Download as pdf or txt
Download as pdf or txt
You are on page 1of 19

Marine Georesources & Geotechnology

ISSN: 1064-119X (Print) 1521-0618 (Online) Journal homepage: http://www.tandfonline.com/loi/umgt20

GMDH-GEP to predict free span expansion rates


below pipelines under waves

Mohammad Najafzadeh & Farid Saberi-Movahed

To cite this article: Mohammad Najafzadeh & Farid Saberi-Movahed (2018): GMDH-GEP
to predict free span expansion rates below pipelines under waves, Marine Georesources &
Geotechnology, DOI: 10.1080/1064119X.2018.1443355

To link to this article: https://doi.org/10.1080/1064119X.2018.1443355

Published online: 12 Mar 2018.

Submit your article to this journal

Article views: 2

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at


http://www.tandfonline.com/action/journalInformation?journalCode=umgt20
MARINE GEORESOURCES & GEOTECHNOLOGY
https://doi.org/10.1080/1064119X.2018.1443355

none defined

GMDH-GEP to predict free span expansion rates below pipelines under waves
Mohammad Najafzadeha and Farid Saberi-Movahedb
a
Department of Water Engineering, Faculty of Civil and Surveying Engineering, Graduate University of Advanced Technology, Kerman, Iran;
b
Department of Mathematics, Faculty of Science and Modern Technology, Graduate University of Advanced Technology, Kerman, Iran

ABSTRACT ARTICLE HISTORY


In this research, group method of data handling (GMDH) as a one of the self-organized approaches is Received 28 November 2017
utilized to predict three-dimensional free span expansion rates around pipeline due to waves. The GMDH Accepted 16 February 2018
network is developed using gene-expression programming (GEP) algorithm. In this way, GEP was KEYWORDS
performed in each neuron of GMDH instead of polynomial quadratic neuron. Effective parameters on the Free span expansion;
three-dimensional scour rates include sediment size, pipeline geometry, and wave characteristics gene-expression
upstream of pipeline. Four-dimensionless parameters are considered as input variables by mean of programming; group
dimensional analysis technique. Furthermore, scour rates along the pipeline, vertical scour rate, and method of data handling;
additionally scour rates in the left and right of pipeline are determined as output parameters. Results of pipeline; scour rate
the proposed GMDH-GEP models for the training stages and testing ones are evaluated using various
statistical indices. Performances of the GMDH-GEP models are compared with artificial neural network
(ANN), GEP, GMDH, and traditional equations-based regression models. Moreover, sensitivity analysis and
parametric study are conducted to perceive influences of different input parameters on the three-
dimensional scour rates.

1. Introduction submarine pipelines (Bernetti et al. 1990; Chiew 1990; Hansen


1992; Liang, Cheng, and Li 2005; Cheng et al. 2009; Zang et al.
Over the past four decades, scouring phenomena below
2009; Wu and Chiew 2012; Cheng et al. 2014). Knowledge of
pipelines has been considered as one of the most remarkable
three-dimensional scour developments is relatively restricted.
issues in fields of ocean and coastal engineering. Local scour
In spite of the reported experimental databases, it is essential
may be localized along the pipeline because of piping mech-
to yield an accurate prediction for the three-dimensional scour
anism and uneven seabed after installing the pipeline. As local
rates below pipelines. In case of experimental studies, it cannot
scour spreads along the pipeline, free spans are expanded. Free
be expected to find comprehensively influences of various
spans are occasionally characterized as culprits for pipeline
parameters on the scouring.
failures in the piping engineering due to the fact that free-
Empirical equations-based regression models are limited to
spanned pipelines are more susceptible to structural damages
the range of experimental databases. Therefore, it should be
related to over-stressing and human activities. From investiga-
said that empirical equations occasionally have not high
tions of the pipeline safety, it can be pointed out that free
generalization capacity for predicting the maximum scour
spans are useful to provide the pipeline stability. Moreover,
depth below pipelines.
free spans can lead to natural self-burial into the scour holes
In the recent decade, data mining models and data driven
formed below the pipelines (Palmer 1996; Sumer and Fredsøe
ones, as artificial neural networks (ANNs), adaptive neuro-
2002). A schematic diagram for free span development due to
fuzzy inference systems (ANFIS), linear genetic programming
local scour is shown in Figure 1. From Figure 1, it can be
(LGP), classification and regression tree (CART), genetic
found that the scour rate is propagated in two opposite direc-
programming (GP), gene-expression programming (GEP),
tions along the pipeline. One of the most significant cases in
and multivariate adaptive regression splines (MARS) were
design of pipelines is the appraising the free span length and
successfully applied to predict the scour depth below pipelines.
the free span growth rates during the storms, which both of
These predictive methods have performed promisingly good
them are contingent upon the flow characteristics, sediment
validations for observed data sets, compared to the regression
properties, and pipeline properties. The present level of knowl-
models (Azamathulla and Ghani 2010; Azamathulla, Guven,
edge on the three-dimensional scour developments still
and Demir 2011; Azamathulla and Zakaria 2011; Etemad-
requires further investigations (Sumer and Fredsøe 2002;
Shahidi, Yasa, and Kazeminezhad 2011; Zanganeh, Yeganeh-
Cheng et al. 2009).
Bakhtiary, and Bakhtyar 2011; Azamathulla and Yusoff 2013;
Several experimental investigations and numerical ones
Yasa and Etemad-Shahidi 2014; Haghiabi 2016). The artificial
were performed to characterize three-dimensional scouring
intelligence approaches for evaluation of scour rates around
process, local scour depth, and time scale of local scour at

CONTACT Mohammad Najafzadeh moha.najafzadeh@gmail.com; m.najafzadeh@kgut.ac.ir Department of Water Engineering, Faculty of Civil and Surveying
Engineering, Graduate University of Advanced Technology, P.O.Box 76315-116, Kerman, Iran.
Color versions of one or more of the figures in the article can be found online at www.tandfonline.com/umgt.
© 2018 Taylor & Francis
2 M. NAJAFZADEH AND F. SABERI-MOVAHED

gene-expression programming (GMDH-GEP) for predicting


three-dimensional scour propagation speed (free span expan-
sion rate) along the pipeline. The most striking improvement
of this study is to run a GEP model in each neuron of GMDH
rather than a quadratic polynomial. In this case, GMDH will
no longer need to update weighting coefficients of polynomial
neurons through the training stages by algorithms such as
back propagation (BP), particle swarm optimization (PSO),
and genetic algorithm (GA) to present an improved structure
of GMDH, as can be found in literature (Amanifard et al.
Figure 1. Three-dimensional scour process below pipeline (re-drawn from
2014; Najafzadeh, Barani, and Azamathulla 2014; Najafzadeh,
Cheng et al. 2009). Barani, and Hessami-Kermani 2014a; Shaghaghi et al. 2017). It
is conceivable that volume of computations for GMDH net-
work will decrease. In this study, the GMDH-GEP is coded
using the MATLAB programming language. Qualitative
submarine pipelines under waves and current have not been results of this innovation are compared with GMDH, GEP,
applied yet. ANN, and traditional equations. Ultimately, influence of pipe-
Recently, GMDH networks have been applied to predict line embedment depth, shields parameters, and flow incidence
the scour depth around offshore structures (Najafzadeh, angle on the scour propagation will be elucidated.
Barani, and Azamathulla 2014; Najafzadeh, Barani, and
Hessami-Kermani 2014a). Performing the GMDH networks
produced relatively more accurate scour depth prediction 2. Effective parameters on the three-dimensional
than conventional nonlinear regression approaches. In fact, scour rates below pipelines under waves
the main feature of the GMDH networks is to build analytical
functions within feed forward network in form of quadratic In the recent years, a myriad of efforts have been made to
polynomial whose weighting coefficients are obtained using investigate three-dimensional pipeline scour by virtue of
regression method (Kalantary, Ardalan, and Nariman-Zadeh understanding the lateral propagation of the scour hole along
2009; Najafzadeh, Barani, and Azamathulla 2014; Najafzadeh, the pipeline. Values of free span depend on dimensions of the
Barani, and Hessami-Kermani 2014a). Applying the GMDH scour hole. For example, the longitudinal propagation of a
networks-based iterative and evolutionary algorithms pipeline scour hole (VH) in terms of its lateral dimension is
provided prosperous performances in various field of sciences ratio of variation of scour hole length, LL, to time interval,
such as evaluation of explosive cutting process and tool life dt, expressing dLL/dt. Similarly, vertical scour rate, VV, is
testing in gun drilling, constructing optimal educational test, dS/dt in which dS and S is scour depth.
modeling of semiconductor, control engineering, chemical Basically, scour depth which is vertical distance between
process, scour depth prediction below pipeline, prediction original bed level of sediment and deepest point of scour hole,
of uplift capacity of suction caisson, and stable open channel varying during the time. As scour depth remains constant with
design (Astakhov and Glitsky 2005; Witczak et al. 2006; increase of time, reaching ultimate scour depth (Cheng et al.
Abdel-Aal and El-Alfy 2009; Jamali et al. 2009; Lin 2012; 2014).
Abdolrahimi, Nasernejad, and Pazuki 2014; Najafzadeh, From laboratory studies on the three-dimensional scour
Barani, and Azamathulla 2014; Najafzadeh, Barani, and rates due to waves, three-dimensional scour rates depend on
Hessami-Kermani 2014a; Masoumi Shahr-Babak, Khanjani, the pipeline geometry, wave properties, and physical charac-
and Qaderi 2016; Shaghaghi et al. 2017). From these applica- teristics of bed material (Chiew 1990; Liang, Cheng, and Li
tions, one of the useful features of the GMDH networks is 2005; Cheng et al. 2009; Zang et al. 2009; Wu and Chiew
analytical equations which can be obtained using the second 2012, Wu and Chiew 2013; Cheng et al. 2014). In this way,
order polynomials (or quadratic polynomial). In addition, the relationship function between the three-dimensional scour
volume of computations for the GMDH networks are less propagation and effective parameters can be expressed as
than the GP, ANFIS, and ANNs models because only six follows:
weighting coefficients should be yielded in each neuron of f ðVH ; VL ; VR ; VV ; e; D; Uw ; d50 ; T; tan /; a; q; qs ; l; gÞ ¼ 0
the GMDH network. Meantime, application of GEP model
ð1Þ
with two input and one output parameter instead of
performance of the quadratic polynomial with six weighting where VH ¼ the scour rate along the pipeline; VL ¼ scour rate at
coefficients can be effectively captured results in comparison left of pipeline; VR ¼ scour rate at right of pipeline; VV ¼
with GMDH model. vertical scour rate; e ¼ embedment depth; D ¼ pipe diameter;
According to the previous research, the scour depth below Uw ¼ maximum undisturbed orbital rate at the sea bottom just
pipeline by artificial intelligence approaches for various con- above the waves boundary; d50 ¼ median sediment size;
ditions of flow and bed sediments was assessed. Three-dimen- T ¼ wave period; φ ¼ repose angle for bed sediments, α’ ¼ flow
sional scour rates below pipelines exposed to the waves have incident angle to the pipeline (angle of attack); ρ ¼ mass density
not been evaluated yet by artificial intelligence models. In this of water; qs ¼ mass density of bed material; μ ¼ dynamic
study, structure of the GMDH network is improved based on viscosity of water; and g ¼ acceleration due to gravity.
MARINE GEORESOURCES & GEOTECHNOLOGY 3

