Neural Comput & Applic

DOI 10.1007/s00521-015-2072-z

ORIGINAL ARTICLE

Developing a hybrid PSO–ANN model for estimating the ultimate

bearing capacity of rock-socketed piles
Danial Jahed Armaghani1 • Raja Shahrom Nizam Shah Bin Raja Shoib1 •

Koohyar Faizi2 • Ahmad Safuan A. Rashid1

Received: 17 May 2015 / Accepted: 5 October 2015

Ó The Natural Computing Applications Forum 2015

Abstract Rock-socketed piles are commonly used in systems. After constructing and modelling these systems,
foundations built in soft ground, and thus, their bearing selected performance indices including the coefficient of
capacity is a key issue of universal concern in research, determination (R2), root-mean-square error, variance
design and construction. The accurate prediction of the account for and total ranking were used to identify the best
ultimate bearing capacity (Qu) of rock-socketed piles is a models and compare the obtained results. This analysis
difficult task due to the uncertainty surrounding the various revealed that the hybrid PSO–ANN model offers a higher
factors that affect this capacity. This study was aimed at degree of accuracy compared to conventional ANN for
developing an artificial neural network (ANN) model, as predicting the Qu of rock-socketed piles. However, the
well as a hybrid model based on both particle swarm developed model would be most useful in the preliminary
optimisation (PSO) and ANN, with which to predict the Qu stages of pile design and should be used with caution.
of rock-socketed piles. PSO, a powerful population-based
algorithm used in solving continuous and discrete optimi- Keywords Rock-socketed pile
Ultimate bearing
sation problems, was here employed as a robust global capacity
ANN
PSO
Hybrid model
search algorithm to determine ANN weights and biases and
thereby improve model performance. To achieve the study
aims, 132 piles socketed in various rock types as part of the 1 Introduction
Klang Valley Mass Rapid Transit project, Malaysia, were
investigated. Based on previous related investigations, Pile foundations are utilised to transfer structural loads
parameters with the most influence on Qu were identified deep into the ground and thereby achieve construction
and utilised in the modelling procedure of the intelligent stability, with the proper estimation of pile bearing
capacity thus of prime importance when designing
geotechnical structures. Many experimental and theoretical
& Danial Jahed Armaghani methods previously developed for the prediction of pile
[email protected] capacity incorporate assumptions associated with the
Raja Shahrom Nizam Shah Bin Raja Shoib parameters that control the ultimate bearing capacity (Qu)
[email protected] [1, 2]. However, due to the complex behaviour of piles,
Koohyar Faizi almost none of the available methods/models provide
[email protected] accurate predictions [3], with many also site specific [4].
Ahmad Safuan A. Rashid The most reliable technique used to determine pile
[email protected] bearing capacity is the static load test (SLT), although the
1 performance of this test is time-consuming, expensive and
Department of Geotechnics and Transportation, Faculty of
Civil Engineering, Universiti Teknologi Malaysia, UTM difficult [5, 6]. One of the most innovative methods
Skudai, 81310 Johor, Malaysia developed for the prediction of pile bearing capacity is high
2
School of Civil Engineering, University of Birmingham, strain dynamic testing (HSDT), which is based on wave
Birmingham, UK propagation theory and implemented via the use of a pile

123
 Neural Comput & Applic

driving analyser (PDA). The HSDT procedure is typically demonstrated the feasibility of ANN in predicting the ABC
standardised based on American Standards Test Methods of pipe piles in sandy soil, with the influential parameters
[7]. In previous investigations, a close agreement was found considered for network construction including the effective
between bearing capacity values predicted by PDA and pile length and diameter, unit weight, soil cohesion and
those predicted by SLT [5]. Furthermore, PDA (HSDT) is internal friction angle, as well as standard penetration test
also faster and more economical compared to SLT [6]. (SPT) results. A gene expression programming (GEP)
However, in order to obtain a reliable result, multiple PDA model was introduced and proposed by Alkroosh and
tests are required for each construction project. As a result, Nikraz [29] in order to estimate the capacity of driven piles
decreasing the number of PDA tests required is highly in cohesive soil. In another study examining pile capacity
desirable since it would reduce total project cost. To this prediction, Momeni et al. [6] predicted the Qu of concrete
end, new techniques such as artificial intelligence (AI) have piles using a hybrid genetic algorithm (GA)-ANN model,
been developed that are able to predict pile bearing capacity successfully demonstrating the high reliability of the
with a higher degree of accuracy in comparison with developed GA-ANN model in predicting Qu.
existing methods, as well as solve complex engineering Although ANNs are able to capture the complex and
problems more rapidly (e.g. [8]). nonlinear relationship between pile bearing capacity and its
AI techniques such as artificial neural networks (ANNs) influencing factors, limitations remain, such as the slow
are commonly employed in solving geotechnical problems rate of learning and the risk of entrapment in local minima
(e.g. [9–14]). ANN and adaptive neuro-fuzzy inference [30]. As a result, the utilisation of optimisation algorithms
system (ANFIS) models were proposed for the estimation such as particle swarm optimisation (PSO) can significantly
of rock strength by Yagiz et al. [15] and Jahed Armaghani enhance ANN performance (e.g. [31]). PSO algorithms
et al. [16], respectively, while Singh et al. [17] reported the represent a powerful population-based stochastic approach
successful use of an ANFIS predictive model in estimating that can be implemented to adjust the weights and biases of
the Young’s modulus of rock. A genetic programming ANN in order to enhance its performance. Several
(GP) technique was developed to predict the backbreak researchers have reported the successful use of hybrid
produced by blasting in a study carried out by Shirani PSO–ANN models in solving engineering problems (e.g.
Faradonbeh et al. [18]. Ground vibration and flyrock [31, 32]).
occurrence were predicted via ANN and imperial com- In the present study, two intelligent systems, namely
petitive algorithm (ICA)-ANN models in studies conducted ANN and hybrid PSO–ANN, were developed to predict the
by Monjezi et al. [19] and Marto et al. [20], respectively. ultimate bearing capacity of rock-socketed piles. The
Ocak and Seker [21] proposed an ANN technique to solve developed models are discussed, and the best examples
the problem of surface settlement caused by tunnelling. A selected for use in engineering applications.
support vector regression (SVR) model was proposed by
Mahdevari et al. [22] for the prediction of the penetration
rate of a tunnel boring machine (TBM). 2 Case study and data collection
In the field of foundation engineering in particular, these
techniques have been extensively utilised. Chan et al. [23] The Klang Valley Mass Rapid Transit (KVMRT) is the
developed an ANN model as an alternative approach to the latest rail-based public transport system to be constructed
establishment of pile driving formulae; in this case, the by the Government of Malaysia aimed at alleviating traffic
network was trained using the pile set, the driving energy congestion in Kuala Lumpur. The first phase of the project
delivered to the pile and the elastic compression of both includes a 51-km rail alignment from Sungai Buloh to
pile and soil, with pile capacity the network output. Lee Kajang. As initially investigated, the project will require the
and Lee [24] reported the successful implementation of an construction of thousands of large-diameter bored piles to
ANN model in predicting the Qu of piles socketed in sandy support both the viaducts and station developments. These
soil. Abu-Kiefa [25] addressed the capability of the ANN structures are to be founded on a wide range of rock types,
technique for the prediction of the axial bearing capacity including granite, sandstone, limestone, schist and phyllite.
(ABC) of driven piles in cohesionless soils. Two soft The location of the KVMRT project is shown in Fig. 1.
computing techniques, namely support vector machine The present study focuses on 132 rock-socketed piles
(SVM) and ANN, were developed in order to estimate the that are mostly underlain by granitic rock of Triassic age,
Qu of concrete spun pipe piles in a study conducted by Pal which dominates the undulating terrain in the study area.
and Deswal [26], who concluded that ANN provided the This granite forms part of the Main Range Granite that has
best predictive model among those tested. An ANN pre- intruded into folded and regionally metamorphosed clastic
dictive model was developed by Shahin et al. [27] to pre- and calcareous Palaeozoic rocks, with the overlying soil
dict the bearing capacity of drilled shafts. Jianbin et al. [28] predominantly consisting of silty clay and sandy silt.

