Professional Documents
Culture Documents
Dry Unit Weight of Compacted Soils Prediction Using GMDH-type Neural Network
Dry Unit Weight of Compacted Soils Prediction Using GMDH-type Neural Network
net/publication/319159054
CITATIONS READS
19 951
4 authors:
Some of the authors of this publication are also working on these related projects:
رﻓﺘﺎر ﺳﺘﻮﻧﻬﺎی داﻧﻬﺎی ﻣﺤﺼﻮرﺷﺪه ﺑﺎ ژﺋﻮﺳﻨﺘﺘﯿﮏ در ﺧﺎک ﻣﺎﺳﻪ اﻟﯿﺪار ﺗﺤﺖ آزﻣﺎﯾﺶ ﺑﺮش ﻣﺴﺘﻘﯿﻢView project
All content following this page was uploaded by Alireza Ardakani on 21 August 2017.
To cite this article: Alireza Ardakani & Afshin Kordnaeij (2017): Soil compaction parameters
prediction using GMDH-type neural network and genetic algorithm, European Journal of
Environmental and Civil Engineering
1. Introduction
Soil compaction occurs when soil particles are pressed together, reducing the pore space between
them. This increases the weight of solids per unit volume of soil. Generally, compaction increases the
strength characteristics of the soil, reduces its compressibility and permeability.
The maximum dry density (γd,max) and optimum moisture content (ωopt) of soils are determined using
standard Proctor test (Proctor, 1933). These compaction parameters are very important for geotechni-
cal engineering and earth structures such as earth dams, bridge abutments, highway embankments
and the fills behind retaining walls. Soil is compacted in the project site according to the compaction
parameters determined in the laboratory. For a wide variety of earth structures, the minimum accept-
able dry density (γd) is usually specified at 95 percent of the γd,max for the standard Proctor compaction
(ASTM D698, 2003).
Soil type is an important factor that affects the soil compaction (Das, 2010). The compaction pro-
cess has significant differences for cohesive soil vs. cohesionless soils. The shapes and the positions
of the compaction curves change as the texture of the soils varies from coarse to fine (Terzaghi, Peck,
& Mesri, 1996). The major difference is that the cohesive soils are typically very moisture-dependent
and cohesionless soils are not (Rout, 2009). As standard Proctor test is time consuming and laborious,
several researchers have tried to find prediction models which can predict approximately the values of
𝜔opt (%):
𝜔opt = 0.24 LL + 0.63 PL − 3.13 (Iraqi soils) (6) AI-Khafaji (1993)
𝜔opt = 0.14LL + 0.54PL (US soils) (7) AI-Khafaji (1993)
𝜔opt = 0.92PL (8) Sridharan and Nagaraj (2005)
𝜔opt = 0.8442 PL + 0.1076 (9) Gunayin (2009)
𝜔opt = 0.3802 LL + 2.4513 (10) Gunayin (2009)
𝜔opt = 0.323 LL + 0.157 PL (11) Gunayin (2009)
γd,max and ωopt from more simply determined index properties of natural soil. Table 1 presents several
types of empirical correlations in this field.
Almost all of the empirical correlations for the compaction parameters prediction have been estab-
lished based on regression analysis. Many factors may affect the compaction parameters (γd,max and
ωopt). Hence any statistical method using for the estimation of the γd,max and ωopt may contain some
deviation. Computational intelligence (CI) techniques can be considered as efficient methods. They
determine the structure of a prediction model by automatically learning from data. CI has different
well-known branches such as support vector machines (SVM), artificial neural network (ANN), fuzzy
inference system (FIS), etc. These methods have been used to model complex relationships between
input and output data-sets in geotechnical engineering (Edincliler, Cabalar, & Cevik, 2013; Kalantary &
Kordnaeij, 2012; Kang, Li, Wang, & Li, 2016; Žlender, Jelušič, & Boumezerane, 2012). Despite the good
performance of these CI methods, they are considered black-box models and are not capable of gener-
ating practical prediction equations. This is a fundamental disadvantage that limits their practicability
(Mohammadzadeh, Bolouri Bazaz, & Alavi, 2014). In order to overcome this disadvantage, Group Method
of Data Handling (GMDH) type NN has been used in this article.
The GMDH-type NN is a powerful identification technique and can be used to model complex sys-
tems, where unknown relationships exist between variables, without having specific knowledge of
processes as other NNs. GMDH is aimed at identifying the functional structure of a model hidden in
the empirical data (Ivakhnenko, 1971).