Furthermore, in Eq. (1), raw variables of u; q; qs ; and l are Table 1. Ranges of input–output parameters for the scour rate prediction.
taken into account as constant parameters and their values Parameters Range
are 32°, 1,000 (kg/m3), 2,650 (kg/m3), and 0.001 (Pa .s), H(m) 0.13–0.17
respectively. T(s) 1.5–2
Uw(m/s) 0.29–0.45
From the use of predictive models for scour modeling e(mm) 5–20
below pipelines, it was concluded that use of grouped nondi- VH ðmm=sÞ 1.6–4.8
mensional parameters, as input–output variables of artificial VL ðmm=sÞ 1.6–5.7
VR ðmm=sÞ 1.5–5.7
intelligence approaches, produced better predictions of scour VV ðmm=sÞ 0.745–2.935
depth than that of dimensional parameters (Etemad-Shahidi, hw 0.18–0.3
Yasa, and Kazeminezhad 2011; Azamathulla and Yusoff KC 8.7–18
e/D 0.1–0.4
2013; Najafzadeh, Barani, and Azamathulla 2014; Najafzadeh, aðradÞ 0–0.7068
Barani, and Hessami-Kermani 2014a; Haghiabi 2016). Hence, VH� 1.27–5.162
the following functions were generated using Buckingham VL� 1.19–4.527
VR� 1.11–4.49
theorem: VV� 0.592–2.405
VH
VH� ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi3ffiffiffi
ð gððqs =qÞ 1Þd50 =D tan /Þ ð2Þ back to the original flume bed are achieved through two
¼ f ðhw ; KC; e=D; sin aÞ 1:20 slopes on the either ends of the sand pit. The upstream
VR end of the sand pit was 13.5 m from the flow inlet, and the
VR� ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi3ffiffiffi downstream end was 6.5 m from the flow outlet. A clear pipe-
ð gððqs =qÞ 1Þd50 =D tan /Þ ð3Þ
line with smooth surface, diameter of 50 mm, and wall thick-
¼ f ðhw ; KC; e=D; sin aÞ ness of 8 mm was utilized in their study. The flow attack angles
VL examined in the experiments were α¼0°, 15°, 30°, and 45°.
VL� ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi3ffiffiffi
ð gððqs =qÞ 1Þd50 =D tan /Þ ð4Þ Also, median sediment size and relative density of sediments
grain ðqs =qÞ were 0.37 and 2.7 mm, respectively. Cheng
¼ f ðhw ; KC; e=D; sin aÞ
et al. (2014) reported that critical Shields parameter for motion
VV of the bed sediments utilized was 0.037, evaluated based on the
VV� ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi3ffiffiffi
ð gððqs =qÞ 1Þd50 =D tan /Þ ð5Þ Soulsby (1997) approach. All experimental programs have
¼ f ðhw ; KC; e=D; sin aÞ been conducted in the live bed status. In addition, raw datasets
to build nondimensional parameters were presented in
which hw and KC are Shields parameter under waves and Appendix A. Data set used to model the scour rates are newly
Keulegan–Carpenter number, respectively: reported and were not utilized by other researchers. A
hw ¼ sw =qgðqs =q 1Þd50 ð6Þ grouped-dimensionless equations given by the GMDH model
would be useful to present physical insights for results of the
KC ¼ Uw T=D ð7Þ
GMDH-GEP networks.
where sw is the wave-induced shear stress on the seabed and is
expressed as follows:
3. Framework of the proposed approaches
1
sw ¼ qfw Uw2 ð8Þ In this section, definitions of the GMDH, GEP, and ANN
2
models are presented. In addition, details related to the devel-
where fw is the wave-induced shear stress on the seabed and opments of the GMDH model based on the GEP are clearly
was proposed by (Soulsby 1997). elucidated.
fw ¼ 0:237r 0:52 ð9Þ
r ¼ Uw T=4pd50 ð10Þ 3.1. Description of GMDH model
In this research, Eqs. (2)–(5) have been only applied to The GMDH is based on the principle of heuristic self-organiz-
develop the GMDH-GEP for predicting the three-dimensional ing as proposed by Ivakhnenko in the 1960s. It is an evolution-
scour rates. The datasets used to predict the scour velocities ary computation technique, which has a series of operations of
were collected from the Cheng et al. (2014) (38 datasets). seeding, rearing, corresponding, selection and rejection of
Table 1 presents the ranges of datasets for modeling the scour seeds corresponding to the determination of the input
rates. Out of the datasets, about 67% (25 data sets) and 33% variables, structure and parameters of model, and selection
(13 datasets) were selected randomly to perform training of model by principle of termination (Ivakhnenko 1971;
and testing stages one for GMDH-GEP models, respectively. Ivakhnenko and Ivakhnenko 2000).
In case of experimental conditions of datasets, Cheng et al. The GMDH network is a very flexible algorithm and it can
(2014) performed three-dimensional scour experiments in a be hybridized by other evolutionary algorithms, such as
wave flume with 50 m in length, 4 m in width, and 2.5 m in genetic algorithm (Amanifard et al. 2008; Mehrara et al.
depth. A concrete sand pit of 4 m long, 4 m wide, and 2009), genetic programming (Iba and de Garis 1996), particle
0.25 m deep was built in the test section. The transitions from swarm optimization (Onwubolu 2008), and back propagations
the original flume bed to test section and from test section (Sakaguchi and Yamamoto 2000; Srinivasan 2008).
4 M. NAJAFZADEH AND F. SABERI-MOVAHED

The formal definition of system identification problem is to difference between actual output, y, and the calculated one,
find a function ^f that can be approximately used instead of ^y, for each pair of xi and xj as input variables is minimized.
actual function f, to predict the output ^y for a given input Indeed, it can be seen that a tree of polynomials is constructed
vector X ¼ ðx1 ; x2 ; x3 ; . . . ; xn Þ as close as possible to its actual using the quadratic form given in Eq. (14) whose weighting
output y. Therefore, given n observation of multi-input- coefficients can be obtained by least-squares sense. In this
single-output data pairs so that way, the weighting coefficients of quadratic function Gi are
obtained to optimally fit the output in the whole set of
yi ¼ f ðxi1 ; xi2 ; xi3 ; . . . ; xin Þði ¼ 1; 2; . . . ; M Þ ð11Þ
input–output data pairs, that is
It is now possible to train the GMDH network to predict
P
M
the output values ^yi for any given input vector ðyi Gi ðÞÞ2
X ¼ ðxi1 ; xi2 ; xi3 ; . . . ; xin Þ, that is E¼
i¼1
! min ð16Þ
M
^yi ¼ ^f ðxi1 ; xi2 ; xi3 ; . . . ; xin Þði ¼ 1; 2; . . . ; M Þ ð12Þ
In the basic form of the GMDH algorithm, all the possibi-
To solve this problem, the GMDH builds the general lities of two independent variables out of total n input vari-
relationship between output and input variables in the form ables are taken to construct the regression polynomial in the
of mathematical description, which is also called reference. form of Eq. (15) that best fits the dependent observations
Through the GMDH network, the square of difference ðyi ; i ¼ 1; 2; . . . ; MÞ in a least-square sense. Consequently,
between the actual output and the predicted one is minimized, Cn2 ¼ nðn 1Þ=2 neurons of quadratic polynomial will be built
that is: up in the first layer of the feed forward network from observa-
M h
X i2 tions fðyi ; xip ; xiq Þ; ði ¼ 1; 2; . . . MÞg for different
^f ðxi1 ; xi2 ; xi3 ; . . . ; xin Þ yi ! min ð13Þ p; q 2 f1; 2; . . . ; ng. In other words, it is now possible to con-
i¼1 struct M data triples fðyi ; xip ; xiq Þ; ði ¼ 1; 2; . . . MÞg from
observation using such p; q 2 f1; 2; . . . ; ng in the form
General connection between inputs and output variables
2 3
can be expressed by a complicated discrete form of the x1p x1q y1
Volterra function, a series in the form of: 4 x2p x2q y2 5 ð17Þ
X
n n X
X n xmp xmq ym
y ¼ w0 þ wi xi þ wij xi xj
i¼1 i¼1 j¼1 Using the quadratic subexpression in the form of Eq. (15)
ð14Þ
X
n X
n X
n for each row of M data triples, the following matrix equation
þ wijk xi xj xk þ . . . can be readily obtained as:
i¼1 j¼1 k¼1
AW ¼ Y ð18Þ
which is known as the Kolmogorov–Gabor polynomial
(Ivakhnenko 1971; Farlow 1984; Sanchez, Shibata, and Zadeh where W is the vector of unknown weighting coefficients of
1997). In the present study, quadratic polynomial of the the quadratic polynomial in Eq. (18)
GMDH network is used that is written as:
Quadratic: W ¼ fw0 ; w1 ; w2 ; w3 ; w4 ; w5 gTr ð19Þ

^y ¼ Gðxi ; xj Þ ¼ w0 þ w1 xi þ w2 xj þ w3 xi xj þ w4 xi2 þ w5 xj2 The superscript Tr represents transpose of matrix.


ð15Þ Y ¼ fy1 ; y2 ; y3 ; . . . ; yM gTr ð20Þ
The GMDH network-based complex theory has many
The vector of observation is values of outputs. It can be
advantages. The first one is to provide adaptive network repre-
readily seen that
sentations which can be tailored to the given problem. Also,
2 3
the GMDH algorithm performed training stage rapidly using 1 x1p x1q x1p x1q 2
x1p 2
x1q
the standard orthogonal least square (OLS) sense. In fact, 6 2 2 7
A ¼ 4 1 x2p x2q x2p x2q x2p x1q 5 ð21Þ
using of OLS method eliminates the need to search for weight- 2 2
1 xmp xmq xmp xmq xmp xmq
ing coefficients of polynomial neurons (Najafzadeh, Barani,
and Hessami-Kermani 2014a).
The least-squares technique from multiple-regression
Although GMDH can generate a systematic procedure for
analysis leads to the solution of the normal equations in the
solution of complex system modeling, it has also several lim-
form of:
itations. For instance, several restrictions of the GMDH mod-
els are related to the selection of input parameters, W ¼ ðATr AÞ 1 ATr Y ð22Þ
inaccuracies in parameter prediction, multi-col-linearity, and
reduction of complexity (Onwubolu 2009). which determines the vector of the best weighting coefficients
This network of connected neurons builds the general of the quadratic Eq. (15) for the whole set of M data triples. It
mathematical relation of inputs and output variables given should be noted that this procedure is repeated for each
in Eq. (14). The weighting coefficients of Eq. (15) are calcu- neuron of the next hidden layer according to the connectivity
lated using regression techniques (Farlow 1984) so that the topology of the network.
MARINE GEORESOURCES & GEOTECHNOLOGY 5