123
Neural Comput & Applic

were excluded, since such data may not characterise

the true intact rock strength.
(c) Underlying the top soil layer up to a maximum depth
of 16.5 m below ground level is a highly weathered
residual material consisting of firm to hard sandy silt,
with SPT N-values ranging from 2 to 48 blows/
300 mm.
(d) Subsoil materials with SPT N-values [50 blows/
300 mm are encountered in almost all areas at depths
ranging from 7.5 to 27.0 m below the existing
ground level and can be typically described as hard
sandy silt/very dense silty sand.
High strain dynamic load testing (HSDLT), also known
as the PDA test, is a fast and reliable method with which to
predict pile bearing capacity based on stress wave propa-
gation through the pile. The concept of using wave prop-
agation through piles to estimate the bearing capacity was
first presented by Smith [34]. However, as Smith’s model
was not successfully implemented to predict the actual
Fig. 1 Location of the KVMRT project energy delivered by the hammer, the use of force and
acceleration records in a simplified model to predict pile
bearing capacity was subsequently reported by a group of
A review of reference materials and a collection of
researchers at Case Western Reserve University (the CASE
available subsurface data were conducted in order to
technique) [35], with the main drawback of their proposed
identify the prevailing geological conditions at the tested
technique being that the static resistance was sensitive to
pile locations and their surrounding areas. Based on site
the CASE damping factor. The full power of wave equation
investigations and borehole results, lithology and stratig-
analysis was first realised when combined with dynamic
raphy were observed to be generally consistent with the
monitoring of the pile during driving [36]. Such combined
geological map, with the underlying subsoil comprising
analysis estimates the bearing capacity by measuring the
residual soils of different rocks (mostly granite). Depth to
force and velocity induced by the hammer blow and does
bedrock is in the range of 0.5 m to more than 14 m below
not require the modelling of the hammer and driving sys-
the prevailing ground levels. Overlying the bedrocks are
tem as in Smith’s model; instead, PDA records are used. To
weathered soil zones and overburden layers. In general, the
determine the pile bearing capacity, the measured force and
overburden soil comprises sandy or gravely silt and/or
velocity combined with wave equation analysis are moni-
gravely or silty sand.
tored and calculated using computer programs.
The subsoil conditions encountered at the pile locations
In the present study, the establishment of an extensive
are typically heterogeneous with colours ranging from light
database for the development of predictive models of pile
brown to brown, grey to greyish-brown and yellowish to
ultimate bearing capacity was achieved by conducting 132
yellowish-brown. A general trend of increasing stiffness/
PDA tests on rock-socketed piles used in the KVMRT
relative density with depth of subsoil was observed. As
project. These tests were carried out using pile driving
investigated from borelogs, the ground conditions can be
analyser equipment produced by Pile Dynamic, Inc. For the
summarised as follows:
development of an appropriate dataset for use in predictive
(a) Rock masses at the pile locations are in the range of models, the selection of input (predictor) parameters plays
slightly weathered to highly weathered. a crucial role. Hence, the most influential parameters
(b) Uniaxial compressive strength (UCS) tests were regarding pile bearing capacity should be selected.
conducted in accordance with the International As mentioned on numerous occasions elsewhere (e.g.
Society of Rock Mechanics [33]; the obtained results [6]), pile geometry including length and diameter is the
were in the range of 23–68 MPa for different rock most important parameter influencing the ultimate pile
types. Furthermore, a range of 12.5–95 was achieved bearing capacity (Qu). Therefore, in order to incorporate
for rock quality designation (RQD) values. It should the effects of both soil and rock layers, two ratios related to
be noted that in the UCS tests, if any sample failed pile geometry were selected: length in soil layer (Ls) to
along fractures or any other defects, then its results socket length (Lr), and total length (Lt) to diameter (D). In

123
 Neural Comput & Applic

addition to pile geometry, several studies have highlighted ANN model are artificial neurons, mathematical functions
other parameters as being influential in predictive models derived from biological neurons. Each artificial neuron
of pile bearing capacity. For example, Tu and Tang [37] receives one or more inputs and sums these inputs to
reported that rock strength can significantly influence Qu, produce an output [41], which is weighted and subse-
while Zhang and Einstein [38] concluded that rock UCS is quently passed through a nonlinear activation function.
the most effective parameter controlling the Qu of rock- The earliest neuron was the Threshold Logic Unit pro-
socketed piles. Therefore, this parameter was selected as a posed by McCulloch and Pitts [42]. A set of parallel
model input in the present study. For the soil layer, several interconnected processing units, called nodes or neurons,
important parameters are mentioned in the literature, such form the basis of an ANN. An activation function is present
as friction angle (/), unit weight (c), cone penetration test along each neuron, with the weights transferring the acti-
(CPT), SPT and cohesion (C). Among these, the SPT N- vation signal between nodes. The data-processing ability of
value was chosen here as an input to incorporate the effect an ANN is thus closely related to its architecture and
of the soil layer on Qu. In summary, Ls/Lr, Lt/D, UCS and weights.
SPT N-value were considered as model inputs to predict The distinguishing property of ANNs is their ability to
the Qu of the rock-socketed piles. These inputs were learn from examples and thereby improve their perfor-
selected due to the fact that a lower number of input mance. ANNs utilise a learning algorithm to learn the
parameters are preferable as it can reduce predictive model relationships between input and output data through an
complexity [39]. A summary of input and output parame- interactive process, with the back-propagation (BP) algo-
ters used in the present study, together with their statistical rithm generally considered the most popular and effective
information, is tabulated in Table 1. Figure 2 shows the learning algorithm for multi-layered networks [43, 44].
input and output parameters used for modelling purposes. Fundamentally, BP learning consists of two passes through
the different layers of the network, namely forward and
backward passes. In the forward pass, the input pattern is
3 Methods for the prediction of pile ultimate applied to the nodes of the network and its effect propa-
bearing capacity gates through the network layer by layer. Finally, an output
is produced as the actual response of the network. If the
In this research, two intelligent models—ANN and hybrid response is different from the desired value, error correc-
PSO–ANN—were applied to estimate the Qu of rock- tion is required to adjust weights and bias levels in the
socketed piles. For this purpose, 132 piles socketed in network in accordance with the error value. Hence, the BP
different rock types (mostly granite) as part of the KVMRT algorithm is used for this purpose.
project were investigated and their influential parameters As a first step in ANN modelling used for predicting Qu,
measured. A database comprising four inputs (Ls/Lr, Lt/D, all datasets should be normalised via the following
UCS and N) and one output (Qu) was prepared to develop equation:
the models. The following sections include an explanation
Xnorm ¼ ðX Xmin Þ=ðXmax Xmin Þ ð1Þ
of these models and their procedures.
where X and Xnorm represent the measured and normalised
3.1 ANN model values, respectively, and Xmin and Xmax are the minimum
and maximum values of the measured parameters,
ANN is a flexible nonlinear function approximation respectively. In the present study, five different datasets
employed to determine the relationship between desired were selected randomly based on several previous studies
input and output data [40]. In contrast to most empirical (e.g. [45]) for training and testing to develop the intelligent
and statistical methods, which require previous knowledge, models. The idea behind using such datasets for testing is
ANNs learn from examples to obtain a close approximation to check the prediction performance of each model and
relationship among the data. The constitutive units in an thus select the best one. Swingler [46] and Looney [47]