In recent years, genetic algorithms have been used in a feed forward GMDH-type NN for each
neuron searching its optimal set of connection with the preceding layer. Over the last few years, the
GMDH-type NN optimised by genetic algorithms have been applied to many geotechnical engineer-
ing problems and have demonstrated some degree of success (Ardalan, Eslami, & Nariman-Zadeh,
2009; Kalantary, Ardalan, & Nariman-Zadeh, 2009; Mola-Abasi, Eslami, & Tabatabaeishorijeh, 2013).
Therefore, this approach can be used in empirical correlation of the compaction parameters as a step
forward in comparison with statistical approaches and investigate the effect of input parameters on
the model output.
Present study aims to develop a GMDH-type NN for the prediction of γd,max and ωopt based on various
soil index parameters such as liquid limit (LL), plastic limit (PL) and fine-grained content (FC) as well as
sand content (SC), without carrying out Proctor tests.
The rest of this paper is organised as: Review of the GMDH-type NN model and application of genetic
algorithms in the topology design of GMDH-type NN, database compilation, evaluation of the compac-
tion parameters using GMDH model, carrying out the sensitivity analysis of the obtained model and
finally, the conclusion of the paper.
EUROPEAN JOURNAL OF ENVIRONMENTAL AND CIVIL ENGINEERING 3
The problem is now to determine a GMDH-type of artificial neural network so that the square of the
differences between the observed and predicted output is minimised as follow:
∑
M
( )
[̂f xi1 , xi2 , xi3 , … , xin − yi ]2 → min (14)
i=1
The general connection between input and output variables can be expressed by a complicated discrete
form of the Volterra functional series, known as the Kolmogorov–Gabor polynomial. Hence:
∑
n
∑
n
∑
n
∑
n
∑
n
∑
n
y = a0 + ai xi + aij xi xj + aijk xi xj xk + … (15)
i=1 i=1 j=1 i=1 j=1 k=1
This full form of mathematical description can be represented by a system of partial quadratic polyno-
mials consisting of only two variables (neurons) in the form of:
( )
y = G xi , xj = a0 + a1 xi + a2 xj + a3 xi xj + a4 xi2 + a5 xj2
̂ (16)
By this means, the partial quadratic description is recursively used in a network of connected neurons
to build the general mathematical relation between inputs and output given in Equation (15). The coef-
ficients ai in Equation (16) are calculated using regression techniques so that the difference between
the observed output, y and the calculated one, ŷ for each pair of xi, xj as input variables is minimised
(Jamali, Nariman-zadeh, Darvizeh, Masoumi, & Hamrang, 2009). Apparently, a tree of polynomials is
constructed using the quadratic form given in Equation (16) whose coefficients are obtained in a least
squares scheme. In this way, the coefficients of each quadratic function Gi are obtained to fit optimally
the output in the whole set of input–output data pairs, that is:
∑M � �
i=1 (yi − Gi xi , xj )2
E= → min (17)
M
In the basic form of the GMDH algorithm, all the possibilities of two independent variables out of the
total n input variables are taken in order to construct the regression polynomial in the form of Equation
4 A. ARDAKANI AND A. KORDNAEIJ
(16) that
( ) best fits the dependent observations (yi , i = 1, 2, … , M) in a least squares sense. Consequently,
n
= n(n−2)
2
neurons will be built up in the first hidden layer of the feed forward network from
2
{( ) }
M) for different
the observations yi , xip , xiq ; (i = 1, 2, 3, … ,{( ) } … , n}. In other words, it
p, q ∈ {1, 2, 3,
is now possible to construct M data triples yi , xip , xiq ; (i = 1, 2, 3, … , M) from observations using
p, q ∈ {1, 2, 3, … , n} in the form of:
⎡ x1p x1q ⋮ y1 ⎤
⎢ x2p x2q ⋮ y2 ⎥
⎢ ⎥
⎢ …… ⋮ … ⎥
⎢ ⎥
⎣ xMp xMq ⋮ yM ⎦
Using the quadratic sub-expression in the form of Equation (16) for each row of M data triples, the
following matrix equation can be readily obtained as:
Aa = Y (18)
a = {a0 , a1 , a2 , a3 , a4 , a5 } (19)
Y = {y1 , y2 , y3 , … , yM }T (20)
It can be readily seen that:
The least squares technique from multiple regression analysis leads to solution of the normal equations,
in the form of:
( )−1
a = AT A AT Y (22)
This determines the vector of the best coefficients of the quadratic Equation (16) for the whole set of M
data triples. Notice that this procedure is repeated for each neuron of the next hidden layer according
to the connectivity topology of the network. However, such a solution directly from normal equations
is rather susceptible to round off errors and, more importantly, to the singularity of these equations.