3.2. Development of the GEP model Table 3. Best fitness function values for the
proposed GEP models.
One of the most recent technique called GEP was developed Parameters Values
which is an extension of the GP approach. The GEP is a search VH� 0.781
model that evolves computer programs in forms of mathemat- VL� 0.159
ical expressions, decision trees, and logical expressions. There VR� 0.267
VV� 0.126
is a fundamental difference between GP and GEP models. In
GP, the individuals are nonlinear entities of different sizes
and shapes, introduced as parse trees, whereas in GEP the Expressions given by the GEP models to evaluate the scour
individuals are also nonlinear entities of various sizes rates were formulated as,
and shapes, known as expression trees (Ferreira 2001, 2006; VH� ¼ 5:092 þ 7:732 cos½ðe=DÞ þ sin a�
Koza 1992.
cosðKCÞ ð24Þ
This research represents GEP models for evaluation of the þ 36:18½ðe=DÞ hw �2 þ0:08942
scour rates around pipelines exposed to the waves. The GEP ½ðe=DÞ sin a�
approach is coded in forms of linear chromosomes, which VL� ¼ þ 7:732 9:89hw ðsin aÞ2 þ0:061KC 0:0613ðe=DÞ
are then expressed into expression trees (ETs). In fact, the 0:0613KC sin a þ 8:845 sin a 8:845ðe=DÞ2
ETs are sophisticated computer programming which are
usually evolved to solve a practical problem, and are selected ð25Þ
accordingly to their fitness at solving that problem. The corre- VR� ¼ ð0:457KC 2:733 sin aÞhw þ 2:722 5:466ðe=DÞ
sponding empirical expressions can be obtained from these sin a
trees structures. A population of the ETs will discover traits, 4:642hw þ 0:9138 sin a þ 4:642
KC
and therefore will adapt to the particular problem they are ð26Þ
used to solve the vast majority of problems in different fields
of science (Ferreira 2001, 2006). VV� ¼ 11:8 þ 12:27 sinðhw sin aÞ 37:39ðhw Þ2
ð27Þ
Development of the GEP approach includes five steps. The þ 10:57ðsin a þ hw þ cosðe=DÞÞ
first step is to select the fitness function, fi, of an individual
program (i). This item is evaluated as follows:
Ct
X � � 3.3. Development of the GMDH model
fi ¼ ðMM �Cði;jÞ Terj �Þ ð23Þ
j¼1 As mentioned in the GMDH definitions section, structure of
the GMDH model is built using least square sense. Four
in which MM, C(ij), and Terj are the selection range, value GMDH models have been developed so as to predict three-
returned by the individual chromosome i for fitness case j, dimensional scour rates. Conspicuously, GMDH networks
the largest value for fitness case j. have three layers with, six neurons in the first layer, three neu-
In the second stage, the set of terminals Ter and the set of rons in the second layer, and output neuron. Using criteria
function f were selected to generate the chromosomes. In this error in each layer, selective neurons are highlighted and other
study, the terminal includes four independent parameters in ones are filtered. Error values for all proposed networks were
form of V � ¼ fhw ; e=D; KC; sin ag. presented in Table 4. For all the scour rates, error values of
To find the appropriate function set, four basic operators selective polynomial neurons in each layer were made bold in
(+, −, *, /) and basic mathematical functions (√, power, sin, Table 4. The criterion of termination for the GMDH is
cos, exp) were applied to predict the local scour depth model- defined by user. In final output, should computational
ing. The third step is to configure the chromosomal architec- error be less than error of criterion, computer program will
ture. The fourth step is selection of liking function. Finally, for be stopped (Najafzadeh, Barani, and Hessami-Kermani
the fifth stage, the sets of genetic operators which cause vari- 2014a). In addition, the values of termination error are given
ation and their rates are selected. The other details related to in Table 4.
the architecture of the GEP modeling were expressed in the Quadratic polynomial neurons extracted from the GMDH
literature (Ferreira 2001, 2006). model were expressed as,
To predict the three-dimensional scour rates, the functional
set and the operational parameters applied in the proposed
GEP models were presented in Table 2. Also, best fitness Table 4. Error values related to the neurons for the proposed GMDH models.
Position of neurons in network VH� VR� VL� VV�
values of the proposed GEP models were given in Table 3.
1st neuron of 1st layer 27.35 26.21 20.054 2.815
2nd neuron of 1st layer 0.779 0.553 0.44 0.126
Table 2. Properties of the proposed GEP model to predict the three-dimensional 3rd neuron of 1st layer 0.808 0.383 0.182 0.0752
scour rates. 4th neuron of 1st layer 0.702 0.542 0.438 0.126
5th neuron of 1st layer 0.85 0.394 0.198 0.0626
Parameters Definition Value 6th neuron of 1st layer 0.823 0.392 0.188 0.0642
P1 Population 100 1st neuron of 2nd layer 0.493 0.0657 0.0391 0.0264
P2 Generation 200 2nd neuron of 2nd layer 0.64 0.142 0.0516 0.027
P3 Max-genes 3 3rd neuron of 2nd layer 0.526 0.129 0.0912 0.0595
P4 Max-depth of tree 3 1st neuron of 3rd layer 0.314 0.0593 0.0311 0.0219
P5 Function set {+,−, square, *, /, sin, cos} Error termination 0.32 0.06 0.033 0.022
6 M. NAJAFZADEH AND F. SABERI-MOVAHED

�1
VH� 2 ¼ 0:06363 þ 0:4554KC 3:7498ðe=DÞ two descendant (lower) neurons. In fact, the GMDH-GEP is
ð28Þ
0:6247ðe=DÞKC 0:01298ðKCÞ þ28:0335ðe=DÞ 2 2 a feed forward network and a performance of a GEP model
�1 has two inputs and one output. Moreover, it is noted that
VH� 3 ¼ 3:882 13:70867ðe=DÞ þ 1:6277 sin a the function set used in the structure of GEP can be composed
ð29Þ
þ 2:318ðe=DÞ sin a þ 28:176ðe=DÞ2 3:1545ðsin aÞ2 of arithmetic operations (+, −, /, �, ^2) and function calls
�1 fex ; x; sin; cos; tan; log; sqrt; ln; powerg. Combination of the
VH� 4 ¼ 14:122 115:047hw þ 6:8871ðe=DÞ GMDH and GEP models is described in six steps as,
69:839ðe=DÞhw þ 274:321ðe=DÞ2 þ25:779ðhw Þ2 (1) For generation of first layer of the GMDH-GEP model,
ð30Þ as seen in Figure 2, x1 and x2 are selected as input parameters
�2 �1 �1 to form an expression GEPx1, x2(x1, x2) which approximates
VH� 1 ¼ 4:2681 þ 0:9693 VH� 2 2:9163 VH� 3 the output z1 using generating of tree structures and
�1 �1 � �1 �2 operational function such as crossover and mutation. In this
5:7337 VH� 2 VH� 3 þ2:7478 VH� 2 ð31Þ way, gene-expression programming model is performed in
� �1 � 2 all neurons of the first layer as,
þ 3:4308 VH� 3
�2 �1 �1 z1 ¼ GEPðx1 ; x2 Þ ð34Þ
VH� 4:3298 VH� 3 þ2:0334 VH� 4
¼ 4:60577
3
� z2 ¼ GEPðx1 ; x3 Þ ð35Þ
�1 �1 �1 � 2
2:943 VH� 3 VH� 4 þ2:2757 VH� 3 ð32Þ In which Z1 and Z2 are the output vectors of the first layer.
� �1 �2 (2) In this step, E value given by Eq. (16) for each neuron in
þ 1:20488 VH� 4 the first layer is computed by comparing between observed
�3 �2 �2 and calculated scour rates.
VH� 1 ¼ 8:497 5:229 VH� 1 þ0:829 VH� 3
�2 �2 � �2 �2 (3) In this stage, as mentioned in Section 3.1, neurons with
7:8181 VH� 1 VH� 3 þ4:4574 VH� 1 ð33Þ lower error of computation, compared to the criteria error
� �2 � 2 values should be selected.
þ 4:10596 VH� 3 (4) Selective neurons in the third stage are taken account as
input vectors for generation of the second layer.
In which superscript and subscript of each parameter (5) In the second layer, a GEP model is performed in each
present few pertaining layer and neuron, respectively. Mean- neuron as conducted in the first layers. Resting of this process
time, selective polynomial neurons extracted by GMDH net- is completely similar to that performed in the second, third,
work for scour rates at right and left of the pipeline and and fourth steps.
vertical scour rates were given in Appendix B. (6) As increased number of layers for the GMDH-GEP
model, a neuron is met as a final output in the latest layer.
In this study, the GMDH-GEP network was constructed
3.4. Development of the GMDH network using GEP
with six neurons in the first layer and based on the previous
model
six steps, successive layers of network are generated and only
In this section, the way of combining GMDH network with the best ones are chosen by the algorithm. The authors of cur-
GEP model is described. In the present research work, the rent research work performed the GMDH-GEP using different
GEP model is performed in each neuron of the GMDH net-
work instead of performing quadratic polynomial. Based on
Table 5. Properties of the proposed GEP model used in structure of the GMDH
Figure 2, x 2 fx1 ; x2 ; x3 ; x4 ; . . . ; xn g is a set of input variables model.
and Intermediate variables z 2 fz1 ; z2 ; z3 ; z4 ; . . . ; zðnÞ g Parameters Definition Value
2
represent simple tree expresses relationship between P1 Population 100
P2 Generation 150
P3 Max-genes 3
P4 Max-depth of tree 3
P5 Function set {+,−, square, *}
P6 Gene recombination rate 0.277
P7 Gene transportation rate 0.277
P8 Number of chromosomes 30
P9 Mutation rate 0.138
P10 Inversion rate 0.546

Table 6. Best fitness function values related to the selective neurons for the
proposed GEP models used in GMDH structures.
Position of neurons in network VH� VR� VL� VV�
1st neuron of 1st layer 0.519 0.578 0.426 0.265
4th neuron of 1st layer 0.582 0.612 0.541 0.245
6th neuron of 1st layer 0.596 0.607 0.458 0.253
1st neuron of 2nd layer 0.317 0.322 0.235 0.167
2nd neuron of 2nd layer 0.321 0.323 0.223 0.17
Figure 2. General structure of the GMDH-GEP network. 1st neuron of 3rd layer 0.302 0.321 0.222 0.162
MARINE GEORESOURCES & GEOTECHNOLOGY 7

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uM
Table 7. Values of setting parameters related to the proposed ANNs. uP
VH� VR� VL� VV�
u ðVi ðPredictedÞ Vi ðObservedÞÞ2
Setting parameters ti¼1
Epoch 30 20 30 20 RMSE ¼ ð37Þ
M
Iteration 6 13 22 10
MSE 1.3 0.0915 0.00109 0.0533 M �� �
100 X �
�Vi ðPredictedÞ Vi ðObservedÞ�
values of the population size and the best properties of the MAPE ¼ � � ð38Þ
M i¼1 Vi ðObservedÞ
GEP formation such as functional set and operational para-
meters being presented in Table 5. Moreover, Table 6 indi- P
M
ðVi ðPredictedÞ Vi ðObservedÞÞ
cated values of best fitness function extracted from optimum i¼1
structures of GMDH-GEP model. BIAS ¼ ð39Þ
M
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P M � �� 2
ð1=MÞ ViðPredictedÞ V Predicted ViðObservedÞ V Observed
i¼1
SI ¼ ð40Þ
P
M
ð1=MÞ ViðObservedÞ
i¼1