Table 1 Summary of input and

Parameter Symbol Unit Category Min Max Mean SD
output parameters used in this
study Soil length to socket length ratio Ls/Lr – Input 0.23 99.04 6.80 9.94
Total length to diameter ratio Lt/D – Input 3.93 37.05 16.20 7.34
Uniaxial compressive strength UCS MPa Input 23 68 46.60 10.33
Standard penetration test N – Input 2 167 53.60 53.70
Ultimate bearing capacity Qu kN Output 12,300 48,500 26,637 8273

123
Neural Comput & Applic

Fig. 2 Input and output parameters used for modelling purposes

recommended that 20 and 25 % of all datasets can be used Table 2 Proposed equation for determining optimum number of
for testing purposes, respectively. Considering these rec- hidden nodes
ommendations, 20 % of the database (26 datasets) was Heuristic Reference
selected randomly for testing, with the remaining 80 % of
data (106 datasets) used for training the constructed mod- B2 9 Ni ? 1 Hecht-Nielsen [52]
els. Random data selection for the proposed models was (Ni ? N0)/2 Ripley [53]
2þN0 Ni þ0:5N0 ðN02 þNi Þ 3 Paola [54]
performed based on ANN code written by the authors. Ni þN0
The most difficult task to be accomplished in ANN 2Ni/3 Wang [55]
modelling is to determine the best network architecture, in pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Ni  N0 Masters [56]
terms of selecting both the proper ANN training algorithm 2Ni Kaastra and Boyd [57]
and the number of hidden nodes in hidden layers [45]. Kannellopoulas and Wilkinson [58]
Many studies have addressed the successful utilisation of
the Levenberg–Marquardt (LM) back-propagation training Ni number of input neuron, N0 number of output neuron
algorithm (e.g. [48]), with a number suggesting that the LM
algorithm converges, while other conventional back-prop- nodes can vary from one to nine. Accordingly, a range of
agation algorithms diverge (e.g. [6, 49]). It is well estab- models were constructed for which the network perfor-
lished that an ANN with one hidden layer can approximate mance was evaluated using the results of both R2 and root-
any continuous function [50]. Furthermore, constructing mean-squared error (RMSE), as shown in Tables 3 and 4,
ANN models with one hidden layer is also of interest due respectively. As also indicated in these tables, each model
to its beneficial effect on reducing model complexity and, was iterated five times. Note that a predictive model with
as a consequence, the likelihood of model over-fitting. higher R2 and lower RMSE values is considered to perform
Based on the above discussion, all intelligent models in the better. According to the obtained average R2 and RMSE
present study were constructed using one hidden layer. values for both training and testing datasets, model no. 5,
According to Sonmez et al. [51], determining neuron with 5 hidden neurons, was found to outperform the other
number(s) in the hidden layer is widely considered the constructed models. Therefore, the architecture of this
most critical task in the establishment of ANN architecture. model (4 9 5 9 1) was selected as the optimum ANN
As displayed in Table 2, various researchers have recom- architecture. After comparing the prediction performances
mended equations with which to obtain the proper number of five different models built using the best ANN config-
of hidden nodes in a hidden layer based on the number of uration (discussed later), the best model was then selected.
input and output parameters. It is worth noting that 2
Ni ? 1 is the upper limit for the number of hidden layer 3.2 PSO–ANN model
neurons needed to map any continuous function, as stated
by Caudill [59]. Based on the equations presented in The particle swarm optimisation (PSO) algorithm was first
Table 2 and considering the number of input and output developed by Kennedy and Eberhart [60] as a bird swarm
parameters used in this research, the number of hidden simulation. The capability of each individual in dealing

123
 Neural Comput & Applic

Table 3 Obtained results of R2 for several ANN models with different hidden nodes
Model Nodes in Network result
no. hidden layers
R2
Iteration 1 Iteration 2 Iteration 3 Iteration 4 Iteration 5 Average
Train Test Train Test Train Test Train Test Train Test Train Test

1 1 0.672 0.610 0.867 0.578 0.674 0.629 0.836 0.641 0.659 0.544 0.742 0.600
2 2 0.809 0.653 0.843 0.671 0.873 0.200 0.761 0.348 0.847 0.663 0.827 0.507
3 3 0.849 0.604 0.869 0.694 0.834 0.809 0.880 0.413 0.763 0.658 0.839 0.636
4 4 0.834 0.737 0.753 0.593 0.796 0.812 0.850 0.808 0.867 0.776 0.820 0.745
5 5 0.843 0.817 0.846 0.808 0.846 0.830 0.850 0.812 0.828 0.837 0.843 0.821
6 6 0.815 0.778 0.839 0.832 0.826 0.799 0.845 0.832 0.805 0.760 0.826 0.800
7 7 0.837 0.769 0.801 0.834 0.828 0.726 0.843 0.748 0.848 0.769 0.831 0.769
8 8 0.808 0.814 0.856 0.677 0.848 0.731 0.819 0.839 0.783 0.852 0.823 0.783
9 9 0.790 0.841 0.867 0.755 0.696 0.712 0.794 0.812 0.743 0.850 0.778 0.794

Table 4 Obtained results of RMSE for several ANN models with different hidden nodes
Model Nodes in Network result
no. hidden layers
RMSE
Iteration 1 Iteration 2 Iteration 3 Iteration 4 Iteration 5 Average
Train Test Train Test Train Test Train Test Train Test Train Test

1 1 0.122 0.203 0.094 0.191 0.147 0.185 0.118 0.173 0.165 0.214 0.129 0.193
2 2 0.097 0.139 0.116 0.148 0.093 0.297 0.108 0.208 0.106 0.163 0.104 0.191
3 3 0.096 0.200 0.100 0.205 0.092 0.181 0.096 0.267 0.105 0.142 0.098 0.199
4 4 0.093 0.182 0.106 0.151 0.097 0.148 0.088 0.127 0.091 0.137 0.095 0.149
5 5 0.095 0.089 0.092 0.135 0.092 0.125 0.091 0.130 0.101 0.083 0.094 0.112
6 6 0.098 0.133 0.099 0.127 0.101 0.138 0.091 0.121 0.110 0.133 0.100 0.130
7 7 0.097 0.129 0.111 0.122 0.099 0.140 0.090 0.139 0.089 0.155 0.097 0.137
8 8 0.105 0.119 0.088 0.166 0.095 0.138 0.099 0.113 0.110 0.098 0.099 0.127
9 9 0.118 0.103 0.084 0.131 0.121 0.152 0.120 0.114 0.109 0.099 0.110 0.120

with the previous experiences of the swarm is known as corresponding value of the objective function for each
swarm intelligence, an ability that directs the swarm particle determining its fitness. After evaluating the fitness
towards the optimum goal. An essential concept of PSO is of the swarm, the velocity of all particles is defined by a
related to the decision-making process of individuals in the function based on the best position of the swarm and the
swarm, with Kennedy and Eberhart [60] stating that each best position of each particle using Eq. 2. Then, the next
individual makes a decision based on the following items: position of all particles is identified by means of their
velocities using Eq. 3. This process iterates until the ter-
1. the personal experiences of an individual that produced
mination criterion of the swarm is met. This criterion can
its best results so far
be based on the number of iterations, a specific accuracy
2. the experiences of other individuals in the swarm that
level or both.
produced the best results in the entire swarm
population !
   !
  