There are two main concepts involved within GMDH-type of artificial neural network design, namely,
the parametric and the structural identification problems. Nariman-Zadeh, Darvizeh, and Ahmad-Zadeh
(2003) present hybrid genetic algorithm (GA) and singular value decomposition (SVD) method to opti-
mally design such polynomial neural network.
a ab
abbc
b bc
abbcadad
c ad
d
advantage, it was possible to present a simple encoding scheme for the genotype of each individual in
the population. The encoding schemes in general structure GMDH-type NN (GS-GMDH) must, however,
demonstrate the ability of representing different length and size of such neural networks. In Figure 1,
neuron ad in the first hidden layer is connected to the output layer by directly going through the sec-
ond hidden layer. Therefore, it is now very easy to notice that the name of output neuron (network’s
output) includes ad twice as abbcadad. In other words, a virtual neuron named adad has been con-
structed in the second hidden layer and used with abbc in the same layer to make the output neuron
abbcadad as shown in the Figure 1. Such repetition occurs whenever a neuron passes some adjacent
hidden layers and connects to another neuron in the next 2nd, or 3rd, or 4th, or … following hidden
layer. In this encoding scheme, the number of repetition of that neuron depends on the number of
passed hidden layers, ̃ n, and is calculated as 2̃n. It is easy to realise that a chromosome such as abab
bcbc, unlike chromosome abab acbc for example, is not a valid one in GS-GMDH network and has to be
simply re-written as abbc (Jamali et al., 2009).
The genetic operators of crossover and mutation can now be implemented to produce two offspring
from two parents. The natural roulette wheel selection method is used for choosing two parents pro-
ducing two offspring. The incorporation of genetic algorithm into the design of such GMDH type neural
network starts by representing each network as a string of concatenated sub-strings of alphabetical
digits. The fitness, Φ, of each entire string of symbolic digits which represents a GMDH-type NN model
is evaluated in the following form:
Φ = 1∕E (23)
Where E is the mean square of error given by Equation (17), which is minimised through the evolutionary
process by maximising the fitness, Φ. The evolutionary process starts by a random generation of an
initial population of symbolic strings, each as a candidate solution. Then, using the genetic operations
of roulette wheel selection, crossover and mutation the entire population of symbolic strings improve
gradually. In this way, GMDH-type NN models with progressively increasing fitness, Φ, are produced
until no further significant improvement is achievable (Atashkari et al., 2007).
The flowchart of proposed method is given in Appendix A.
60 80
70
50
60
40
50
Frequency
Frequency
30 40
30
20
20
10
10
0 0
20 30 40 50 60 70 80 10 20 30 40 50 60
LL (%) PL (%)
40
40
30
30
Frequency
Frequency
20
20
10
10
0 0
0 20 40 60 80 100 0 15 30 45 60 75
FC (%) SC (%)
80 100
80
60
60
Frequency
Frequency
40
40
20
20
0 0
10 12 14 16 18 20 22 5 10 15 20 25 30 35 40 45
γd,max (KPa) ωopt (%)
𝜔opt = −4.9 + 1.333 Y1 + 0.17 Y2 + 0.0043 Y12 + 0.032 Y22 − 0.049 Y1 .Y2 (25a)
Y1
Y2
Figure 3. Evolved structure of generalised GMDH neural network to predict compaction parameters (Yi).
8 A. ARDAKANI AND A. KORDNAEIJ
a b
22 50
Training Training
Testing Testing
20
40
(KPa)
(%)
18
opt
d,max
Estimated ω
30
16
Estimated γ
14 20
12
10
12 14 16 18 20 22 10 20 30 40 50
Measured γ (KPa) Measured ω (%)
d,max opt
Figure 4. The measured compaction parameters obtained from the Proctor test vs. the GMDH estimated values.
