Sei sulla pagina 1di 13

International Journal of Rock Mechanics & Mining Sciences 85 (2016) 174–186

Contents lists available at ScienceDirect

International Journal of
Rock Mechanics & Mining Sciences
journal homepage: www.elsevier.com/locate/ijrmms

Prediction of the uniaxial compressive strength of sandstone


using various modeling techniques
Danial Jahed Armaghani a,n, Mohd For Mohd Amin a, Saffet Yagiz b,
Roohollah Shirani Faradonbeh c, Rini Asnida Abdullah a
a
Department of Geotechnics and Transportation, Faculty of Civil Engineering, Universiti Teknologi Malaysia, 81310, UTM, Skudai, Johor, Malaysia
b
Department of Geological Engineering, Pamukkale University, Denizli, Turkey
c
Young Researchers and Elite Club, South Tehran Branch, Islamic Azad University, Tehran, Iran

art ic l e i nf o a b s t r a c t

Article history: Sandstone blocks were collected from Dengkil site in Malaysia and brought to laboratory, and then intact
Received 10 June 2015 samples prepared for testing. Rock tests, including Schmidt hammer rebound number, P-wave velocity,
Received in revised form point load index, and UCS were conducted. The established dataset is composed of 108 cases. Conse-
18 January 2016
quently, the established dataset was utilized for developing the simple regression, linear, non-linear
Accepted 23 March 2016
multiple regressions, artificial neural network, and a hybrid model, developed by integrating imperialist
competitive algorithm with ANN. After performing the relevant models, several performance indices i.e.
Keywords: root mean squared error, coefficient of determination, variance account for, and total ranking, are ex-
Uniaxial compressive strength amined for selecting the best model and comparing the obtained results. It is obtained that the ICA–ANN
Artificial neural network
model is superior to the others. It is concluded that the hybrid of ICA–ANN could be used for predicting
Imperialist competitive algorithm
UCS of similar rock type in practice.
Non-destructive tests
Point load index & 2016 Elsevier Ltd. All rights reserved.

1. Introduction input parameters is used for estimating the UCS of rocks.16–20


Some of the milestone simple relationships between the UCS and
The UCS of rock is one of the significant parameters required relevant rock properties including the Rn, Vp, Is(50) published in the
for rock related engineering projects like excavation and tunnel- literature are given in Table 1.
ing. The test may be performed directly in the laboratory on rock Besides traditional empirical relations, various computer aid
sample which can be prepared and then tested according to in- techniques including artificial neural network (ANN), adaptive
ternational testing standards such as the American Society for neuro fuzzy inference system (ANFIS), fuzzy logic (FL), fuzzy in-
Testing and Materials (ASTM) or International Society for Rock ference system (FIS), genetic algorithm (GA), particle swarm op-
Mechanics (ISRM). Determining the UCS of rock in a laboratory is timization (PSO) and also hybrid models like ANN-PSO have been
expensive, time consuming, and also needs well-prepared test performed to estimate the UCS of rocks.35–45
sample which difficult to obtain for relatively weak rocks such as, Further, imperialist competitive algorithm (ICA) which was
shale, marl or sandstone. Because of these obstacles, predicting the introduced by Atashpaz-Gargari and Lucas46 is a global search
UCS of rock is often interest of scientists dealing with engineering population-based algorithm. The ICA is an evolutionary compu-
geology and rock mechanics. So, the UCS is traditionally estimated tation that does not need the gradient of the function in its opti-
as a function of physical and mineralogical properties of rocks. mization process. Kaveh and Talatahari47 applied the ICA to solve
Many researchers have introduced empirical equations that is problem of skeletal structures. Nazari-Shirkouhi et al.48 performed
the result of simple or multiple regression analysis to estimate the the ICA to solve the integrated product mix-outsourcing optimi-
UCS of various rocks.1–14 These researches have been performed by zation problem. Taghavifar et al.49 developed both ANN and the
using different rock type and rock properties; due to that, these ICA–ANN systems to predict soil compaction indices. They suc-
empirical relations that depend on rock types may be vary. Fur- cessfully indicated that the network optimized by the ICA shows
ther, multiple regression analysis technique which at least two better performance in comparison with conventional ANN tech-
nique. Marto et al.50 and Hajihassani et al.51 integrated the ICA
n
with ANN to optimize the ANN model for predicting environ-
Correspondence to: Department of Geotechnics and Transportation, Faculty of
Civil Engineering, Universiti Teknologi Malaysia (UTM), 81310 Skudai, Johor,
mental issues of blasting. They stated that hybrid model (ICA–
Malaysia. ANN) is superior to other techniques (i.e., ANN).
E-mail address: danialarmaghani@gmail.com (D. Jahed Armaghani). In fact the ICA is recently introduced algorithm and yet no

http://dx.doi.org/10.1016/j.ijrmms.2016.03.018
1365-1609/& 2016 Elsevier Ltd. All rights reserved.
D. Jahed Armaghani et al. / International Journal of Rock Mechanics & Mining Sciences 85 (2016) 174–186 175

Table 1
Various relationships between the UCS and Rn, Vp, Is(50) in the literature.