v!
new ¼ ~þ
v C 1  pbest ~p þ C 2  gbest ~
p ð2Þ
In the first step of Kennedy and Eberhart’s PSO algorithm,
a certain number of individuals, also known as particles, new ¼ ~
p! p þ v!
new ð3Þ
are placed in the search space in a random pattern. Each
particle represents a feasible solution. The goal of the new , ~,
where v! new and ~
v p! p are new velocity, current velocity,
swarm is determined by an objective function, with the new position and current position of particles, respectively;

123
Neural Comput & Applic

!
C1 and C2 represent pre-defined coefficients; pbest signifies
!

the personal best position of a particle; and gbest denotes
the global best position among all particles. To obtain more
control on the velocity, an inertia weight (w) can be
described in the velocity equation [61]. It is notable that in
the basic PSO proposed by Kennedy and Eberhart [60], no
inertia weight is included. Using the inertia weight, the
contribution ratio of the previous velocity of a particle to
its velocity at the current time stage is determined [62]. The
equation for updated velocity is as follows:
!
 !
v! ~þ C1  ðpbest ~Þ
new ¼ w:v p þ C2  ðgbest ~Þ p ð4Þ
where w denotes the inertia weight. The standard flow
chart of the PSO algorithm is presented in Fig. 3.
The velocity equation has three components: initial,
cognitive and social. Exploration and exploitation, the two
search specifications, are dependent on the velocity of
particles as controlled by the values of C1 and C2 . As a
result, these values control PSO search behaviour. Gener-
ally, C1 ¼ C2 ¼ 2 is set in the early search stage, as
mentioned in the study conducted by Yagiz and Karahan
[63]. If the value of C1 is greater, convergence is achieved
more rapidly, whereas if the value of C2 is increased, new
solutions can be discovered within the search space. Poli
et al. [61] showed that the combination of a low inertia
weight value in the initial step and a high inertia weight
value at convergence produces the best PSO performance.
Several researchers have attempted to enhance the per-
formance and generalisation capabilities of ANNs in both
engineering and science applications through the use of
PSO algorithms (e.g. [64, 65]). Since BP is a local search
learning algorithm, the optimum search process of ANN
may fail and return an unsatisfied solution. In contrast, PSO
is known as a robust global search method and as such can
be employed to adjust the weight and bias of ANNs and
thereby enhance the latter’s performance capacity. At a
local minimum, the use of ANNs increases the probability
of convergence, whereas PSO is able to find a global
minimum. Therefore, a hybrid PSO–ANN model enjoys
the search properties of both ANN and PSO. In such a Fig. 3 Standard PSO flow chart [60]
model, PSO searches for the global minimum in the search
space, with ANN then employing this information to find decreased through changing the particle positions. To
the best results. !
update the velocity equation, values of pbest (i.e. the
The learning procedure involved in hybrid PSO–ANN
lowest error obtained by each particle until that moment)
models is presented in Fig. 4. This procedure starts with !
the initialisation of a group of random particles, during and gbest (i.e. the lowest error obtained by all particles
which the positions of those particles that are representa- until that moment) are used. As a result, a value is pro-
tive of the ANN weights and biases are assigned randomly. duced with which to adjust particle positions to the best
Following this step, using the initial weights and biases (i.e. solutions. Then, using the updated positions, a new error is
the initial positions of particles), the hybrid PSO–ANN is obtained. This procedure continues until the termination
trained and the error between actual and predicted values is criteria are met, with the optimisation output obtained with
!
computed. At each iteration, the computed error is respect to the value of gbest.

123
 Neural Comput & Applic

Fig. 4 PSO–ANN learning

process

A three-layered hybrid PSO–ANN model was here for each nominee swarm size, a value of 2 for both coeffi-
developed to predict Qu, based on the same datasets cients (C1 and C2) and an inertia weight of 0.25. Note that
implemented during ANN modelling. The following sec- during PSO–ANN modelling, the previously obtained ANN
tions describe the process involved in developing the PSO– architecture of 4 9 5 9 1 was used (see Sect. 3.1). The
ANN predictive model. target of the sensitivity analyses was to find the minimum
RMSE and maximum R2 of the network. Table 5 shows the
3.2.1 Swarm size results of sensitivity analyses for the selected training and
testing datasets using swarm sizes in the range of 25–500. In
To develop the PSO–ANN predictive model, the proper this table, the R2 and RMSE values are tabulated in terms of
PSO parameters were determined based on several para- the training and testing datasets used for each model.
metric studies. The first parametric investigation was Generally, an increase in swarm size causes a rise in R2
conducted in terms of swarm size (number of particles). values and a decrease in RMSE values. Since the selection
Whereas a small swarm may fail to converge to a global of the best model was too difficult at this stage, the simple
solution, choosing a large swarm may cause a delay in ranking method proposed by Zorlu et al. [69] was applied to
convergence and an increase in training time. According to obtain the optimum swarm size. Based on this method, each
Mendes et al. [66], swarm size commonly varies from 20 to performance index was ordered for each class, with the best
50, with the optimum number depending on the specifi- index assigned the highest rating. For instance, R2 values of
cations of the problem under study. According to Kennedy 0.663, 0.712, 0.681, 0.705, 0.751, 0.759, 0.796, 0.717,
and Clerc [67], optimum swarm size depends on the 0.744, 0.731, 0.753 and 0.773 were obtained for the training
number of variables for each particle. However, Clerc [68] datasets of models 1–12, respectively (see Table 5), pro-
later showed that the result of this equation might not ducing equivalent ratings of 1, 4, 2, 3, 8, 10, 12, 5, 7, 6, 9
always be a good estimator and thus there is currently no and 11. This procedure was then repeated for the results of
theoretical method available with which to accurately RMSE analysis. After this process, the obtained ratings of
estimate optimum swarm size. As a result, this value is the performance indices for the training and testing datasets
typically established via a sensitivity test conducted on were summed for each model, as shown in the final column
swarm convergence. in Table 5 (total rank). According to the obtained total
A series of sensitivity analyses of swarm size were rank values, model no. 7 was found to provide a higher-
applied in the present study in order to identify the appro- performance capacity compared to the other models, and
priate number of particles in the swarm. These analyses thus, a swarm size of 250 was selected for PSO–ANN
were conducted using an iteration number of 200 repetitions modelling.