√
√
√1 ∑M
RMSE = √ (h − ti )2 (27)
M 1 i
∑M �
h − ti ��
MAD = 1 � i (28)
M
EUROPEAN JOURNAL OF ENVIRONMENTAL AND CIVIL ENGINEERING 9
Table 5. Statistical results for the empirical formulas and GMDH models.
Equation Equation number R RMSE MAD MAPE
𝛾d,max (KPa):
𝛾d,max = 23.94 − 0.2PL − 0.078LL (Iraqi soils) (1) 0.77 1.28 0.99 5.79
𝛾d,max = 22.27 − 0.19PL − 0.029LL (US soils) (2) 0.74 0.97 0.76 4.47
γd,max = 0.23(93.3 – PL) (3) 0.70 1.19 0.89 5.20
𝛾d,max = −0.1008LL + 21.16 (4) 0.76 0.91 0.69 4.05
𝛾d,max = −0.2283PL + 21.88 (5) 0.70 1.04 0.82 4.80
γd,max (GMDH model) (24) 0.90 0.60 0.46 2.70
𝜔opt (%):
𝜔opt = 0.24LL + 0.63PL − 3.13 (Iraqi soils) (6) 0.80 3.97 3.18 17.52
𝜔opt = 0.14LL + 0.54PL (US soils) (7) 0.79 2.95 2.42 13.34
𝜔opt = 0.92PL (8) 0.74 3.87 2.99 16.50
𝜔opt = 0.8442PL + 0.1076 (9) 0.74 3.27 2.66 14.62
𝜔opt = 0.3802LL + 2.4513 (10) 0.76 3.03 2.21 12.18
𝜔opt = 0.323LL + 0.157PL (11) 0.79 3.22 2.40 13.20
ωopt (GMDH model) (25) 0.92 1.81 1.38 7.58
100
80
60
SCF
40
AI-Khafaji (1993), Eq (1)
AI-Khafaji (1993), Eq (2)
Sridharan and Nagaraj
20 (2005), Eq (3)
Gunayin (2009), Eq (4)
Gunayin (2009), Eq (5)
GMDH
0
-20 -10 0 10 20
Er (%)
∑M �
h − ti ��
MAPE = 1∑� Mi × 100 (29)
1 hi
where M denotes the total number of data, the hi and ti are the actual and predicted output values for
the ith outputs, respectively. h̄ i and t̄i are, respectively, the average of the actual and predicted outputs.
The lower the RMSE, MAD and MAPE values the better the model performance. Under ideal conditions
an accurate and precise method gives R of 1.0, RMSE, MAPE and MAD of 0.
In Table 4, the predictability of the GMDH model is statistically given. It is clearly evident that the
evolved GMDH-type NN in terms of simple polynomial equations could successfully model and predict
the output of testing data that has not been used during the training process.
In Table 5, the predictability of the GMDH models for all data-set are statistically compared with the
empirical equations. It can be seen that the least disparities of these parameters are achieved by the
GMDH approach. Therefore, the developed ANN models are more efficient than the existing empirical
10 A. ARDAKANI AND A. KORDNAEIJ
100
80
60
SCF
formulas. A graphic representation of the comparative accuracy of the correlations and the proposed
methods are shown by plotting the Scaled relative error, (Er) vs. Scaled cumulative frequency (SCF).
(ti − hi )
Er (%) = × 100 (30)
hi
As is seen in Figures 5 and 6, broader ranges of prediction are given by all of the previously proposed
equations in comparison to suggested equations. The proposed GMDH models to predict both com-
paction parameters are more accurate in comparison with all of the other empirical correlations.
Where mi and mj are input and output parameters respectively. Rij, which ranges between 0 and 1, shows
the strength of the relation between each input parameter and the model output. In other words, if
the target variable has no relation with the input, then the Rij value is zero, while the value of Rij closer
to 1 indicates the further influence of the input parameter. The obtained strength of relations for the
problem in hand is shown in Figure 7. As displayed in this figure, the LL and PL are the most influential
parameters on the model outputs (γd,max and ωopt).