References Relationship R2 Description

Aufmuth1 UCS ¼0.33(Rn.ρ) 1.35 0.80 25 different lithology


Singh et al.2 UCS ¼2Rn 0.86 Sandstone, siltstone, mudstone etc.,
Sachpazis21 UCS ¼4.29Rn  67.52 0.96 33 different carbonates
Xu et al.22 UCS ¼2.98e (0.06 Rn) 0.95 Mica-schist
4
Tugrul and Zarif UCS ¼8.36Rn  416 0.87 Granite
Kahraman6 UCS ¼9.95Vp1.21 0.69 27 different rock samples
UCS ¼8.41IS(50) þ 9.51 0.72
Sulukcu and Ulusay5 UCS ¼15.3IS(50) 0.64 23 samples in different rock types
Yasar and Erdogan7 UCS ¼0.000004Rn4.29 0.89 13 samples of various carbonate rock types
Vp ¼0.0317UCS þ2.02 0.64
Tsiambaos and Sabatakakis23 UCS ¼7.3IS(50)1.71 0.82 188 samples (limestone, sandstone and marlstones)
Entwisle et al.24 UCS ¼0.78e 0.88Vp 0.53 171 samples of volcanic rock
Kahraman et al.25 UCS ¼10.22 IS(50) þ 24.31 0.75 38 different rock samples
Basu and Aydin8 UCS ¼18IS(50) 0.97 40 granitic rock samples
Yilmaz and Yuksek26 UCS ¼12.4IS(50)  9.0859 0.81 39 gypsum sample sets
Kilic and Teymen10 UCS ¼0.0137Rn 2721 0.93 Different rock types
Yagiz17 UCS ¼0.0028Rn 2.584 0.85 9 different rock types
Yagiz11 UCS ¼0.258Vp3.543 0.92 9 different rock type, sedimentary and meta-sedimentary rocks
UCS ¼49.4Vp  167 0.89
Khandelwal and Singh27 UCS ¼0.1333Vp  227.19 0.96 12 coal samples
Moradian and Behnia28 UCS ¼165 exp( 4.45/Vp) 0.70 64different rock samples
Diamantis et al.29 UCS ¼19.79IS(50) 0.74 32 samples of serpentinite
Mishra and Basu30 UCS ¼14.63IS(50) 0.88 60 samples (granite, schist and sandstone)
Kohno and Maeda31 UCS ¼16.4IS(50) 0.85 44 different rock samples
Khandelwal32 UCS ¼0.033Vp  34.83 0.87 12 samples of a wide rock types
Minaeian and Ahangari33 UCS ¼0.005Vp 0.94 Some samples of weak conglomeratic rock
Kahraman13 UCS ¼2.68e0.93 IS(50) 0.86 32 samples of pyroclastic rocks
Tonnizam Mohamad et al.34 UCS ¼0.032Vp  44.23 0.83 40 samples of soft rocks
Jahed Armaghani et al.14 UCS ¼0.0308Vp  61.61 0.47 45 samples of granitic rocks

Rn: Schmidt hammer Rebound Number; Is(50): Point load test; Vp: p-wave velocity; ρ: Density of the rock.

attempt made to estimate the UCS of rock using it or its hybrid. In to clay forming in their cement matrix. A range of (23.2–66.8 MPa)
the present study, several modeling techniques including LMR, was obtained for UCS results which can be classified as medium to
NLMR, ANN and hybrid ICA–ANN, have been conducted to predict strong according to ISRM.53 Also, obtained results can be classified
the UCS of rock by using rock properties including Rn, Vp, and Is(50). as very low to low, in accordance with Vp (1.57–3.06 km/s).54 The
Furthermore, developed models are compared with several per- value of Rn and Is(50) ranges from 19-43 and 1.23–4.15 MPa, re-
formance indices in accordance with their performance for practice. spectively. The results of Rn, Vp and Is(50) were utilized as input
variables for generating the predictive models by using relevant
modeling techniques. Afterward, the most acceptable and reliable
2. Case study and data construction model were chosen among them to introduce for practice.