123
Neural Comput & Applic

Table 5 Effects of different

Model no. Swarm size Network result Ranking Total rank
number of particles (swarm
size) and their results in Train Test Train Test
predicting Qu 2 2 2
R RMSE R RMSE R RMSE R2 RMSE

1 25 0.663 0.134 0.550 0.136 1 2 4 6 13

2 50 0.712 0.124 0.539 0.138 4 4 3 4 15
3 75 0.681 0.126 0.853 0.095 2 3 12 12 29
4 100 0.705 0.119 0.472 0.185 3 7 1 2 13
5 150 0.751 0.117 0.536 0.137 8 9 2 5 24
6 200 0.759 0.113 0.618 0.131 10 11 6 7 34
7 250 0.796 0.106 0.721 0.115 12 12 11 10 45
8 300 0.717 0.120 0.659 0.139 5 6 9 3 23
9 350 0.744 0.118 0.658 0.120 7 8 8 9 32
10 400 0.731 0.121 0.565 0.139 6 5 5 3 19
11 450 0.753 0.116 0.697 0.113 9 10 10 11 40
12 500 0.773 0.113 0.636 0.121 11 11 7 8 37

3.2.2 Termination criteria number of iterations in the PSO–ANN modelling procedure

to predict Qu.
Termination criteria are defined as the conditions required
to end the iterative process. In general, termination criteria 3.2.3 Coefficients of velocity equation
involve either one of two conditions: the reaching of a
required precision value and the exceedance of a maximum Other influential parameters on PSO include the coeffi-
number of iterations. Whereas the first criterion can be cients of velocity, C1 and C2. Based on the original coef-
easily set, determining a constant number for the second ficients proposed by Kennedy and Eberhart [60] and the
condition is not easy. A practical way in which to deter- modified coefficients of Clerc and Kennedy [70], a series of
mine this number is to compare the results of the analysis candidates for C1 and C2 were utilised in the present study.
of different iteration numbers. Typically, the iteration Indeed, although the original values (C1 = C2 = 2) of
number is lower for problems with simple search spaces these coefficients have been implemented on several
and higher for problems involving complex search spaces. occasions elsewhere (e.g. [63]), a range of different values
In the present study, the maximum number of iterations have also been applied to solve problems in the field of
was used as a termination criterion. As mentioned above, geotechnical engineering. For example, C1 = 1.333 and
training time is gradually increased with an increase in the C2 = 2.667 were suggested by Hajihassani et al. [71] to
maximum number of iterations. However, based on pre- predict air-overpressure resulting from quarry blasting,
vious investigations, there is no common method with while C1 = 1.714 and C2 = 2.286 were employed to pre-
which to achieve the maximum number of iterations. dict geotechnical problems by both Jahed Armaghani et al.
Therefore, a series of sensitivity analyses were applied to [72] and Momeni et al. [32]. Considering these results,
the PSO–ANN network to find this value. These analyses further sensitivity analysis is likely needed in order to
were conducted by setting a fixed iteration number of 1000, select the best coefficient combination. Table 6 displays
a value of 2 for both coefficients of the velocity equation different combinations of C1 and C2 and the obtained
(C1 and C2) and an inertia weight of 0.25. The target of this results. As can be seen from this table, 12 models were
sensitivity analysis was to monitor the changes in the cost constructed and the related results utilised for training and
function (RMSE) taking place after each iteration. Figure 5 testing datasets based on R2 and RMSE. In constructing
displays the results of sensitivity analyses regarding itera- these models, the obtained parameters from the previous
tion number for the swarm size range of 25–500. Accord- sections were used. As before, a simple ranking method
ing to this figure, whereas significant changes took place was employed to identify the best combinations. According
during the first few iterations, only moderate changes were to the total rank values presented in Table 6, model no. 6
observed up to iteration no. 400, after which point no (C1 = 2.286, C2 = 1.714) outperformed the other models
significant variation in the RMSE results was recorded. and thus these values were selected as the coefficients of
Therefore, a value of 400 was selected as the maximum the velocity equation in this study.

123
 Neural Comput & Applic

Fig. 5 The effect of number of iterations on network performance

Table 6 Effects of different combinations of C1 and C2

Model no. C1 C2 C1 ? C2 Network result Ranking Total rank
Train Test Train Test
R2 RMSE R2 RMSE R2 RMSE R2 RMSE

1 0.8 3.2 4 0.691 0.121 0.796 0.119 3 4 7 6 20

2 1.333 2.667 4 0.774 0.115 0.533 0.111 9 8 3 8 28
3 1.714 2.286 4 0.765 0.112 0.812 0.103 8 9 9 10 36
4 3.2 0.8 4 0.728 0.120 0.515 0.155 6 5 2 2 15
5 2.667 1.333 4 0.879 0.098 0.871 0.095 11 11 10 12 44
6 2.286 1.714 4 0.907 0.089 0.876 0.098 12 12 11 11 46
7 2.5 2.5 5 0.725 0.116 0.888 0.105 5 7 12 9 33
8 2 2 4 0.788 0.110 0.798 0.105 10 10 8 9 37
9 1.75 1.75 3.5 0.756 0.115 0.455 0.122 7 8 1 5 21
10 1.5 1.5 3 0.716 0.122 0.585 0.146 4 3 4 3 14
11 1.25 1.25 2.5 0.686 0.124 0.763 0.124 2 2 6 4 14
12 1 1 2 0.657 0.137 0.699 0.112 1 1 5 7 14

3.2.4 Inertia weight Based on these values (0.25, 0.5, 0.75 and 1), four PSO–
ANN models were built, as presented in Table 7. These
Further sensitivity analyses were performed to find the models were constructed using the PSO parameters
optimum value of inertia weight (w) in the velocity equa- obtained in the previous sections and considering the
tion. A range of different inertia weights (0.25, 0.5. 0.75 suggested ANN-derived architecture. Analytical results
and 1) were proposed by Clerc and Kennedy [70], while were recorded based on R2 and RMSE values, with a rank
Momeni et al. [32] also suggested an inertia weight of 0.25 value assigned to each R2 and RMSE for the training and
in predicting the UCS of granite and limestone samples. testing datasets. Considering the total rank values presented

123
Neural Comput & Applic

Table 7 Effects of different

Model no. Inertia weight Network result Ranking Total rank
inertia weights on network
performance Train Test Train Test
2 2 2
R RMSE R RMSE R RMSE R2 RMSE