Figure 8 shows the effect of changing or error in measuring the input parameters (inaccuracies)
on the proposed models to predict the compaction parameters. Various constant change rates (–10
to 10%) are selected in the study. For every input neuron, mean absolute percent error (MAPE) in the
EUROPEAN JOURNAL OF ENVIRONMENTAL AND CIVIL ENGINEERING 11
γd,max ωopt
1
0.95
0.90
0.85
Rij
0.80
0.75
0.70
LL PL FC SC
Input parameter
γd,max ωopt
3.1
LL PL FC SC LL PL SC FC
8.5
2.9
MAPE (%)
MAPE (%)
2.8
2.7 7.58
-10 -5 0 5 10 -10 -5 0 5 10
Change in input (%) Change in input (%)
Figure 8. Effect of the change in the input parameters on the output MAPE.
outputs, as a result of change in the input neuron, is observed. It can be noticed from the Figure 8 that
the developed model to predict γd,max is considerably influenced by changing or error in measuring
LL value and the MAPE increases greatly by changing it. For example 10% error in measuring LL in the
laboratory may cause about 3% error (MAPE ≈ 3) in predicting γd,max by proposed model. Then it requires
more accuracy to measure this parameter in laboratory. As seen in Figure 8, the proposed model to
predict ωopt is notably affected by changing all three parameters LL, PL and FC. For example 10% error
in determining each of these parameters (LL, PL or FC) in the laboratory may cause about 8% error
(MAPE ≈ 8) for prediction of ωopt by derived model. Therefore, by increasing the accuracy in determining
each of these parameters in laboratory, the ωopt predicted by the model will be more accurate.
7. Conclusions
In present study, an attempt has been made to predict the compaction parameters of soils by neural
network simulations using a database consisting of 212 soil samples. Using the same database the
12 A. ARDAKANI AND A. KORDNAEIJ
predictability of the proposed models is statistically given. Furthermore, the sensitivity analyses of the
obtained models were used to study the influence of input parameters on models outputs. The results
indicate that:
• The results obtained from GMDH-type neural network in this study agreed well with the measured
γd,max and ωopt. Therefore, reliable predicting capabilities were obtained.
• The proposed GMDH models to predict the compaction parameters are more accurate in compar-
ison with all of the other empirical correlations.
• The sensitivity analysis shows that the LL and PL of soil are the most influential parameters on the
model outputs (γd,max and ωopt).
• The proposed models to predict γd,max is considerably influenced by error in measuring LL value
and the mean absolute percent error (MAPE) increases greatly by changing it. Then it needs more
accuracy to measure this parameter in laboratory.
• The proposed model to predict ωopt is considerably influenced by error in measuring LL, PL and
FC. Then for more accurate prediction of ωopt by proposed GMDH model, it needs more accuracy
to measure each of these parameters in laboratory.
It may thus be concluded that all predictions based on empirical approaches require the determi-
nation of input parameters, which are prone to uncertainties and inaccuracies. Therefore, a correlation
based on all of the soil parameters can be more useful to limit these uncertainties.
Nomenclature
Disclosure statement
No potential conflict of interest was reported by the authors.
EUROPEAN JOURNAL OF ENVIRONMENTAL AND CIVIL ENGINEERING 13
References
Al-Khafaji, A. N. (1993). Estimation of soil compaction parameters by means of Atterberg limits. Quarterly Journal of
Engineering Geology and Hydrogeology, 26, 359–368. doi:10.1144/GSL.QJEGH.1993.026.004.10
Ardalan, H., Eslami, A., & Nariman-Zadeh, N. (2009). Piles shaft capacity from CPT and CPTu data by polynomial neural
networks and genetic algorithms. Computers and Geotechnics, 36, 616–625. doi:10.1016/j.compgeo.2008.09.003
ASTM D698. (2003). Standard test methods for laboratory compaction characteristics of soil using standard effort. West
Conshohocken, PA: American Society for Testing and Materials (ASTM) International.
Atashkari, K., Nariman-Zadeh, N., Gölcü, M., Khalkhali, A., & Jamali, A. (2007). Modelling and multi-objective optimization
of a variable valve-timing spark-ignition engine using polynomial neural networks and evolutionary algorithms. Energy
Conversion and Management, 48, 1029–1041. doi:10.1016/j.enconman.2006.07.007
Das, B. M. (2010). Principles of geotechnical engineering (7th ed.). Stamford, CT: Cenage Learning.
Edincliler, A., Cabalar, A. F., & Cevik, A. (2013). Modelling dynamic behaviour of sand–waste tires mixtures using Neural
Networks and Neuro-Fuzzy. European Journal of Environmental and Civil Engineering, 17, 720–741. doi:10.1080/19648
189.2013.814552
Farlow, S. J. (1984). Self-organizing method in modelling: GMDH type algorithm (Vol. 54). New York, NY: CRC Press/Marcel
Dekker.