3.5. Development of the artificial neural network where V, V,� and M are scour rates around pipeline, average
of the scour rate around pipeline, and few samples,
Artificial neural network-based feed forward model and back
respectively.
propagation learning approach could be considered as a subset
of multilayer perceptron neural networks (MLP-NNs). The
topology of FFBP-NN is built as a set of neurons jointed by
links in several layers. The fundamental structure of the ANNs 4.1. Evaluation of the proposed GMDH-GEP models
model consists of an input layer, hidden layers and an output The three-dimensional scour rates have been predicted using
layer (or as output neuron) (Hornik, Stichcombe, and White the GMDH-GEP networks. Performances of results for the
1989). In this study, FFBP-NN models were trained using training and testing stages were presented in Table 8. Also,
levenberg-marquardt (MT) to predict three-dimensional scour qualitative comparing results of training and testing have been
rates around the pipeline. Four FFBP-NN models have been illustrated in Figures 3–6. From the training stages, Table 8
designed for the scour rates evaluation at right and left of the and Figure 3 indicated that the GMDH-GEP network pre-
pipeline, longitudinal scour rates, and vertical scour rates. dicted the longitudinal scour rate with relatively higher accu-
These models were developed by MATLAB 7.6.0 (R2008a) soft- racy (RMSE ¼ 0.309 and MAPE ¼ 9.22) and the correlation
ware. In addition, to stop program, MSE (mean square error) as coefficient of 0.95. For predicting the VL� parameter, the
a criterion of error is used. Setting parameters applied in ANN GMDH-GEP network provided relatively same accuracy
modeling were presented in Table 7. Each of improved FFBP-
NN model included four neurons in input layer, two hidden
layers with eight neurons, and one output neuron. Table 8. Results of performances for proposed models.
Training stage R RMSE MAPE BIAS SI
VH� (GMDH-GEP) 0.95 0.309 9.22 0 0.115
4. Results and implementations VL� (GMDH-GEP) 0.96 0.31 7.78 0 0.087
VR� (GMDH-GEP) 0.94 0.32 12.47 0 0.131
The statistical results of the GMDH-GEP networks, GMDH, VV� (GMDH-GEP) 0.94 0.162 10.07 0 0.128
GEP, and ANN models are presented in this section. In addition, Testing stage
� 0.93
VH (GMDH-GEP) 1.244 15.77 −0.158 0.165
results of performances were compared with those obtained VL� (GMDH-GEP) 0.96 0.358 13.037 0.127 0.123
using empirical equations-based regression approaches. VR� (GMDH-GEP) 0.90 0.97 17.18 0.376 0.125
Correlation coefficient (R), root mean square error (RMSE), VV� (GMDH-GEP) 0.92 0.215 16.92 0.11 0.159
VH� (GMDH) −0.38 1.792 62.71 0.207 0.675
mean absolute percentage of error (MAPE), BIAS, and scatter VL� (GMDH) 0.92 0.623 18.92 −0.251 0.197
index (SI) which are commonly used to evaluate errors VR� (GMDH) 0.932 0.626 19.14 −0.034 0.211
prediction of the scour rates in training and testing stages: VV� (GMDH) 0.876 0.237 21.65 0.0063 0.2
VH� (GEP) −0.255 1.67 55.41 0.202 0.675
X
M VL� (GEP) 0.97 0.291 8.15 −0.078 0.099
ðViðObservedÞ V ðObservedÞ ÞðViðPredictedÞ V ðPredictedÞ Þ �
VR (GEP) 0.97 0.212 6.57 0.0005 0.076
i¼1 VV� (GEP) 0.92 0.187 16.92 0.021 0.16
R¼ vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi VH� (FFBP-NN) −0.161 1.15 51.9 0.324 0.437
uX
u M VL� (FFBP-NN) 0.932 0.407 12.97 −0.079 0.142
u
u ðViðObservedÞ V ðObservedÞ Þ2 VR� (FFBP-NN) 0.962 0.289 8.8 −0.092 0.093
u i¼1 VV� (FFBP-NN) 0.948 0.171 11.81 0.078 0.13
u
u Eq. (47) −0.02 1.99 47.58 1.18 0.64
t � X ðV
M
iðPredictedÞ V ðPredictedÞ Þ2 Eq. (48) 0.405 0.95 38.07 0.0657 0.385
i¼1
Eq. (49) 0.938 1.73 53.75 −1.56 0.264
Eq. (50) 0.807 1.66 54.11 −1.59 0.165
ð36Þ Eq. (51) 0.87 0.66 46.15 −0.57 0.264
8 M. NAJAFZADEH AND F. SABERI-MOVAHED

Figure 3. Performances of the GMDH-GEP model for prediction of the VH� .

(RMSE ¼ 0.31 and MAPE ¼ 7.78) as statistical error para-


meters obtained from the VH� . Moreover, Figure 4 illustrated
that GMDH-GEP yielded relatively lower accuracy of the
scour rate prediction in term of R (0.94) in comparison with
that seen in Figure 5. From Table 8 and Figures 5 and 6, stat- Figure 5. Performances of the GMDH-GEP model for prediction of the VL� .
istical values in terms of R and SI indicated relatively same
accuracy for prediction of the vertical scour rate and scour rate
In the VR� prediction, results of statistical error parameters
at right of pipeline. The RMSE and MAPE values given for
demonstrated that the GMDH-GEP model produced more
predicting the VR� were 0.32 and 12.47 whilst these parameters
accurate prediction of scour rate (RMSE ¼ 0.97, BIAS ¼ 0.376,
for prediction of the VV� were yielded 0.162 and 10.7,
and SI ¼ 0.125) than those obtained using the VH� (RMSE ¼
respectively.
1.24, BIAS ¼ −0.158, and SI ¼ 0.165). In addition, error index
In the testing stages, Table 8 indicated that GMDH-GEP
of the MAPE given by the VR� is as same value as that obtained
networks were successful in prediction of the longitudinal
by the VL� . For the VV� prediction, the GMDH-GEP model
scour rate, scour rates at right and left of pipeline, and vertical
demonstrated remarkably good performance in terms of
scour rate. From Table 8 and Figure 3, the GMDH-GEP has
RMSE (0.215) and BIAS (0.11) compared to those yielded
provided the accurate longitudinal scour rate with RMSE of
for prediction of the VH� , VL� , and VR� .
1.244, MAPE of 15.77, and the correlation coefficient of
Due to the limitation of data points (38 datasets), six
0.93. In addition, statistical parameters of SI and BIAS were
weighting coefficients of a polynomial neuron (Eq. 15) cases
obtained 0.165 and 0.414, respectively. Table 8 and
over fitting. To prevent this process, a GEP was performed
Figures 4–6 showed that GMDH-GEP provided the VL� with
in each neuron of the GMDH network instead building of
relatively higher accuracy in terms of R ¼ 0.96 and MAPE ¼
polynomial neuron with six coefficients. The second one is
13.037, compared to the VR� (R ¼ 0.92 and MAPE ¼ 17.18)
related to the physical insights of the GMDH-GEP structure.
and VV� (R ¼ 0.90 and MAPE ¼ 16.92). The BIAS (0.127)
The proposed GMDH-GEP models have three layers (3-2-1)
and SI (0.123) parameters indicated compromisingly good
and only six polynomial neurons. In fact, selection of neurons
performance in testing stage for the VL� prediction.

Figure 4. Performances of the GMDH-GEP model for prediction of the VR� . Figure 6. Performances of the GMDH-GEP model for prediction of the VV� .
MARINE GEORESOURCES & GEOTECHNOLOGY 9

in the first layer based on the error criteria prevents over


fitting process and complexity of structure in future layers.
One of the useful features of GMDH networks is analytical
equations which can be obtained using the second order poly-
nomials (or quadratic polynomial). For instance, analytical
equations given by the GMDH-GEP network for predicting
the VH� were presented as follows:
ðVH� Þ11 ¼ GEPðsin a; e=DÞ ¼ 4:778
pffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffi
þ 4:298 sin a 4:298e=D 12:2ðe=D sin aÞ e=D
pffiffiffiffiffiffiffiffiffi
þ 3:426 sin aðsin a 3:4182Þ
ð41Þ
ðVH� Þ14 ¼ GEPðsin a; KCÞ ¼ 0:01975
þ 0:2245ðsin aÞ2 KC þ 0:3387 sin a ð42Þ
þ 0:3887KC 0:3597 sin a 0:005583ðKCÞ2
ðVH� Þ16 ¼ GEPðsin a; hw Þ ¼ 18:71 sin ahw 10
pffiffiffiffiffi ð43Þ Figure 7. The Proposed structure of the GMDH-GEP network for prediction of
2 the VH� .
þ 105:3ðsin aÞ hw 195 hw
qffiffiffiffiffiffiffiffiffiffiffiffi
ðVH Þ1 ¼ GEPððVH Þ1 ; ðVH Þ4 Þ ¼ 3:43 þ 0:6549 ðVH� Þ11
� 2 � 1 � 1
performance with R ¼ −0.38 and MAPE ¼ 62.71 compared
2:197ðVH� Þ11
0:01683ððVH� Þ11 Þ0:25 ð44Þ with the GMDH-GEP model (R ¼ 0.93 and MAPE ¼ 15.77).
qffiffiffiffiffiffiffiffiffiffiffiffi In addition, for evaluating the VL� , the GMDH-GEP produced
þ 0:06093ðVH� Þ14 ðVH� Þ11 better estimation (RMSE ¼ 0.623 and MAPE ¼ 18.92) than the
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
GMDH approach (RMSE ¼ 0.358 and MAPE ¼ 13.037).
ðVH� Þ22 ¼ GEPððVH� Þ11 ; ðVH� Þ16 Þ ¼ 1:35 þ 0:004854 ðVH� Þ11 ðVH� Þ16
rffiq
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi R, RMSE, and BIAS values given by the GMDH network for
qffiffiffiffiffiffiffiffiffiffiffiffi the VR� prediction indicated a relative superior to the GMDH-
� 1 1
þ 0:08482ðVH Þ1 ðVH� Þ1 0:0179 ðVH� Þ11 GEP model in this case. Meantime, in the VV� prediction,
ð45Þ statistical error parameters obtained by the GMDH-GEP
approach demonstrated more efficient performance (MAPE ¼
ðVH Þ31 ¼ GEPððVH� Þ21 ; ðVH� Þ22 Þ ¼ 0:000708 16.92 and RMSE ¼ 0.215) than the GMDH network (RMSE ¼
rq
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffi 0.237 and MAPE ¼ 21.65).
þ 0:03305 ðVH� Þ22 0:3318ðVH� Þ21 ðVH� Þ22 ð46Þ Performance of the GMDH-GEP indicated notably accurate
prediction of the longitudinal scour rates (RMSE ¼ 1.244 and
þ 0:5176ðVH� Þ21 ððVH� Þ21 þ ðVH� Þ22 Þ
MAPE ¼ 15.77) compared with those resulted using the GEP
The proposed structure of GMDH-GEP network contain- model (RMSE ¼ 1.67 and MAPE ¼ 55.41). In the VV� predic-
ing three selective neurons in the first layer, 2 selective tion, the GMDH-GEP and GEP have relatively same efficiency
neurons in the second layer, and a neuron in output layer in terms of R, MAPE, and SI values. Meantime, the GEP mod-
(or 3-2-1) for predicting the VH� was represented in Figure 7. els proposed to evaluate the scour rates at left and right of the
Analytical equations obtained by the GMDH-GEP model for pipeline are relatively superior to the proposed GMDH-GEP
predicting the VV� ,VR� , and VL� were given in Appendix C. From methods.
the GMDH-GEP network, a quadratic polynomial with tree Statistical results of the ANN model showed considerably
structure has been generated by development of the GEP higher error for the VH� prediction (RMSE ¼ 1.154, MAPE ¼-
model for each neuron of the GMDH network. 51.9, and BIAS ¼ 0.324) in comparison with the GMDH-GEP
model (RMSE ¼ 1.144, MAPE ¼ 15.77, and BIAS ¼ 0.158). In
the VL� prediction, R and RMSE values obtained by the
GMDH-GEP model indicated more accurate performance
4.2. Comparison of the GMDH-GEP with GMDH, GEP,
than the FFBP-NN approach (R ¼ 0.932 and RMSE ¼ 0.407).
and ANN models
Regarding the MAPE values, the proposed GMDH-GEP and
In this section, statistical results of the GMDH-GEP FFBP-NN models have same performance in this case. For
approach have been compared with those obtained using prediction of the VL� , statistical error parameters of BIAS
the proposed GMDH, GEP, and ANNs models. Quantitative and SI demonstrated that FFBP-NN model produced relatively
performances of the proposed approaches were summarized good results in comparison with the GMDH-GEP model. Stat-
in the Table 6. istical results given in Table 8 showed which FFBP-NN pro-
Performances of the GMDH model for evaluation of the vided relatively perfect performances for the VR� and VV�
three-dimensional scour rates demonstrated lower accuracy evaluations, compared to the GMDH-GEP model. Scatter
in comparison with the GMDH-GEP. For VH� prediction, plots of VH� , VL� , VR� , and VV� for the proposed models in testing
proposed GMDH technique gave remarkable inaccurate stage have been illustrated in Figures 8–11, respectively.
10 M. NAJAFZADEH AND F. SABERI-MOVAHED