1 0.25 0.910 0.081 0.864 0.081 2 3 1 1 7

2 0.5 0.849 0.092 0.898 0.075 1 1 2 2 6
3 0.75 0.922 0.073 0.918 0.063 3 4 4 4 15
4 1 0.924 0.084 0.902 0.068 4 2 3 3 12

in Table 7, it can be concluded that by setting the inertia Table 8 Normalised testing data for set 1
weights equal to 0.75, the best network performance can be Dataset no. Lp/D Ls/Lr N-SPT UCS (Mpa) Total (kN)
obtained.
1 8.8 1.5 6.6 40 31,000
3.2.5 Network architecture 2 7.9 1.2 7.5 34 20,500
3 11.9 4.6 6.0 24 14,000
In the last step of PSO–ANN modelling, as with the ANN 4 15.8 5.8 19.5 27 24,750
predictive model, network architecture was determined in 5 13.7 6.9 20.0 43 32,190
order to train the system. To achieve a better model com- 6 11.1 3.5 4.7 34 19,550
parison, the previously suggested ANN architecture of 7 21.0 7.3 115.0 43 30,171
4 9 5 9 1 was employed in PSO–ANN modelling to 8 10.4 3.0 156.0 47 36,712
predict the Qu of the rock-socketed piles. Considering the 9 16.8 0.4 4.3 30 22,931
obtained parameters from the previous steps and using the 10 18.9 3.6 117.5 32 31,580
same five randomly selected datasets employed in ANN 11 22.3 6.4 7.1 30 15,730
modelling, five PSO–ANN models were constructed. As an 12 27.2 9.3 12.7 29 17,910
example, normalised testing data for set 1 are presented in 13 12.3 2.6 4.3 26 17,350
Table 8, with the selection of the best PSO–ANN model 14 29.2 9.4 12.7 25 15,310
discussed in the following section. 15 25.5 6.3 12.7 29 20,608
16 7.5 2.0 9.0 36 25,000
17 5.4 1.3 9.0 42 29,000
4 Evaluation of the proposed intelligent systems 18 9.4 2.3 9.0 39 19,200
19 13.9 2.5 9.1 53 27,000
In the present study, two intelligent systems were devel- 20 14.8 10.8 29.9 24 12,300
oped to predict the Qu of rock-socketed piles. For this 21 27.7 24.2 6.0 31 14,050
purpose, all data (132 datasets) were randomly selected to 22 6.9 1.9 145.0 50 45,000
produce five different datasets for the development of the 23 10.0 2.8 118.9 38 31,100
predictive models. These five datasets were applied at the 24 4.3 0.4 113.0 40 31,000
final stage for each predictive model, with the respective 25 25.7 6.7 2.0 32 15,500
results for the training and testing datasets also obtained. 26 11.5 1.1 118.3 55 34,020
Selected performance indices, including R2, RMSE and
variance account for (VAF), were calculated as follows in
order to evaluate the prediction performance of the
developed predictive models:
PN where y, y0 and y~ are the measured, predicted and mean
2 ðy y0 Þ2
R ¼ 1 Pi¼1 ð5Þ y values, respectively; N is the total number of data; and
N
i¼1 ðy y~Þ2 P is the number of predictors. Theoretically, the model is

var ðy y0 Þ
 excellent if R2 is equal to one, VAF is equal to 100, and
VAF ¼ 1  100 ð6Þ RMSE is zero.
var ðyÞ
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi The results of performance index (including R2, RMSE
u N and VAF) analysis of the developed predictive models for
u1 X
RMSE ¼ t ðy y0 Þ2 ð7Þ the five randomly selected datasets based on training and
N i¼1
testing are shown in Table 9. According to this table, the

123
 Neural Comput & Applic

Table 9 Performance indices of each model and their rank values for predictive models
Method Model R2 RMSE VAF Rating for R2 Rating for RMSE Rating for VAF Rank value

ANN Training 1 0.843 0.095 84.203 3 3 2 8

Training 2 0.846 0.092 84.392 4 4 3 11
Training 3 0.846 0.092 84.432 4 4 4 12
Training 4 0.850 0.091 84.786 5 5 5 15
Training 5 0.828 0.101 82.469 2 2 1 5
Testing 1 0.817 0.089 80.609 3 4 4 11
Testing 2 0.808 0.135 77.636 1 1 1 3
Testing 3 0.830 0.125 79.519 4 3 3 10
Testing 4 0.812 0.130 78.439 2 2 2 6
Testing 5 0.837 0.083 83.395 5 5 5 15
PSO–ANN Training 1 0.930 0.063 93.025 3 3 3 9
Training 2 0.935 0.055 93.473 5 5 5 15
Training 3 0.921 0.063 91.957 2 3 2 7
Training 4 0.932 0.061 93.075 4 4 4 12
Training 5 0.910 0.068 90.981 1 2 1 4
Testing 1 0.932 0.058 92.854 4 5 3 12
Testing 2 0.924 0.107 84.329 2 1 1 4
Testing 3 0.932 0.071 92.228 4 2 2 8
Testing 4 0.938 0.059 93.348 5 4 5 14
Testing 5 0.930 0.063 92.977 3 3 4 10

performance index results are very similar and thus Table 10 Total rank values for
Method Model Total
selecting the best model cannot be achieved on this basis predictive models obtained from
rank
five different datasets
alone. To overcome this difficulty, Zorlu et al. [69] sug-
ANN 1 19
gested the use of a simple ranking method in which a
2 14
ranking value is calculated and assigned to each training
3 22
and testing dataset separately, as discussed above (see
4 21
Table 9). Note that to achieve a better comparison between
results, the obtained normalised ANN and PSO–ANN 5 20
model results were here converted to correspond to the pre- PSO–ANN 1 21
normalised data. Total rank values of training and testing 2 19
datasets for the developed predictive models are shown in 3 15
Table 10. According to this table, models no. 3 and 4 4 26
exhibited the best performance in terms of Qu prediction 5 14
via the ANN and PSO–ANN techniques, respectively.
When considering both training and testing datasets, the 5 Sensitivity analysis
prediction performances of the PSO–ANN models were
significantly better than those of the ANN models. Sensitivity analysis was carried out in order to determine
The identified relationships between the measured and the relative influence of each parameter in the network
predicted Qu of the rock-socketed piles using the ANN and system using the cosine amplitude method [73]. All data
PSO–ANN models are illustrated in Figs. 6 and 7, pairs were utilised to build a data array X as follows:
respectively. According to these figures, the PSO–ANN
model, with R2 values of 0.932 and 0.938 for training and X ¼ fx1 ; x2 ; x3 ; . . .; xi ; . . .; xn g ð8Þ
testing data, respectively, achieved a better prediction Variable xi in array X is a length vector of m as:
performance than the ANN model, with equivalent R2
values of 0.846 and 0.830. This suggests the superiority of xi ¼ fxi1 ; xi2 ; xi3 ; . . .; xim g ð9Þ
the PSO–ANN technique in comparison with the employed The following equation expresses the strength of the rela-
ANN technique. tionship (rij) between datasets Xi and Xj :

123
Neural Comput & Applic

Fig. 6 Measured and predicted values of Qu obtained via the ANN model for training and testing datasets

Fig. 7 Measured and predicted values of Qu obtained via the PSO–ANN model for training and testing datasets

6 Summary and conclusion

In this study, an attempt was made to develop intelligent

systems aimed at predicting the Qu of rock-socketed piles.
To this end, 132 rock-socketed piles constructed as part of
the KVMRT project, Malaysia, were investigated and the
influential parameters on pile Qu assessed. Using four
inputs, namely Ls/Lr, Lt/D, UCS and N-SPT, and both ANN
and PSO–ANN modelling procedures, predictive models
were designed and developed for Qu estimation. Based on a
simple ranking method and the results of selected perfor-
mance indices, the best ANN and PSO–ANN models were
Fig. 8 Strength of relationship for each input parameter
identified.
Considering both the training and testing datasets, the
Pm
xik xjk prediction performance (in terms of R2) of the PSO–ANN
Pm k¼12 Pm 2ffi
rij ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð10Þ model (R2 = 0.932) was better than that of the ANN model
k¼1 xik k¼1 xik
(R2 = 0.846). Moreover, taking the testing datasets alone,
Figure 8, which shows the strengths of the relationships a similar pattern was again obtained (R2 = 0.938 and 0.830
(rij values) between the input and output parameters, reveals for PSO–ANN and ANN models, respectively). The
that among all inputs, UCS has the most influence on pile Qu. recorded PSO–ANN model results illustrate the high