Günaydın, O. (2009). Estimation of soil compaction parameters by using statistical analyses and artificial neural networks.
Environmental Geology, 57, 203–215. doi:10.1007/s00254-008-1300-6
Ivakhnenko, A. G. (1971). Polynomial theory of complex systems. Systems, man and cybernetics. IEEE Transactions on, SMC
-1, 364–378. doi:10.1109/TSMC.1971.4308320
Jamali, A., Nariman-zadeh, N., Darvizeh, A., Masoumi, A., & Hamrang, S. (2009). Multi-objective evolutionary optimization
of polynomial neural networks for modelling and prediction of explosive cutting process. Engineering Applications of
Artificial Intelligence, 22, 676–687. doi:10.1016/j.engappai.2008.11.005
Kalantary, F., Ardalan, H., & Nariman-Zadeh, N. (2009). An investigation on the Su–NSPT correlation using GMDH type neural
networks and genetic algorithms. Engineering Geology, 104, 144–155. doi:10.1016/j.enggeo.2008.09.006
Kalantary, F., & Kordnaeij, A. (2012). Prediction of compression index using artificial neural network. Scientific Research and
Essays, 7, 2835–2848. doi:10.5897/SRE12.297
Kang, F., Li, J. S., Wang, Y., & Li, J. (2016). Extreme learning machine-based surrogate model for analyzing system reliability
of soil slopes. European Journal of Environmental and Civil Engineering, 1–22. doi:10.1080/19648189.2016.1169225
Kolay, E., & Baser, T. (2014). Estimating of the dry unit weight of compacted soils using general linear model and multi-layer
perceptron neural networks. Applied Soft Computing, 18, 223–231. doi:10.1016/j.asoc.2014.01.033
Mohammadzadeh, S. D., Bolouri Bazaz, J., & Alavi, A. M. (2014). An evolutionary computational approach for formulation
of compression index of fine-grained soils. Engineering Applications of Artificial Intelligence, 33, 58–68. doi:10.1016/j.
engappai.2014.03.012
Mola-Abasi, H., Eslami, A., & Tabatabaeishorijeh, P. (2013). Shear wave velocity by polynomial neural networks and genetic
algorithms based on geotechnical soil properties. Arabian Journal for Science and Engineering, 38, 829–838. doi:10.1007/
s13369-012-0525-6
Momeni, M., Nazir, R., Jahed Armaghani, D., & Maizir, H. (2014). Prediction of pile bearing capacity using a hybrid genetic
algorithm-based ANN. Measurement, 57, 122–131. doi:10.1016/j.measurement.2014.08.007
Nariman-Zadeh, N., Darvizeh, A., & Ahmad-Zadeh, G. R. (2003). Hybrid genetic design of GMDH-type neural networks using
singular value decomposition for modelling and prediction of the explosive cutting process. Proceedings of the Institution
of Mechanical Engineers, Part B: Journal of Engineering Manufacture, 217, 779–790. doi:10.1243/09544050360673161
Proctor, R. R. (1933). Fundamental principles of soil compaction. Engineering News Record, 111, 245–248.
Rout, S. K. (2009). Prediction of relative density of sand with particular reference to compaction energy (Doctoral dissertation),
National Institute of Technology, Rourkela.
Sridharan, A., & Nagaraj, H. B. (2005). Plastic limit and compaction characteristics of finegrained soils. Proceedings of the
Institution of Civil Engineers - Ground Improvement, 9, 17–22.
Terzaghi, K., Peck, R. B., & Mesri, G. (1996). Soil mechanics in engineering practice. John Wiley and Sons, New York.
Tokar, S. A., & Johnson, P. A. (1999). Rainfall-runoff modeling using artificial neural networks. Journal of Hydrologic Engineering,
4, 232–239. doi:10.1061/(ASCE)1084-0699(1999)4:3(232)
Yao, X. (1999). Evolving artificial neural networks. Proceedings of the IEEE, 87, 1423–1447. doi:10.1109/5.784219
Žlender, B., Jelušič, P., & Boumezerane, D. (2012). Planning geotechnical investigation using ANFIS. Geotechnical and
Geological Engineering, 30, 975–989. doi:10.1007/s10706-012-9520-7
14 A. ARDAKANI AND A. KORDNAEIJ