Figure 8. Performances of the proposed models for prediction of the VH� .


Figure 10. Performances of the proposed models for prediction of the VR� .

4.3. Evaluation of empirical equation and regression


models square, � in these equations. To prevent the complexity of
analytical equations, function sets of exp, sin, cos were
Cheng et al. (2014) proposed an empirical equation for not used. Application of the function sets obtained group-
prediction of longitudinal scour rate as follows: dimensionless equations for prediction of the scour rates.
� e � pffiffiffiffiffiffi Lack of validation for Eq. (47) is related to the nonlinear
VH� ¼ 11:3 1 ð1 þ sinðaÞÞ h5=3
w KC ð47Þ
D regression technique and it cannot produce physical features
Eq. (47) was used for evaluating the VH� parameter. From of scour rates. Meantime, new traditional equations-based
Table 6, performances of Eq. (47) produced higher error of regression models to predict three-dimensional scour rates
the scour rate prediction in terms of RMSE ¼ 1.99 and MAPE ¼ were extracted using the same datasets utilized earlier for
47.58 than those obtained using the GMDH-GEP model developing the GMDH-GEP network. Applying the least
(RMSE ¼ 1.244 and MAPE ¼ 15.77). Other statistical error squares sense to those datasets produced equations as,
parameters given by Eq. (47) were presented in Table 8. VH� ¼ 256:723ðhw Þ1:313 KC 1:054
ðe=DÞ0:02 ð1 þ sin aÞ 0:103

A grouped Eqs. (41)–(46) represented the VH� values based ð48Þ


on inherent properties of the GMDH network and tree struc- :0:579 0:37 0:308 1:288
ture of the gene-expression programming. VL� ¼ 1:756ðhw Þ KC ðe=DÞ ð1 þ sin aÞ ð49Þ
0:409 0:69 0:464 1:356
Thorough performing the GEP in each neuron of the VR� ¼ 0:468ðhw Þ KC ðe=DÞ ð1 þ sin aÞ ð50Þ
GMDH, selection of function sets were adjusted based on :0:494 0:397 0:355 1:33
VV� ¼ 0:661ðhw Þ KC ðe=DÞ ð1 þ sin aÞ ð51Þ
mathematical shape of Eq. (47) and polynomial neuron of
the GMDH (Eq. 15). There are simple function sets of +, −,

Figure 9. Performances of the proposed models for prediction of the VL� . Figure 11. Performances of the proposed models for prediction of the VV� .
MARINE GEORESOURCES & GEOTECHNOLOGY 11

Performances of the proposed nonlinear regression equa- (RMSE ¼ 1.33, MAPE ¼ 38.7, and SI ¼ 0.461) variable is the
tions were given in Table 8. Qualitative statistical results of most effective parameter on the scour rates at left of pipeline.
Eqs. (48)–(51) indicated remarkably higher error of the scour Finally, hw was assigned as most effective variable (RMSE ¼
rate predictions in comparison with the proposed GMDH- 0.762, MAPE ¼ 51.63, and SI ¼ 0.65) on the vertical scour rate
GEP models. Eq. (48) provided the longitudinal scour rate whereas e/D has the least influence on the VV� . Statistical error
around pipeline with RMSE ¼ 2.72 and MAPE ¼ 99.97, com- parameters yielded from sensitivity analysis are given in
pared to Eqs. (41)–(46). From Figure 8, testing results of the Table 9.
proposed models indicated that Eqs. (47)–(48) based on
regression technique produced considerable under (or over)
prediction of longitudinal scour rate compared with other 6. Parametric study
approaches. Also, Figure 9 illustrated that Eq. (49) cannot cap-
ture efficient performance for the scour rate prediction at left In this section, variations of three-dimensional scour rates ver-
of the pipeline in comparison with the proposed GMDH-GEP sus input parameters are investigated. Scour rate values ðVH� Þ
approach. From Table 8 and Figure 10, Eq. (50) produced more predicted by Eq. (47) for a ¼ 0� 45� were compared with
considerably over prediction (MAPE ¼ 2409.7 and BIAS ¼ experimental data sets. Cheng et al. (2014) used such compar-
63.85) of the scour rates at right of the pipeline than those isons for highlighting efficiency of the proposed equation. The
obtained using the GMDH-GEP model. In addition, first comparison is related to the a ¼ 0� and KC ¼ 8.7. In this
Figure 11 showed that Eq. (51) gave the vertical scour rates with range, Eq. (47) has under-predicted the VH� for e/D between
more precise prediction than other regression methods. 0.2 and 0.4. BIAS error obtained by Eq. (47) is −1.611. It is
clear that the GMDH-GEP provided the VH� more efficient
results (BIAS ¼ −0.145) for a ¼ 0� and KC ¼ 8.7 than
Eq. (47). From Figure 12, GMDH-GEP indicated that
5. Sensitivity analysis
variations of the VH� versus e/D has gone through an upward
To find the quantitatively relative significance of each input trend, whereas Eq. (47) failed to predict VH� . For a ¼ 0� and
parameter on the three-dimensional scour rates, the GMDH- KC ¼ 15.8, Figure 13 indicates that VH� has decreased from
GEP model was utilized to conduct a sensitivity analysis. about 1.75 in e/D ¼ 0.1 to 1.5 in e/D ¼ 0.2 then VH� has
The analysis are performed such that, one parameter of increased, ending at roughly 4.5 in e/D ¼ 0.4. And one more,
Eqs. (2)–(5) is removed each time to evaluate the effect of that Eq. (47) remarkably over-predicted the VH� for e/D values
particular input on output. For Eq. (2), results of the analysis between 0.1 and 0.3. GMDH-GEP model produced higher
demonstrated which KC (R ¼ −0.786, BIAS ¼ 0.144, and SI ¼ accuracy of scour rates prediction (BIAS ¼ 0.119), compared
0.588) is the most effective parameter on the VH� whereas the e/ to that resulted by Eq. (47) (BIAS ¼ 1.89).
D (R ¼ 0.252, BIAS ¼ 0.0526, and SI ¼ 0.45) has the least From Figure 14, it can be said that GMDH-GEP could not
influence on the dimensionless longitudinal scour rate around present a precise prediction of VH� for e/D ¼ 0.1 whereas per-
pipeline. Other effective input parameters on the VH� were missible degree of precision was seen for e/D between 0.2 and
fixed as hw and sin α, respectively. 0.4. Furthermore, the GMDH-GEP network had an acceptable
In addition, for Eq. (3), hw (RMSE ¼ 1.53, MAPE ¼ 42.15, error value of BIAS (0.811) in comparison with Eq. (47)
and SI ¼ 0.518) was determined as most effective parameter (BIAS ¼ 2.36).
on the VR� and while sin α with RMSE ¼ 0.845, MAPE ¼ 18.61, Variations of the VH� versus e/D for a ¼ 15� and KC ¼ 15.8
and SI ¼ 0.252 has the least effects on the scour rate at right of were shown in Figure 15. For e/D ¼ 0.2–0.3, Eq. (47) over-
pipeline. predicted the VH� values whereas under-prediction for e/D of
Performance of sensitivity analysis for Eq. (4) indicated that 0.4 was seen. From Figure 15, it can be said that the GMDH-
KC parameter has the most influence on the VL� in terms of GEP model indicated existence of physical insights between
RMSE ¼ 1.707, MAPE ¼ 58.53, and SI ¼ 0.59 whereas sin α

Table 9. Statistical results of the sensitivity analysis.


Function R RMSE MAPE BIAS SI
VH� ¼ f ðKC; e=D; sin aÞ −0.061 1.49 62.08 −0.275 0.577
VH� ¼ f ðhw ; e=D; sin aÞ −0.786 1.49 58.15 0.144 0.588
VH� ¼ f ðhw ; KC; sin aÞ 0.252 1.14 43.83 0.0526 0.45
VH� ¼ f ðhw ; KC; e=DÞ 0.007 1.24 49.46 0.1095 0.487
VR� ¼ f ðKC; e=D; sin aÞ 0.95 1.53 42.15 −0.128 0.518
VR� ¼ f ðhw ; e=D; sin aÞ 0.905 1.028 23.47 −0.191 0.342
� 0.77
VR ¼ f ðhw ; KC; sin aÞ 1.24 38.191 0.142 0.418
VR� ¼ f ðhw ; KC; e=DÞ 0.46 0.845 18.61 −0.401 0.252
VL� ¼ f ðKC; e=D; sin aÞ 0.852 1.56 54.37 0.141 0.553
VL� ¼ f ðhw ; e=D; sin aÞ 0.952 1.707 58.53 −0.403 0.59
VL� ¼ f ðhw ; KC; sin aÞ 0.822 1.56 49.06 −0.647 0.507
VL� ¼ f ðhw ; KC; e=DÞ 0.645 1.33 38.7 −0.328 0.461
VV� ¼ f ðKC; e=D; sin aÞ 0.932 0.762 51.63 0.0178 0.65
VV� ¼ f ðhw ; e=D; sin aÞ 0.948 0.468 38.33 −0.044 0.399
VV� ¼ f ðhw ; KC; sin aÞ 0.659 0.463 28.5 −0.0095 0.396
VV� ¼ f ðhw ; KC; e=DÞ 0.498 0.498 41.74 0.036 0.425 Figure 12. Variations of VH� parameter versus e/D for KC ¼ 8.7 and a ¼ 0� .
12 M. NAJAFZADEH AND F. SABERI-MOVAHED

Figure 13. Variations of VH� parameter versus e/D for KC ¼ 15.8 and a ¼ 0� . Figure 15. Variations of VH� parameter versus e/D for KC ¼ 15.8 and a ¼ 15� .

