, pp. 87–100. Muhs, H., Elmiger, R., Weiß, K., 1969. Sohlreibung und Grenztragfähigkeit unter lotrecht
Bullnheimer, B., Hartl, R.F., Strauss, C., 1999. A new rank based version of the ant system, und schräg belasteten Einzelfundamenten. Deutsche Forschungsgesellschaft für
a computational study. Central European Journal for Operations Research and Bodenmechanik (DEGEBO), Berlin.HEFT 62.
Economics 7 (1), 25–38. Padmini, D., Ilamparuthi, K., Sudheer, K.P., 2007. Ultimate bearing capacity prediction of
Caro, G.Di, Dorigo, M., 1998. Mobile agents for adaptive routing. Proceedings of 31st shallow foundations on cohesionless soils using neurofuzzy models. Computers
Hawaii Conference on Systems Sciences (HICSS-31). and Geotechnics 35, 33–46.
Cerato, A.B., 2005. Scale Effects Of Shallow Foundation Bearing Capacity On Granular Perloff, William H., Baron, William, 1976. Soil Mechanics: Principles and Applications.
Material, Ph.D. dissertation, University of MASSACHUSETTS AMHERST, 38. Ronald Press Company, New York.
Changfu, C., Xiaonan, G., Yisun, W., 2003. Adaptive colony algorithm and its application to Picton, Phil, 1994. Introduction to Neural Networks. Macmillan Press, UK.
the slope engineering. Journal of Zhejiang University (Engineering science) 37 (5), Prandtl, L., 1921. Über die Eindringungsfestigkeit (Härte) plastischer Baustoffe und die
566–569 (in Chinese). Festigkeit von Schneiden (On the penetrating strengths (hardness) of plastic
Corne, D., Dorigo, M., Glover, F. (Eds.), 1999. New Ideas in Optimization. McGraw-Hill, UK. construction materials and the strength of cutting edges). Zeitschrift für
Das, Braja, 1999. Principles of Foundation Engineering. International Thomson Comp., p. 156. Angewandte Mathematik und Mechanik 1 (1), 15–20.
De Beer, E.E., 1965. The scale effect on the phenomenon of progressive rupture in Rahman, M.S., Wang, J., Deng, W., Carter, J.P., 2001. A neural network model for the
cohesionless soil. Proceedings of the 6th international conference on soil uplift capacity of suction caissons. Computers and Geotechnics 28 (4), 269–287.
mechanism and foundation engineering, 2, pp. 13–17. Reeves, C.R. (Ed.), 1995. Modern Heuristic Techniques for Combinatorial Optimization.
Delisle, P., Krajecki, M., Gravel, M., Gagné, C., 2001. Parallel implementation of an ant McGraw-Hill, UK.
colony optimization metaheuristic with openmp, International Conference on Reissner, H., 1924. Zum Erddruckproblem (Concerning the earth-pressure problem).
Parallel Architectures and Compilation Techniques. Proceedings of the 3rd Proceedings 1st International Congress of Applied Mechanics, Delft, 295–311.
European Workshop on OpenMP (EWOMP'01), Barcelona, Spain. Shahin, M.A., Jaksa, M.B., Maier, H.R., 2000. Predicting the settlement of shallow
Dorigo, M., Maniezzo, V., Colorni, A., 1991. Positive feedback as a search strategy, foundations on cohesionless soils using back-propagation neural networks,
Technical Report N. 91-016. Politecnico di Milano. Research Report No. R 167. The University of Adelaide, Adelaide.
Dorigo, M., Maniezzo, V., Colorni, A., 1996. The ant system: Optimization by a colony of Shahin, M. A., Jaksa, M.B., and Maier, H. R., 2009. Recent advances and future challenges
cooperating agents. IEEE Transactions on Systems, Man and Cybernetics—Part B 26 (1), for artificial neural systems in geotechnical engineering applications, Advances in
1–13. Artificial Neural Systems. doi:10.1155/2009/308239.
38 A. Kalinli et al. / Engineering Geology 117 (2011) 29–38

Sivakugan, N., Eckersley, J.D., Li, H., 1998. Settlement predictions using neural networks. Processing, 11 IPPS/SPDP'99 Workshops. Lecture Notes in Computer Science, 1586.
Australian Civil Engineering Transactions CE40, 49–52. Springer-Verlag, London, UK, pp. 239–247.
Steenfelt, J.S., 1977. Scale effect on bearing capacity factor Nγ. Proceedings of the 9th Tatsuoka, F., Okahara, M., Tanaka, T., Tani, K., Morimoto, T., Siddiquee, M.S.A., 1991.
international conference on soil mechanism and foundation engineering, Tokyo, Progressive failure and particle size effect in bearing capacity of footing on sand.
Japan, 1, pp. 749–752. Proceedings of the ASCE Geotechnical Engineering Congress, New York 27, 788–802.
Stützle, T., 1998. Parallelization strategies for ant colony optimization. In: Eiben, A.E., Back, Taylor, R.N., 1995. Geotechnical Centrifuge Technology, 1st Edition. Chapman & Hall,
T., Schoenauer, M., Schwefel, H.P. (Eds.), Parallel Problem Solving from Nature- PPSN London, U.K.
V: 5th International Conference, Amsterdam. Lecture Notes in Computer Science, Teh, C.I., Wong, K.S., Goh, A.T.C., Jaritngam, S., 1997. Prediction of pile capacity using
1498. Springer-Verlag, Heidelberg, Germany, pp. 722–731. September. neural networks. Journal of Computing in Civil Engineering, ASCE 11 (2), 129–138.
Stützle, T., Dorigo, M., 1999. ACO algorithms for quadratic assignment problem. In: Corne, Terzaghi, K., 1943. Theoretical Soil Mechanics. John Wiley & Sons, New York.
D., Dorigo, M., Glover, F. (Eds.), New Ideas in Optimization. McGraw-Hill, pp. 33–50. Transportation Research Board (TRB), 1999. Use of artificial neural networks in
Stützle, T., Hoos, H.H., 1997. The MAX–MIN ant system and local search for the traveling geomechanical and pavement systems. Transportation Research Circular (E-C012),
salesman problem. In: Baeck, T., Michalewicz, Z., Yao, X. (Eds.), Proceedings of the IEEE 1–18 December.
International Conference on Evolutionary Computation (ICEC'97), pp. 309–314. Vesic, A.S., 1973. Analysis of ultimate loads of shallow foundations. Journal of The Soil
Subrahmanyam, G., 1967. The effect of roughness of footings on bearing capacity. Journal Mechanics and Foundations Division 99 (1), 45–73.
of the International Society of Soil Mechanics and Foundation Engineering 6, 33–45. Weiß, K., 1970. Der Einfluß der Fundamentform auf die Grenztragfähigkeit flachge-
Talbi, E.-G., Roux, O., Fonlupt, C., Robillard, D., 1999. Parallel ant colonies for combi- gründeter Fundamente. Deutsche Forschungsgesellschaft für Bodenmechanik
natorial optimization problems. In: Rolim, J., et al. (Ed.), Parallel and Distributed (DEGEBO), Berlin. HEFT 65.

You might also like