An investigated area is located in Dengkil, Selangor, Malaysia,


where are about 35 km to the south of Kuala Lumpur and 13 km to 3. Modeling techniques
the north of Kuala Lumpur International Airport (KLIA). The area is
under active development, and in fact, the data used for this study 3.1. Artificial neural network (ANN)
was collected for verifying the excavability of rock. The area
composed of the Kenny Hill (KH) Formation52 that is Carbonifer- The ANN is first trained through processing numerous input
ous series. Sandstone, sedimentary, is the main rock type in the patterns and corresponding outputs. The network is capable of
formation; however, phyllite, slate and shale are also locally ob- recognizing similarities when they are presented with a new input
served in the field. parameter after appropriately predicting the output pattern. The
To obtain the aim of the study, more than 100 sandstone blocks ANNs can identify similarities in inputs, even though a certain
samples were taken from the field and brought to the laboratory. input might never have been recognized until that time. Because
Further, those blocks were cored and prepared to obtain the of this property, it has excellent interpolation capabilities, in par-
standard sample for individual test in accordance with Interna- ticular once input data is noisy (not exact).55
tional Society for Rock Mechanics.53 Further, prepared samples are An ANN should be trained prior to the interpretation of new
tested and the database is established. For each test, total 108 information. Although, there are several algorithms for training
samples are prepared and the test including Rn, Vp, Is(50), and UCS ANNs, back-propagation (BP) algorithm can be defined as the most
was performed on them. versatile technique among them.56 This algorithm makes available
Sandstone, type of rock studied herein, displays typical gran- the most effective learning procedure for the multilayer neural
ular texture with mineral grains size vary from 0.06 to 2 mm, and networks. Due to the mentioned fact, BP is a well-known algo-
well cemented. Further, typical petrographic study has shown that rithm to train ANNs.15,42,57
the sandstone is composed of 85% mineral quartz and 15% clay as Generally, the feed-forward BP includes three different layers
cement. The quartz is slightly fractured and exhibits sub-rounded i.e. input, hidden and output which are connected to each other. In
to angular shape grains, traces of feldspars were also noted in the fact, the outputs of neurons or nodes of the input layer are sent to
thin section of samples. nodes in hidden layer as input, and then by implementing similar
As a result of tests, the rock showed relatively low strength due procedure, they transfer to the last layer, which is output layer. The
176 D. Jahed Armaghani et al. / International Journal of Rock Mechanics & Mining Sciences 85 (2016) 174–186

n
E = 0. 5 ∑ (tk − Ok )2 (7)
k=1

The network training is a process of arriving at an optimum


weight space for the network. Using the following rule, the stee-
pest descent error surface is made:
⎛ δE ⎞
∇Wjk = − η ⎜⎜ ⎟⎟
⎝ δWjk ⎠ (8)