VH� and e/D parameters. Eq. (47) produced lower accuracy of


scour rates prediction in term of BIAS error (1.28) than that
obtained using the GMDH-GEP (BIAS ¼ 0.213). GMDH-
GEP showed variations of VH� versus e/D which had an upward
trend whereas Eq. (47) had an opposite trend.
Figure 16 illustrated efficiency of the GMDH-GEP network
and Eq. (47) to characterize physical behavior of VH� and dif-
ferent values of e/D for a ¼ 15� and KC ¼ 18. Figure 16 indi-
cated that GMDH-GEP predicted erratic trend, decreasing
from roughly 3.3 in e/D ¼ 0.1 to approximately 2.25 in e/D ¼
0.2, then increasing to about 4.7 in. Next, VH� has plummeted
to about 2.5 in e/D ¼ 0.4. Moreover, Eq. (47) was indicative of
downward trend which produced VH� with relatively high
over-prediction for all the ratio of e/D except 0.3. The
GMDH-GEP network provided relatively better agreement
Figure 16. Variations of VH� parameter versus e/D for KC ¼ 18 and a ¼ 15� .
with observed data sets (BIAS ¼ −0.0136) than Eq. (47)
(BIAS ¼ 0.551).
For a ¼ 30� and KC ¼ 15.8, Figure 17 indicated that (BIAS ¼ 0.777). From graphical results of GMDH-GEP, VH�
Eq. (47) gave over-prediction of VH� values for e/D ¼ 0.1–0.3, decreased from 3.5 in e/D ¼ 0.1 to 2.5 in e/D ¼ 0.3 after that
whereas it under-predicted VH� variable for e/D ¼ 0.4. it had a slight increase, reaching roughly 2.7 in e/D ¼ 0.4.
Results of performance shown which the GMDH-GEP Figure 18 demonstrated that maximum over-prediction
model provided lower error of scour rates prediction produced by Eq. (47) was seen in e/D ¼ 0.1 for a ¼ 30� and
(BIAS ¼ −0.188), compared to that obtained by Eq. (47) KC ¼ 18 whereas a puny amount of under prediction was
observed by the GMDH-GEP. Eq. (47) gave lower accuracy

Figure 14. Variations of VH� parameter versus e/D for KC ¼ 18 and a ¼ 0� . Figure 17. Variations of VH� parameter versus e/D for KC ¼ 15.8 and a ¼ 30� .
MARINE GEORESOURCES & GEOTECHNOLOGY 13

Figure 18. Variations of VH� parameter versus e/D for KC ¼ 18 and a ¼ 30� .
Figure 20. Variations of VR� parameter with KC for a ¼ 0� .

(BIAS ¼ 2.1) of the VH� prediction than that yielded by the


GMDH-GEP network (BIAS ¼ 0.337). Moreover, for e/D of
0.2 and 0.3, Eq. (47) had similar trend. GMDH-GEP predicted
VH� with permissible level of accuracy for e/D ¼ 0.2 whereas
GMDH-GEP had over prediction for e/D ¼ 0.3–0.4.
For a ¼ 45� and KC ¼ 15.8, Figure 19 indicates that VH� ,
estimated by the GMDH-GEP, has decreased from roughly
3.75 in e/D ¼ 0.1 to about 2.5 in e/D ¼ 0.2 and after that, pre-
dicted values of VH� had an upward trend, reaching approxi-
mately 3.25 in e/D ¼ 0.3. Eq. (47) over-predicted in e/
D ¼ 0.1–0.2. Conversely, better performance of Eq. (47) can
be seen for e/D ¼ 0.3. The GMDH-GEP model indicated sig-
nificantly lower error (BIAS ¼ 0.053) than Eq. (47)
(BIAS ¼ 1.44).
To obtain new contribution of this research, effects of KC
and e/D on the variations of scour rates for a ¼ 0� were inves-
tigated. Variations of VR� versus KC values for different e/D Figure 21. Variations of VL� parameter with KC for a ¼ 0� .
between 0.1 and 0.4 were illustrated in Figure 20. From
Figure 20, scour rate of the VR� increased with increase in Furthermore, variations of VR� versus e/D values for differ-
KC parameter from 8.7 to 18. For KC ¼ 8.7–18, VR� values ent KC between 8.7 and 18 were illustrated in Figure 23. From
increased with decrease in e/D parameter. It can be generally Figure 23, for a constant KC value, it can be said that the VR�
said that, for a constant value of KC, the less embedded ratio decreased with increase in e/D parameter from 0.1 to 0.4. It
is, the higher VR� will become. These behaviors were seen for should be noted that for a constant value of e/D, the higher
scour rates of the VL� and VV� in Figures 21 and 22. KC is, the higher VR� will become. Variations of the VL� and

Figure 19. Variations of VH� parameter versus e/D for KC ¼ 15.8 and a ¼ 45� . Figure 22. Variations of VV� parameter with KC for a ¼ 0� .
14 M. NAJAFZADEH AND F. SABERI-MOVAHED

produce more accurate general patterns among output para-


meters and independent ones rather than applying Eq. (47),
as an empirical equation. The GMDH-GEP model provided
perfectly physical insights of scour rates with e/D and KC para-
meters in terms of design curves which can be useful for prac-
tical applications.

7. Conclusion
In this paper, the GMDH-GEP network was applied to predict
the three-dimensional scour rates below pipeline due to waves.
Development of the GMDH-GEP model was conducted using
four nondimensional input parameters including Shields para-
meter under waves, Keulegan–Carpenter number, angle of
attack, and ratio of embedment depth to pipe diameter.
Performance of training stage indicated that the GMDH-
Figure 23. Variations of VR� parameter with embedment depth for a ¼ 0� . GEP network predicted the longitudinal scour rate with MAPE
of 0.33. Error functions in terms of MAPE (0.269) and RMSE
(RMSE ¼ 0.31) produced the scour rate at left of pipeline with
relatively same accuracy as statistical error parameters
obtained from the VH� . In addition, R and SI indicated
relatively same accuracy for prediction of VV� and VR� .
In the testing stages, the proposed GMDH-GEP networks
were successful in prediction of the longitudinal scour rate,
scour rates at right and left of pipeline, and vertical scour rates.
For instance, the proposed GMDH-GEP model yielded an
accurate longitudinal scour rate (RMSE ¼ 1.244 and MAPE ¼
2.47). Performances of Eq. (47) provided the VH� parameter
with higher error of MAPE (5.07) than the VH� calculated by
the GMDH-GEP network (MAPE ¼ 2.47).
Statistical results of the proposed methods showed that the
GMDH-GEP network predicted longitudinal scour rates with
striking superiority to the ANN, GEP, GMDH, and empirical
equations. In addition, for other scour rates, GMDH-GEP
models had relative good performances, compared to the
others approaches.
Figure 24. Variations of VL� parameter with embedment depth for a ¼ 0� .
Effects of KC and e/D parameters on the variations of three-
dimensional scour rates have been evaluated. For different
VV� parameters versus e/D values for different KC values have ranges of e/D, analytical Eqs.(41)–(46) given by the GMDH-
been demonstrated in Figures 24 and 25, respectively. GEP network produced more accurate prediction of the VH�
From Figures 20–25, a good many design curve were pro- (BIAS ¼ −0.0136) for a ¼ 15� and KC ¼ 18, compared to
posed to estimate score rates below pipeline and additionally, Eq. (47) (BIAS ¼ 0.551). In addition, for a ¼ 45� and
graphical results indicated that use of GMDH-GEP could KC ¼ 15.8, the GMDH-GEP model demonstrated remarkably
lower error (BIAS ¼ 0.053) than Eq. (47) (BIAS ¼ 1.44).
Scour rates in the right and left of pipeline and vertical
scour rates generally decreased with increase in e/D parameter
for the KC ¼ 8.7 and 18. Moreover, scour rates of VL� increased
with increase in KC values for e/D ¼ 0.1–0.4. Performance of
the GMDH-GEP model indicated that physical insights
between variations of VR� and VV� versus KC values were simi-
lar to those represented between VL� and KC.
Applying the gene-expression programming in form of tree
structures for developing the GMDH network was proven and
GMDH-GEP models were used successfully as a powerful soft
computing tool for predicting the free span expansion rates in
three-dimensional directions below pipelines due to waves. In
case of practical applications, design curves given in different
ranges of input variables can be useful to evaluate the scour rates
Figure 25. Variations of VV� parameter with embedment depth for a ¼ 0� . for engineering applications in marine environments.
MARINE GEORESOURCES & GEOTECHNOLOGY 15

References Kalantary, F., H. Ardalan, and N. Nariman-Zadeh. 2009. An investigation