123
 Neural Comput & Applic

reliability of this new predictive model in estimating the rocks and comparison of their performances. Eng Comput.
ultimate bearing capacity of rock-socketed piles. By doi:10.1007/s00366-015-0410-5
17. Singh R, Kainthola A, Singh TN (2012) Estimation of elastic
implementing sensitivity analysis, it was found that UCS is constant of rocks using an ANFIS approach. Appl Soft Comput
the most effective input regarding pile ultimate bearing 12(1):40–45
capacity. 18. Shirani Faradonbeh R, Monjezi M, Jahed Armaghani D (2015)
Genetic programing and non-linear multiple regression tech-
Acknowledgments The authors would like to extend their sincere niques to predict backbreak in blasting operation. Eng Comput.
gratitude and appreciation to the Universiti Teknologi Malaysia doi:10.1007/s00366-015-0404-3
(UTM) for all their support that made this research possible. 19. Monjezi M, Hasanipanah M, Khandewal M (2013) Evaluation
and prediction of blast-induced ground vibration at Shur River
Dam Iran, by artificial neural network. Neural Comput Appl
22:1637–1643
References 20. Marto A, Hajihassani M, Jahed Armaghani D, Tonnizam
Mohamad E, Makhtar AM (2014) A novel approach for blast-
1. Coyle HM, Castello RR (1981) New design correlations for piles induced flyrock prediction based on imperialist competitive
in sand. J Geotech Eng ASCE 107:965–986 algorithm and artificial neural network. Sci World J, Article ID
2. Shahin MA (2010) Intelligent computing for modeling axial 643715
capacity of pile foundations. Can Geotech J 47(2):230–243 21. Ocak I, Seker SE (2013) Calculation of surface settlements
3. Kordjazi A, Nejad FP, Jaksa MB (2014) Prediction of ultimate caused by EPBM tunneling using artificial neural network, SVM,
axial load-carrying capacity of piles using a support vector and Gaussian processes. Environ Earth Sci 70(3):1263–1276
machine based on CPT data. Comput Geotech 55:91–102 22. Mahdevari S, Shahriar K, Yagiz S, Shirazi MA (2014) A support
4. Randolph MF (2003) Science and empiricism in pile foundation vector regression model for predicting tunnel boring machine
design. Geotechnique 53(10):847–875 penetration rates. Int J Rock Mech Min Sci 72:214–229
5. Likins G, Rausche F (2004) Correlation of CAPWAP with static 23. Chan WT, Chow YK, Liu LF (1995) Neural network: an alter-
load tests. In: Proceedings of the seventh international conference native to pile driving formulas. Comput Geotech 17(2):135–156
on the application of stress wave theory to piles, The Institute of 24. Lee IM, Lee JH (1996) Prediction of pile bearing capacity using
Engineers Malaysia artificial neural networks. Comput Geotech 18(3):189–200
6. Momeni E, Nazir R, Jahed Armaghani D, Maizir H (2014) Pre- 25. Abu-Kiefa M (1998) General regression neural networks for
diction of pile bearing capacity using a hybrid genetic algorithm- driven piles in cohesionless soils. J Geotech Geoenviron Eng
based ANN. Measurement 57:122–131 124(12):1177–1185
7. American Society for Testing and Materials (2010) Standard test 26. Pal M, Deswal S (2008) Modeling pile capacity using support
method for high-strain dynamic testing of piles, D 4945-08, vector machines and generalized regression neural network.
Annual Book of ASTM Standards. ASTM, Philadelphia, PA, 4, J Geotech Geoenviron Eng 134(7):1021–1024
08 27. Shahin MA, Jaksa MB, Maier HR (2009) Recent advances and
8. Shahin MA, Jaksa MB, Maier HR (2001) Artificial neural net- future challenges for artificial neural systems in geotechnical
work application in geotechnical engineering. Aust Geomech engineering applications. Adv Artif Neural Syst. doi:10.1155/
36:49–62 2009/308239
9. Gokceoglu C, Zorlu K (2004) A fuzzy model to predict the 28. Jianbin Z, Jiewen T, Yongqiang S (2010) An ANN model for
uniaxial compressive strength and the modulus of elasticity of a predicting level ultimate bearing capacity of PHC pipe pile. Earth
problematic rock. Eng Appl Artif Intell 17(1):61–72 and Space, pp 3168–3176
10. Khandelwal M, Singh TN (2007) Evaluation of blast-induced 29. Alkroosh I, Nikraz H (2012) Predicting axial capacity of driven
ground vibration predictors. Soil Dyn Earthq Eng 27(2):116–125 piles in cohesive soils using intelligent computing. Eng Appl
11. Yagiz S, Gokceoglu C, Sezer E, Iplikci S (2009) Application of Artif Intell 25(3):618–627
two non-linear prediction tools to the estimation of tunnel boring 30. Eberhart RC, Shi Y (1998) Evolving artificial neural networks.
machine performance. Eng Appl Artif Intell 22(4):808–814 In: Proceedings of the international conference on neural net-
12. Mishra DA, Basu A (2013) Estimation of uniaxial compressive works and brain, pp PL5–PL13
strength of rock materials by index tests using regression analysis 31. Tonnizam Mohamad E, Jahed Armaghani D, Momeni E, Alavi
and fuzzy inference system. Eng Geol 160:54–68 Nezhad Khalil Abad SV (2014) Prediction of the unconfined
13. Singh R, Vishal V, Singh TN, Ranjith PG (2013) A comparative compressive strength of soft rocks: a PSO-based ANN approach.
study of generalized regression neural network approach and Bull Eng Geol Environ. doi:10.1007/s10064-014-0638-0
adaptive neuro-fuzzy inference systems for prediction of uncon- 32. Momeni E, Jahed Armaghani D, Hajihassani M, Amin MFM
fined compressive strength of rocks. Neural Comput Appl 23(2): (2015) Prediction of uniaxial compressive strength of rock sam-
499–506 ples using hybrid particle swarm optimization-based artificial
14. Verma AK, Singh TN (2011) Intelligent systems for ground neural networks. Measurement 60:50–63
vibration measurement: a comparative study. Eng Comput 33. ISRM (2007) In: Ulusay R, Hudson JA (eds) The complete ISRM
27(3):225–233 suggested methods for rock characterization, testing and moni-
15. Yagiz S, Sezer EA, Gokceoglu C (2012) Artificial neural net- toring: 1974–2006. Suggested methods prepared by the com-
works and nonlinear regression techniques to assess the influence mission on testing methods, International Society for Rock
of slake durability cycles on the prediction of uniaxial com- Mechanics
pressive strength and modulus of elasticity for carbonate rocks. 34. Smith EAL (1960) Pile-driving analysis by the wave equation.
Int J Numer Anal Methods 36:1636–1650 J Soil Mech Found Div 4:35–61
16. Jahed Armaghani D, Mohamad ET, Hajihassani M, Yagiz S, 35. Goble GG, Rausche F, Moses F (1970) Dynamics studies on the
Motaghedi H (2015) Application of several non-linear prediction bearing capacity of piles. Final report to the Ohio department of
tools for estimating uniaxial compressive strength of granitic highways, Case Western Reserve University, Cleveland, Ohio