where η stands for the learning rate parameter and E signifies the
error function. The update of weights for the (n þ1)th pattern is
given as
Wjk(n+1) = Wjk(n) + □Wjk(n) (9)
Fig. 1. Back-propagation neural network.58 A comparable logic is applied to connections between the
hidden layers and output ones. For each pair of training case, this
number of hidden layer(s) and their neurons (nodes) are depen- procedure is repeated. Each process of the whole training patterns
is called as epoch. Then the process is repeated as many cycles as
dent on the given problem. Trial-and-error procedure is a common
required until the error is within specified goal of the user.
method to determine them.
Except the input layer, the whole nodes within the BP network
3.2. Imperialist competitive algorithm (ICA)
are accompanied with a transfer function and a bias node. The bias
is similar to a weight, except that it contains a constant input of 1,
The ICA was developed by Atashpaz-Gargari and Lucas46 to be
whereas the transfer function does a filtration on the summed utilized for solving various optimization problems. It has indicated
signals that are received from this neuron. Transfer functions are superior ability to obtain the global optima and convergence rate.59–
applied based on the purpose of ANN. Computed output vectors, 63
An individual in ICA are called countries and imperialists are
which are corresponding to the solution, are generated by output defined as some of the best countries and the remaining countries
layer. Typically, the input and output data are denoted as vectors organize the colonies of the imperialists. The basis for designing ICA
that are known as training pairs. As mentioned earlier, for the is the socio-politic competition among imperialists in real world for
whole training pairs in the data set, the process continues until the attracting more colonies. In ICA, those individuals that have the least
network error is converged to a threshold that is defined by a costs are considered as the more powerful imperialists, hence pos-
corresponding error function like root mean square error (RMSE). sessing more colonies. The implemented operators by ICA are
The jth node, in the hidden layer, is linked to several inputs, as named as assimilation, revolution, and competition.
shown in Fig. 1. Colonies in ICA during assimilation and revolution, can poten-
tially reach to a better state in comparison with those of corre-
xi =(x1, x2 , x3,…,x n) (1) sponding imperialists and are able to control the empire.61 Im-
Within the hidden layer, the net input values will be: perialists during competition try to adopt more colonies the whole
empires attempt to possess the colonies of other empires. After
n
that, weak empires crumble gradually and the empires with
Netj = ∑ xiwij + θj higher power can increase their power. The ICA procedure can be
i=1 (2)
explained in detail as follows:
where x i signifies the input units, wij represents the weights on the
connection of the ith input and jth neuron, θj denotes the bias node 3.2.1. Forming the initial empires
(optional) and n stands for the number of input units. A logarith- By generating a series of random solutions which are named as
mic sigmoid function computes the output from the hidden layer: countries, the ICA procedure is started. In an Nvar dimension op-
timization problem, a country is represented by a 1  Nvar array:
Oj = f (Netj )=1/[1 + e−(Netj+ θj)] (3)
country=[P1, P2, P3,…,P Nvar ] (10)
The total input to the kth unit is
where pi represents numerous parameters e.g. economy, culture,
n
language, and religion. Subsequently, through the evaluating the
Netk = ∑ wjkOj + θk
j=1 (4)

where θk signifies the bias node, and wjk denotes the weight be-
tween the jth neuron and the kth output. As a result, the total
output from kth unit will be:
Ok = f (Netk ) (5)

In ANN systems, by using weights and thresholds, networks


calculate their own output patterns. Then, a comparison is made
between the actual output and the desired output. As a result, at
any output in layer k, the error is:
eI = tk − Ok (6)

where tk represents the desired output and Ok stands for the real
output. The following equation is shown function of total error: Fig. 2. Schematic of forming initial imperialists and colonies.46
D. Jahed Armaghani et al. / International Journal of Rock Mechanics & Mining Sciences 85 (2016) 174–186 177

Fig. 3. Assimilation procedure in ICA.46

function f in variables (P1, P2, P3,… ,PNvar ), cost of countries can be


achieved as:
costi = f (countryi) = f (P1, P2, P3,…,P Nvar ) (11)

To produce N countries (Ncountry), a certain number of countries


with lowest costs are selected as imperialists (Nimp) and other
countries (colonies) are signified by Ncol. It should be noted that,
Ncountry and Nimp are choices in the hands of designer. According to
the initial power of empires, all colonies are distributed among
them, in this way, the normalized cost can be achieved as follows:
Cn = maxi{ci} − cn (12)

where cn is the cost of the nth imperialist, max i{ci} is the max-
imum cost among imperialists and Cn is the normalized cost of the
Fig. 5. ICA flowchart.46
nth imperialist. Considering the normalized costs of each empire,
pn, the parameter of imperialist normalized power can be obtained 3.2.2. Assimilation and competition procedure
based on: Colonies are attracted to their corresponding imperialists in the
process of assimilation as shown schematically in Fig. 3. Based on
Cn
pn = this figure, the distances between imperialists and their colonies
Nimp
∑i = 1 Ci (13) are signified by d and the colonies move toward their imperialist
by x units, where x is a random number with uniform distribution
The initial number of colonies, i.e. the Nth imperialist, is cal- as given by
culated by
x □U (0,β × d) (15)
N . C .n = round{pn . (Ncol )} (14)
where β can be considered as 2 (according to many researchers).
where N.C.n stands for the number of colonies for the nth im- In addition, for enhancing the search properties of the algorithm,
perialist. Considering the empires’ N.C. number and using a ran- the deviation in the movement of a colony toward its corre-
dom process, colonies are distributed among various empires. sponding imperialist is shown in Fig. 3. Parallel to the previous
Formation of initial empires and their colonies are displayed in parameter (i.e., x), θ is a random parameter with uniform dis-
Fig. 2. In Fig. 2, more colonies can be adopted by the powerful tribution:
imperialists or bigger stars. θ□U (−γ , γ ) (16)

where γ denotes a parameter whose values are determinant of the


amount of deviation. According to the original study of ICA [46], a
value of π/4 for θ can get satisfactory result. The total power of
empire is computed as follows:

T . C .n = Cost(imperialist n) + ξ
mean{Cost(colonies of empiren)} (17)

where T.C.n signifies the total cost of the n empire and ξ re-
th

presents a value very close to zero.