on the Su-NSPT correlation using GMDH type neural networks and
Abdel-Aal, R. E., and E.-S. M. El-Alfy. 2009. Constructing optimal edu- genetic algorithms. Eng. Geol. 104 (1–2):144–55. doi:10.1016/j.
cational tests using GMDH-based item ranking and selection. Neuro- enggeo.2008.09.006.
computing 72:1184–97. doi:10.1016/j.neucom.2008.02.004. Koza, J. R. 1992. Genetic programming: On the programming of computers
Abdolrahimi, S., B. Nasernejad, and G. Pazuki, 2014. Prediction of par- by means of natural selection. Cambridge, MA: The MIT Press.
tition coefficients of alkaloids in ionic liquids based aqueous biphasic Liang, D. F., L. Cheng, and F. Li. 2005. Numerical modelling of scour
systems using hybrid group method of data handling (GMDH) neural below a pipeline in currents. Part II: Scour simulation. Coast. Eng.
network. J. Mol. Liq. 191:79–84. doi:10.1016/j.molliq.2013.11.033. 52 (1):43–62.
Astakhov, V. P., and V. V. Glitsky. 2005. Tool life testing in gun drilling: An Lin, J.-S. 2012. A novel design of wafer yield model for semiconductor
application of the group method of data handling (GMDH). Int. J. Mach. using a GMDH polynomial and principal component analysis. Expert.
Tool. Manu. 45:509–17. doi:10.1016/j.ijmachtools.2004.09.003. Syst. Appl. 39 (5):6665–71. doi:10.1016/j.eswa.2011.09.146.
Azamathulla, H., and A. Ghani. 2010. Genetic programming to predict Masoumi Shahr-Babak, M., M. J. Khanjani, and K. Qaderi. 2016. Uplift
river pipeline scour. J. Pipeline Syst. Eng. Pract. 1 (3):127–32. capacity prediction of suction caisson in clay using a hybrid
doi:10.1061/(asce)ps.1949-1204.0000060. intelligence method (GMDH-HS). Appl. Ocean Res. 59:408–16.
Azamathulla, H. M., A. Guven, and Y. K. Demir. 2011. Linear genetic doi:10.1016/j.apor.2016.07.005.
programming to scour below submerged pipeline. Ocean Eng. Mehrara, M., A. Moeini, M. Ahrari, and A. Erfanifard. 2009. Investigating
38:995–1000. doi:10.1016/j.oceaneng.2011.03.005. the efficiency in oil futures market based on GMDH approach. Expert
Azamathulla, H. M. and M. A. M. Yusoff. 2013. Soft computing for Syst. Appl. 36 (4):7479–83.
prediction of River pipeline scour depth. Neural. Compt. Appl. Naghashnejad, M., N. Amanifard, and H. M. Deylami. 2014. A predictive
23 (7–8):2465–69. doi:10.1007/s00521-012-1205-x. model based on a 3-D computational approach for film cooling
Azamathulla, H. M. and N. A. Zakaria. 2011. Prediction of scour bellow effectiveness over a flat plate using GMDH-type neural networks. Heat
submerged pipeline crossing a river using ANN. Water. Sci. Technol. and Mass Transfer 50 (1):139–49.
23 (7–8):2225–30. doi:10.2166/wst.2011.459. Najafzadeh, M., G. A. Barani, and H. M. Azamathulla. 2014a. Prediction
Bernetti, R., R. Bruschi, V. Valentini, and M. Venturi. 1990. Pipelines of pipeline scour depth in clear-water and live-bed conditions using
placed on erodible seabeds. Proc. 9th International Conference on Off- group method of data handling. Neural Comput. Appl. 24 (3–4):
shore Mechanics and Artic Engineering, ASME, Houston, TX, vol. V, 629–35. doi:10.1007/s00521-012-1258-x.
pp. 155–64. Najafzadeh, M., G. A. Barani, and M. R. Hessami-Kermani. 2014b.
Cheng, L., K. Yeow, Z. Zang, and F. Li. 2014. 3D scour below pipelines Estimation of pipeline scour due to waves by the group method of data
under waves and combined waves and currents. Coastal Eng. handling. J. Pipeline Syst. Eng. Pract., ASCE 5 (3):06014002.
83:137–49. doi:10.1016/j.coastaleng.2013.10.006. doi:10.1061/(asce)ps.1949-1204.0000171.
Cheng, L., K. Yeow, Z. Zhang, and B. Teng. 2009. Three-dimensional Onwubolu, G. C. 2008. Design of hybrid differential evolution and group
scour below offshore pipelines in steady currents. Coastal Eng. method in data handling networks for modeling and prediction. Inf.
56 (5–6):577–90. doi:10.1016/j.coastaleng.2008.12.004. Sci. 178:3618–34. doi:10.1016/j.ins.2008.05.013.
Chiew, Y. M. 1990. Mechanics of local scour around submarine pipelines. Onwubolu, C. G. 2009. Hybrid self-organizing modeling systems. Berlin,
J. Hydraul. Eng. 116 (4):515–29. doi:10.1061/(asce)0733-9429(1990) Heidelberg: Springer-Verlag, 7–11.
116:4(515). Palmer, A. C. 1996. A flaw in the conventional approach to stability
Etemad-Shahidi, A., R. Yasa, and M. H. Kazeminezhad. 2011. Prediction design of pipelines. Proc. Offshore Pipeline Technology Conf.,
of wave-induced scour depth under submarine pipelines using Amsterdam, the Netherlands.
machine learning approach. Appl. Ocean Res. 33:54–59. doi:10.1016/ Sakaguchi, A. and T. Yamamoto. 2000. A GMDH network using back
j.apor.2010.11.002. propagation and its application to a controller design. J. IEEE 4:2691–97.
Farlow, S. J. (Ed.). 1984. Self-organizing method in modeling: GMDH type Sanchez, E., T. Shibata, and L. A. Zadeh. 1997. Genetic algorithms and
algorithm. New York: Marcel Dekker Inc. fuzzy logic systems. River Edge, NJ, USA: World Scientific.
Ferreira, C. 2001. Gene expression programming: A new adaptive algor- Shaghaghi, S., H. Bonakdari, A. Gholami, I. Ebtehaj, and M.
ithm for solving problems (PDF). Complex Syst. 13 (2):87–129. Zeinolabedini. 2017. Comparative analysis of GMDH neural network
Ferreira, C. 2006. Gene-expression programming: Mathematical Modeling based on genetic algorithm and particle swarm optimization in stable
by an Artificial Intelligence. Berlin: Heidelberg, New York: Springer, 21. channel design. Appl. Math. Comput. 313:271–86. doi:10.1016/j.
Haghiabi, A. H. 2016. Prediction of river pipeline scour depth using mul- amc.2017.06.012.
tivariate adaptive regression splines. J. Pipeline Syst. Eng. Pract. 8 Soulsby, R. 1997. Dynamics of marine sands: a manual for practical appli-
(1):04016015. doi:10.1061/(asce)ps.1949-1204.0000248. cations. London: Thomas Telford.
Hansen, E. A. 1992. Scour below pipelines and cables. A Simple model. Srinivasan, D. 2008. Energy demand prediction using GMDH networks.
11th Offshore Mechanics and Artic Engineering Conference. Pipeline Neuro Comput. 72 (1–3):625–29. doi:10.1016/j.neucom.2008.08.006.
Technol., ASME, Calgary, Canada, vol. V-A, pp. 133–38. Sumer, B. M., and J. Fredsøe. 2002. The mechanics of scour in the marine
Hornik, K., M. Stichcombe, and H. White. 1989. Multi layer feed forward environment. Singapore: World Scientific.
networks are universal approximation. Neural Netw. 2:359–66. Witczak, M., J. Korbicz, M. Mrugalski, and R. Patton. 2006. A GMDH
doi:10.1016/0893-6080(89)90020-8. neural network-based approach to robust fault diagnosis: Application
Iba, H., and H. de Garis, 1996. Extending genetic programming with to the DAMADICS benchmark problem. Control Eng. Pract.
recombi-native guidance. In Advances in genetic programming, ed. 14 (6):671–83. doi:10.1016/j.conengprac.2005.04.007.
P. Angeline, K. Kinnear, vol. 2. Cambridge: MIT Press. Wu, Y., and Y. M. Chiew. 2012. Three-dimensional scour at submarine
Ivakhnenko, A. G. 1971. Polynomial theory of complex systems. pipelines. J. Hydraul. Eng. 138 (9):788–95. doi:10.1061/(asce)hy.1943-
IEEE Trans. Syst. Man, Cybern. 4:364–78. doi:10.1109/tsmc.1971. 7900.0000583.
4308320. Wu, Y., and Y. M. Chiew. 2013. Mechanics of three-dimensional pipeline
Ivakhnenko, A. G., and G. A. Ivakhnenko. 2000. Problems of further scour in unidirectional current. J. Pipeline Syst. Eng. Pract. 4 (1):3–10.
development of the group method of data handling algorithms, Part doi:10.1061/(asce)ps.1949-1204.0000118.
1. Pattern Recognit. Image Anal. 110:187–94. Yasa, R., and A. Etemad-Shahidi. 2014. Classification and regression trees
Jamali, A., N. Nariman-Zadeh, A. Darvizeh, A. Masoumi, and S. approach for predicting current-induced scour depth under pipelines.
Hamrang. 2009. Multi-objective evolutionary optimization of J. Offshore Mech. Arct. Eng. 136 (1):011702. doi:10.1115/1.4025654.
polynomial neural networks for modelling and prediction of explosive Zang, Z., L. Cheng, M. Zhao, D. Liang, and B. Teng. 2009. A numerical
cutting process. Eng. Appl. Artif. Intel. 22:676–87. doi:10.1016/j. model for onset of scour below offshore pipelines. Coast. Eng. 56 (4):
engappai.2008.11.005. 458–66. doi:10.1016/j.coastaleng.2008.10.001.
16 M. NAJAFZADEH AND F. SABERI-MOVAHED

�2 �1 �1
Zanganeh, M., A. Yeganeh-Bakhtiary, and R. Bakhtyar. 2011. Combined VR� 1 ¼ 1:5311 þ 0:4436 VR� 3 þ1:3804 VR� 4
particle swarm optimization and fuzzy inference system model for esti- �1 �1 � �1 � 2
mation of current-induced scour beneath marine pipelines. J. Hyroinf. þ 0:48597 VR� 3 VR� 4 0:17021 VR� 3 ðB:4Þ
13 (3):558–73. doi:10.2166/hydro.2010.101. � �1 �2
0:36279 VR� 4
Appendix A. Experimental datasets used to �2 �1 �1
develop the GMDH-GEP networks VR� 3 ¼ 1:9386 0:64409 VR� 4 0:3767 VR� 5
�1 �1 � �1 �2
þ 1:0802 VR� 4 VR� 5 0:2303 VR� 4 ðB:5Þ
Table A. Datasets to predict the three-dimensional scour rates. � �1 �2
e Alpha H T Uw VH VR VL Vv 0:340818 VR� 5
(mm) (°) (m) (s) (m/s) (mm/s) (mm/s) (mm/s) (mm/s)
5 0 0.13 1.5 0.29 3.6 3.4 3.5 1.6776 �3 �2 �2
10 0 0.13 1.5 0.29 4.8 2.8 3.8 2.2368 VR� 1 ¼ 0:152 þ 1:3567 VR� 1 0:5281 VR� 3
15 0 0.13 1.5 0.29 2.6 2.6 2.6 1.2116 �2 �2
20 0 0.13 1.5 0.29 2 1.8 1.9 0.932 0:4622 VR� 1 VR� 3 ðB:6Þ
5 0 0.17 1.8 0.44 6.3 5.1 5.7 2.9358 � �2 �2 � �2 � 2
10 0 0.17 1.8 0.44 5.4 4.4 4.9 2.5164 þ 0:11302 VR� 1 þ0:38561 VR� 3
15 0 0.17 1.8 0.44 4.6 4 4.3 2.1436
20 0 0.17 1.8 0.44 3.7 2.8 3.2 1.7242
5 0 0.15 2 0.45 6.5 5.7 6.1 3.029 for VL� prediction:
10 0 0.15 2 0.45 5.3 5.5 5.4 2.4698 �1
15 0 0.15 2 0.45 4.7 4.7 4.7 2.1902 VL� 3 ¼ 2:995 þ 6:034ðe=DÞ þ 1:9983 sin a
20 0 0.15 2 0.45 3 3.7 3.4 1.398
15 15 0.13 1.5 0.29 1.6 2.7 2.2 0.7456 þ 0:80296ðe=DÞ sin a 19:887ðe=DÞ2 6:9993ðsin aÞ2
20 15 0.13 1.5 0.29 1.6 1.5 1.5 0.7456
5 15 0.17 1.8 0.44 5.2 5.6 5.4 2.4232 ðB:7Þ
10 15 0.17 1.8 0.44 4.1 4.5 4.3 1.9106 �
� 1
15 15 0.17 1.8 0.44 4 4.4 4.2 1.864 VL 4 ¼ 18:694 þ 180:822ðhw Þ 4:0136ðe=DÞ
20 15 0.17 1.8 0.44 2.6 3.1 2.8 1.2116
10 15 0.15 2 0.45 4.4 5.2 4.8 2.0504 17:664ðe=DÞhw 344:9063ðhw Þ2 þ9:1463ðe=DÞ2
15 15 0.15 2 0.45 3.7 4.3 4 1.7242
20 15 0.15 2 0.45 3.4 3 3.2 1.5844
ðB:8Þ
5 30 0.13 1.5 0.29 2.2 3 2.6 1.0252 �
� 1
10 30 0.13 1.5 0.29 2.2 1.4 1.8 1.0252 VL 6 ¼ 11:0709 þ 108:612hw þ 3:243 sin a
5 30 0.17 1.8 0.44 3.9 4 3.9 1.8174
10 30 0.17 1.8 0.44 3.2 3.6 3.4 1.4912 12:249hw sin a 201:248ðhw Þ2 3:1565ðsin aÞ2
15 30 0.17 1.8 0.44 2.7 3.1 2.9 1.2582 ðB:9Þ
20 30 0.17 1.8 0.44 2.3 2.6 2.5 1.0718
5 30 0.15 2 0.45 3.4 4.1 3.8 1.5844 �
� 2