123
Neural Comput & Applic

36. Fellenius BH (1984) Wave equation analysis and dynamic 55. Wang C (1994) A theory of generalization in learning machines
monitoring. Deep Found J 1:49–55 with neural application. PhD thesis, The University of Pennsyl-
37. Tu J, Tang A (2011) Grey correlation analysis of bearing capacity vania, USA
of rock-socketed pile in soft soil area. In: 2011 international 56. Masters T (1994) Practical neural network recipes in C??.
conference on multimedia technology (ICMT), IEEE, Academic Press, Boston
pp 1081–1085 57. Kaastra I, Boyd M (1996) Designing a neural network for fore-
38. Zhang L, Einstein HH (1998) End bearing capacity of drilled casting financial and economic time series. Neurocomputing
shafts in rock. J Geotech Geoenviron Eng 124(7):574–584 10:215–236
39. Jahed Armaghani D, Tonnizam Mohamad E, Momeni E, Nar- 58. Kanellopoulas I, Wilkinson GG (1997) Strategies and best
ayanasamy MS, Mohd Amin MF (2014) An adaptive neuro-fuzzy practice for neural network image classification. Int J Remote
inference system for predicting unconfined compressive strength Sens 18:711–725
and Young’s modulus: a study on Main Range granite. Bull Eng 59. Caudill M (1988) Neural networks primer part III. Al Expert
Geol Environ. doi:10.1007/s10064-014-0687-4 3(6):53–59
40. Monjezi M, Mehrdanesh A, Malek A, Khandelwal M (2013) 60. Kennedy J, Eberhart RC (1995) Particle swarm optimization. In:
Evaluation of effect of blast design parameters on flyrock using Proceedings of the IEEE international conference on neural net-
artificial neural networks. Neural Comput Appl 23(2):349–356 works (Perth, Australia), IEEE Service Center, Piscataway,
41. Monjezi M, Ahmadi Z, Varjani AY, Khandelwal M (2013) pp 1942–1948
Backbreak prediction in the Chadormalu iron mine using artificial 61. Poli R, Kennedy J, Blackwell T (2007) Particle swarm opti-
neural network. Neural Comput Appl 23(3–4):1101–1107 mization an overview. Swarm Intell 1:33–57
42. McCulloch WS, Pitts W (1943) A logical calculus of the ideas 62. Bansal JC, Singh PK, Saraswat M, Verma A, Jadon SS, Abraham
immanent in nervous activity. Bull Math Biophys 5:115–133 A (2011) Inertia Weight strategies in particle swarm optimiza-
43. Tawadrous AS, Katsabanis PD (2007) Prediction of surface tion. Third World Congress on nature and biologically inspired
crown pillar stability using artificial neural networks. Int J Numer computing. IEEE, Piscataway, pp 640–647
Anal Methods 31(7):917–931 63. Yagiz S, Karahan H (2011) Prediction of hard rock TBM pene-
44. Khandelwal M, Singh TN (2009) Correlating static properties of tration rate using particle swarm optimization. Int J Rock Mech
coal measures rocks with P-wave velocity. Int J Coal Geol Min Sci 48(3):427–433
79:55–60 64. Gordan B, Jahed Armaghani D, Hajihassani M, Monjezi M
45. Jahed Armaghani D, Mohamad ET, Hajihassani M, Abad SANK, (2015) Prediction of seismic slope stability through combination
Marto A, Moghaddam MR (2015) Evaluation and prediction of of particle swarm optimization and neural network. Eng Comput.
flyrock resulting from blasting operations using empirical and doi:10.1007/s00366-015-0400-7
computational methods. Eng Comput. doi:10.1007/s00366-015- 65. Bashir ZA, El-Hawary ME (2009) Applying wavelets to short-
0402-5 term load forecasting using PSO-based neural networks. IEEE
46. Swingler K (1996) Applying neural networks: a practical guide. Trans Power Syst 24(1):20–27
Academic Press, New York 66. Mendes R, Cortes P, Rocha M, Neves J (2002) Particle swarms
47. Looney CG (1996) Advances in feed-forward neural networks: for feed forward neural net training. In: Proceedings of the IEEE
demystifying knowledge acquiring black boxes. IEEE Trans international joint conference on neural networks, Honolulu, HI,
Knowl Data Eng 8(2):211–226 USA, 12–17 May 2002, pp 1895–1899
48. Ceryan N, Okkan U, Kesimal A (2013) Prediction of unconfined 67. Kennedy J, Clerc M (2006) Standard PSO 2006. http://www.
compressive strength of carbonate rocks using artificial neural particleswarm.info/Standard-PSO-2006.c
networks. Environ Earth Sci 68(3):807–819 68. Clerc M (2011) Standard particle swarm optimisation, from 2006
49. Hagan MT, Menhaj MB (1994) Training feed forward networks to 2011. clerc.maurice.free.fr/pso/SPSO_descriptions.pdf (2011-
with the Marquardt algorithm. IEEE Trans Neural Netw 07-13 version)
5:861–867 69. Zorlu K, Gokceoglu C, Ocakoglu F, Nefeslioglu HA, Acikalin S
50. Baheer I (2000) Selection of methodology for modeling hys- (2008) Prediction of uniaxial compressive strength of sandstones
teresis behavior of soils using neural networks. J Comput Aided using petrography-based models. Eng Geol 96(3–4):141–158
Civil Infrastruct Eng 5(6):445–463 70. Clerc M, Kennedy J (2002) The particle swarm explosion, sta-
51. Sonmez H, Gokceoglu C, Nefeslioglu HA, Kayabasi A (2006) bility, and convergence in a multi-dimensional complex space.
Estimation of rock modulus: for intact rocks with an artificial IEEE Trans Evol Comput 6(1):58–73
neural network and for rock masses with a new empirical equa- 71. Hajihassani M, Jahed Armaghani D, Sohaei H, Mohamad ET,
tion. Int J Rock Mech Min Sci 43:224–235 Marto A (2014) Prediction of airblast-overpressure induced by
52. Hecht-Nielsen R (1987) Kolmogorov’s mapping neural network blasting using a hybrid artificial neural network and particle
existence theorem. In: Proceedings of the first IEEE international swarm optimization. Appl Acoust 80:57–67
conference on neural networks, San Diego, CA, USA, pp 11–14 72. Jahed Armaghani D, Hajihassani M, Mohamad ET, Marto A,
53. Ripley BD (1993) Statistical aspects of neural networks. In: Noorani SA (2014) Blasting-induced flyrock and ground vibra-
Barndoff-Neilsen OE, Jensen JL, Kendall WS (eds) Networks and tion prediction through an expert artificial neural network based
chaos-statistical and probabilistic aspects. Chapman & Hall, on particle swarm optimization. Arab J Geosci 7:5383–5396
London, pp 40–123 73. Yang Y, Zang O (1997) A hierarchical analysis for rock engi-
54. Paola JD (1994) Neural network classification of multispectral neering using artificial neural networks. Rock Mech Rock Eng
imagery. MSc thesis, The University of Arizona, USA 30:207–222

123