As illustrated in Fig. 4, the weakest empire is represented as
Empire 1 and, Empires 2 to N strive to attract the weakest colony
of Empire 1. Attracting probability of the target colony can be
determined based on computing the normalized total power of the
nth empire N.T.C.n:
N . T . C .n = maxi{T . C .i } − T . C .n (18)
Fig. 4. Competition among empires to attract the colony of weakest empire.46
178 D. Jahed Armaghani et al. / International Journal of Rock Mechanics & Mining Sciences 85 (2016) 174–186

2. Then random vector R, with a size similar to vector P, is created.


the arrays of vector R are the random numbers between [0–1]
with uniform distribution.
R = [r1, r2, r3,…,rNimp] (21)
3. Afterwards, vector D is generated as follows

D = P − R = [D1, D2 , D3, …, D Nimp]

= [Pp1 − r1, Pp2 − r2, Pp3 − r3, …, Pp N − rNimp]


imp (22)

At the end of ICA procedure, weak empires collapse one after


another until only one empire remains, which possesses all
countries as its colonies. The flowchart of the ICA algorithm is il-
lustrated in Fig. 5.
So, besides simple regression (SR), LMR, NLMR and ANN
Fig. 6. Relationship between Schmidt hammer rebound number and UCS.
models, hybrid model (ICA–ANN) was developed to obtain the best
prediction model for estimating the UCS of rocks herein. Detail
step and progress of purposed models are given in following
section.

4. Application of techniques for solving problem

A series of SR analyses are performed in the first stage of the


model construction. These analyses were conducted between the
UCS and individual parameter including Rn, Vp and Is(50). The result
obtained from SR analysis is not good enough to be used for sol-
ving the problem. Due to that, to propose the better model for
estimating the UCS of rock, linear and non-linear multiple re-
gression analysis, ANN and hybrid model (ICA–ANN) techniques
were also performed. The modeling process of each approach is
described herein.
Fig. 7. Relationship between p-wave velocity and UCS.

4.1. Simple regression analysis (SR)

In order to examine the weight of each input parameter, the


simple regression analyses were performed between the UCS and
relevant parameters including Rn, Vp and Is(50), and then new
equations introduced for predicting the UCS of rock. To develop
equations with higher performance prediction, several SR analyses
e.g. exponential, linear, logarithmic and power were implemented.
To check the reliability of the developed equations, their coef-
ficient of determination (R2) values were obtained and compared.
As a result, the power relationships gives relatively better results
for predicting UCS as a function of Rn and Vp with correlation
coefficients of 0.445 and 0.484 respectively. Further, the linear
relationship between the UCS and Is(50) with correlation coefficient
of 0.521 provides slightly higher performance capacity to estimate
the UCS. The purposed relationships between the UCS and relevant
Fig. 8. Relationship between point load index and UCS.
parameters of the rock are given in Figs. 6–8. Moreover, the de-
veloped equations for predicting UCS using Rn, Vp and Is(50), are
presented in (Eqs. (23)–25), respectively. Although, the developed
N . T . C .n relationships are statistically meaningful, nevertheless, they are
Ppn = N not practically good enough for estimating UCS. These relation-
∑i =imp
1 N. T . C .i (19) ships showed that multi input parameters are needed to estimate
Through the following procedure, the above-mentioned co- the UCS of rock, because of that, various modeling techniques such
lonies are allocated to the other empires: as linear and non-linear multiple regression analysis, ANN and
ICA–ANN were also constructed and performed.
1. Vector P of size 1*Nimp is shaped based on the above-mentioned UCS = 3.002R n0.801 (R2 = 0.445) (23)
probabilities;
P = [Pp1, Pp2, Pp3,…,P PNimp] (20) UCS = 17.783Vp1.099 (R2 = 0.484) (24)
D. Jahed Armaghani et al. / International Journal of Rock Mechanics & Mining Sciences 85 (2016) 174–186 179

Table 2
The developed LMR and NLMR equations for estimating UCS of the rock.