� 1

� 1
10 30 0.15 2 0.45 3.1 3.4 3.2 1.4446 VL 1 ¼ 1:5311 þ 0:4436 VL 3 þ1:3804 VL 4
15 30 0.15 2 0.45 2.3 3.1 2.7 1.0718 �1 �1 � �1 � 2
20 30 0.15 2 0.45 2.3 2.3 2.3 1.0718 0:48597 VL� 3 VL� 4 0:1702 VL� 3 ðB:10Þ
5 45 0.13 1.5 0.29 1.9 1.7 1.8 0.8854 �
10 45 0.13 1.5 0.29 1.8 1.4 1.6 0.8388 �1 � 2
5 45 0.17 1.8 0.44 2.8 2.7 2.7 1.3048 0:3627 VL� 4
10 45 0.17 1.8 0.44 2.5 1.9 2.2 1.165 �2 �1 �1
15 45 0.17 1.8 0.44 2 1.6 1.8 0.932 VL� 2 ¼ 2:0854 1:76041 VL� 3 þ0:4765 VL� 6
5 45 0.15 2 0.45 2.4 3.1 2.8 1.1184 �
10 45 0.15 2 0.45 2.3 2.3 2.3 1.0718 �1 �1 �1 �2
0:4721 VL� 3 VL� 6 þ0:2588 VL� 3 ðB:11Þ
� �1 �2
Appendix B. Proposed quadratic polynomial 0:18151 VL� 6
neurons for evaluation of the VR� , VL� , and VV� . �3 �2 �2
VL� 1 ¼ 0:3323 þ 0:6193 VL� 1 þ0:619805 VL� 2
for VR� prediction: �2 �2 � �2 �2
�1 0:21282 VL� 1 VL� 2 þ0:14147 VL� 1 ðB:12Þ
VR� 3 ¼ 2:9956 þ 6:034ðe=DÞ þ 1:9983 sin a � �2 �2
þ 0:8029ðe=DÞ sin a 19:887ðe=DÞ2 6:999ðsin aÞ2 þ 0:032 VL� 2

ðB:1Þ and for VV� prediction:


�1 �1
VR� 4 ¼ 18:694 þ 180:822ðhw Þ 4:0136ðe=DÞ VV� 3 ¼ 1:8779 þ 1:7417ðe=DÞ 1:3569 sin a
2 2
17:664ðe=DÞhw 344:906ðhw Þ þ9:1463ðe=DÞ þ 1:686ðe=DÞ sin a 9:8302ðe=DÞ2 0:5928ðsin aÞ2
ðB:2Þ ðB:13Þ
�1 �
� 1
VR� 5 ¼ 0:82139 þ 0:3897KC þ 2:658 sin a VV 5 ¼ 0:37082 þ 0:08642ðKCÞ þ 0:05924ðsin aÞ
0:1717KC sin a 0:00812ðKCÞ2 3:3502ðsin aÞ2 0:071992ðKCÞ sin a þ 0:000153ðKCÞ2 ðB:14Þ
ðB:3Þ 0:18602ðsin aÞ2
MARINE GEORESOURCES & GEOTECHNOLOGY 17

�1
VV� 6 ¼ 10:7108 þ 102:284hw þ 0:0145 sin a Selective polynomial neurons of the GMDH-GEP for
3:9073hw sin a 202:149ðhw Þ 2
0:07318ðsin aÞ 2 prediction of the VR� can be expressed as follows:
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ðB:15Þ ðVR� Þ11 ¼ GEPðsin a; e=DÞ ¼ 3:718 þ 132:3 sin ae=D

� 2

� 1 qffip
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðC:7Þ
VV 1 ¼ 0:22239 0:3747 VV 3 pffiffiffiffiffiffiffiffiffi
�1 �1 �1 76:55 e=D 50:27e=D sin a
þ 0:7576 VV� 5 þ0:6312 VV� 3 VV� 5 ðB:16Þ
� � �2 � � �2
� 1 � 1
þ 0:06323 VV 3 0:3543 VV 5 ðVR� Þ14 ¼ GEPðsin a; KCÞ ¼ 0:839 þ 0:2846ðsin a KCÞ
�2 �1 �1 pffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffi
VV� 2 ¼ 0:29061 0:32087 VV� 3 þ0:5938 VV� 6 � sin a þ 0:4889 sin a þ 0:4889KC2
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
�1 �1 � �1 �2 0:008815ð9:925 þ KCÞ KC sin a
þ 0:65344 VV� 3 VV� 6 þ0:033 VV� 3 ðB:17Þ
� � ðC:8Þ
�1 2
0:3056 VV� 6
ðVR� Þ16 ¼ GEPðsin a; hw Þ ¼ 9:772 55:62
�3 �2 �2 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffi pffiffiffiffiffi
VV� 1 ¼ 0:3084 29:826 VV� 1 þ31:3419 VV� 2 � sin a:hw þ 97:26hw 178:7 hw þ 8:928 sin a hw
�2 �2 � �2 � 2
þ 71:94046 VV� 1 VV� 2 21:9718 VV� 1 ðC:9Þ
ðB:18Þ
� � �2 ðVR� Þ21 ¼ GEPððVR� Þ11 ; ðVR� Þ14 Þ ¼ 1:456
2
50:1641 VV� 2 rffiq
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
qffiffiffiffiffiffiffiffiffiffiffi
1 1
1:37ðVR� Þ1 1:37 ðVR� Þ1 þ 0:03059 ðVR� Þ11
qffiffiffiffiffiffiffiffiffiffiffi
0:3392ðVR� Þ11 : ðVR� Þ11
Appendix C. Selective polynomials of the GMDH-
GEP networks for predicting the three-dimensional ðC:10Þ
scour rates
ðVR� Þ22 ¼ GEPððVR� Þ11 ; ðVR� Þ16 Þ ¼ ð0:1112ðVR� Þ11 0:4836Þ
Selective polynomial neurons of the GMDH-GEP for qffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
prediction of the VL� can be expressed as follows: � ðVR� Þ16 þ 0:05444 4:2677ðVR� Þ11 þ 0:00907
rqffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ðVL� Þ11 ¼ GEPðsin a; e=DÞ ¼ 4:896 3:068 sin a 43:068e=D
pffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffi � ðVR� Þ16 þ 0:6896
2:259 sin a þ 3:426 sin ae=D 8:667 e=D
ðC:11Þ
ðC:1Þ
ðVL� Þ14 ¼ GEPðsin a; KCÞ ¼ 2:282 þ 0:6139KC þ 1:318 sin a ðVR� Þ31 ¼ GEPððVR� Þ21 ; ðVR� Þ22 Þ ¼ ð 1:46 0:8748ððVR� Þ21 Þ2 Þ
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffi
þ 0:09676KC2 0:112 sin a þ KC � ðVR� Þ21 þ 0:0786 ðVR� Þ21 þ 4:97 ðVR� Þ22
ðC:2Þ 4:97ðVR� Þ22 þ 3:58
ðVL� Þ16 ¼ GEPðsin a; hw Þ ¼ ð 15:82hw ðC:12Þ
qpffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
þ 2:20Þ sin a 3:2 þ 242:5 hw þ 29:44hw Selective polynomial neurons of the GMDH-GEP for pre-
ðC:3Þ diction of the VV� can be expressed as follows:
pffi
ðVL� Þ21 ¼ GEPððVL� Þ11 ; ðVL� Þ14 Þ ¼87:91 þ 0:476ðVL� Þ11 ðVL� Þ14 ðVV� Þ11 ¼ GEPðsin a; e=DÞ ¼ 1:862 13:08 sin a
qffiffiffiffiffiffiffiffiffiffiffi qffip
ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffi ðC:13Þ
pffiffiffiffiffiffiffiffiffi
0:3538ðVL� Þ14 0:1222 ðVL� Þ11 þ 1:19ðVL� Þ14 � sin a 35:51 e=D þ 14:64 sin a
qffiffiffiffiffiffiffiffiffiffiffi
þ 1:190 ðVL� Þ11 ðVV� Þ14 ¼ GEPðsin a; KCÞ ¼ 0:346 þ 0:0904KC
pffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ðC:4Þ 0:9288 sin a þ 0:8384 sin a 0:005211 sin a:KC
ðVL� Þ22 ¼ GEPððVL� Þ11 ; ðVL� Þ16 Þ ¼ 0:889 0:09935 ðC:14Þ
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffi
� ðVL� Þ16 ðVL� Þ11 0:04159ðVL� Þ11 ðVL� Þ16 ðVL� Þ11 ðVV� Þ16 ¼ GEPðsin a; hw Þ ¼ 11
qffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffi ðC:15Þ
þ 0:2135ðVL� Þ11 ðVL� Þ16 4:041 sin ahw þ 97:26hw 200:8 hw

ðC:5Þ ðVV� Þ21 ¼ GEPððVV� Þ11 ; ðVV� Þ14 Þ ¼ 0:471


qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffi
ðVL Þ31 ¼ GEPððVL� Þ21 ; ðVL� Þ22 Þ þ 0:2466 ðVV� Þ11 ðVV� Þ14 þ 0:4336 ðVV� Þ14
¼ ð0:289 0:02969ððVL� Þ22 Þ2 ÞðVL� Þ21 ðC:6Þ qffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffi
qffiffiffiffiffiffiffiffiffiffiffi 0:2296ðVV� Þ11 ðVV� Þ14 ðVV� Þ14 0:3392ðVV� Þ11 ðVV� Þ11
þ 0:602 þ 0:26 ðVL� Þ22
ðC:16Þ
18 M. NAJAFZADEH AND F. SABERI-MOVAHED

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ðVV� Þ22 ¼ GEPððVV� Þ11 ; ðVV� Þ16 Þ ¼ 0:4783 þ 0:2452 ðVV� Þ31 ¼ GEPððVV� Þ21 ; ðVV� Þ22 Þ ¼ 0:1919 9:7522ðVV� Þ21
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffi
� ðVV� Þ11 ðVV� Þ16 0:224ðVV� Þ11 ðVV� Þ16 ðVV� Þ16 ðC:17Þ þ ð 15:98 ððVV� Þ22 ÞððVV� Þ21 Þ2 þ 0:3787 ðC:18Þ
qffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
þ 0:1046 ðVV� Þ16 � ðVV� Þ21 ðVV� Þ22 þ 0:03235

You might also like