Model Dataset no. Developed relationships Training R2 Testing R2

LMR 1 UCS = 0. 422 × R n + 6. 357 × Vp + 7. 224×Is(50) þ0.997 0.608 0.681


2 UCS = 0.333 × Rn + 8.624 × Vp + 5.984×Is(50) þ1.340 0.604 0.659
3 UCS = 0.277 × Rn + 8.401 × Vp + 6.881×Is(50) þ1.539 0.628 0.602
4 UCS = 0.447 × Rn + 7.944 × Vp + 6.487×Is(50) –1.490 0.622 0.625
5 UCS = 0.452 × Rn + 8.485 × Vp + 6.384×Is(50) þ3.092 0.650 0.452

NLMR 1 UCS = 0.915Rn0.835+6.356Vp1.016+7.045×Is(50)  1.744 0.608 0.684


2 UCS = 0.022Rn1.670+2.539Vp1.801+5.268×Is(50)+14.755 0.602 0.646
3 UCS = 0. 368R n0.957+13. 974Vp0.755+6. 5×Is(50) –5.563 0.628 0.609
4 UCS = 0.965Rn0.817+27.150Vp0.434+7.324×Is(50)  26.138 0.623 0.621
5 UCS = 2.306Rn 0.657
+14.334Vp 0.753
+6.075×Is(50)  17.477 0.651 0.445

Table 3
Results of R2 and RMSE for various ANN models.

Performance index Model no. Nodes in hidden layers Network result

Iteration 1 Iteration 2 Iteration 3 Iteration 4 Iteration 5 Averaged

Train Test Train Test Train Test Train Test Train Test Train Test

2
R 1 1 0.522 0.445 0.497 0.643 0.665 0.401 0.660 0.479 0.777 0.422 0.620 0.474
2 2 0.645 0.606 0.660 0.256 0.734 0.419 0.678 0.323 0.747 0.414 0.692 0.395
3 3 0.645 0.579 0.703 0.448 0.728 0.414 0.735 0.347 0.693 0.451 0.700 0.445
4 4 0.692 0.489 0.743 0.578 0.764 0.416 0.766 0.806 0.769 0.752 0.746 0.599
5 5 0.691 0.629 0.769 0.621 0.762 0.529 0.706 0.889 0.780 0.783 0.741 0.684
6 6 0.821 0.767 0.839 0.757 0.850 0.769 0.830 0.808 0.812 0.852 0.830 0.790
7 7 0.780 0.814 0.867 0.760 0.819 0.776 0.821 0.812 0.809 0.764 0.819 0.785

RMSE 1 1 0.210 0.189 0.182 0.162 0.168 0.201 0.158 0.203 0.145 0.190 0.173 0.189
2 2 0.175 0.167 0.160 0.310 0.155 0.204 0.161 0.239 0.138 0.193 0.158 0.223
3 3 0.172 0.177 0.152 0.181 0.148 0.196 0.158 0.241 0.154 0.187 0.157 0.196
4 4 0.17 0.193 0.155 0.179 0.151 0.194 0.152 0.133 0.144 0.137 0.154 0.167
5 5 0.165 0.171 0.112 0.175 0.153 0.183 0.169 0.085 0.138 0.126 0.147 0.148
6 6 0.128 0.148 0.103 0.151 0.099 0.133 0.115 0.132 0.119 0.086 0.113 0.130
7 7 0.135 0.127 0.096 0.148 0.109 0.142 0.121 0.134 0.115 0.129 0.115 0.136

UCS = 11.613IS (50) + 18.478 (R2 = 0.521) multiple regression models for predicting the UCS of rock as re-
(25)
commended in Refs. [38,67,68]. In the literature, various percen-
tages like 20%, 25% and a range of (20–30%) of whole datasets were
4.2. Linear and non-linear multiple regression techniques suggested for testing datasets.69–71 By considering these sugges-
tions, 80% (86 datasets) and 20% (22 datasets) of whole datasets
The regression analysis is a statistical tool that can be applied to were chosen as training and testing datasets, respectively. Note
examine the relationships between variables. In this technique, the that, to choose random data for developing predictive models, an
relationship between independent (predictor) variable and de- ANN code written by authors was used. Using the constructed
pendent (output) variable is systematically determined in the form datasets, five linear multiple equations have been developed as
of a function.64 By performing linear multiple regression (LMR) presented in Table 2.
technique, a linear multiple relationship between input and out- It is found that, the R2 values in ranges of (0.604–0.650) and
put parameters can be obtained, while non-linear multiple re- (0.452–0.681) for training and testing datasets, respectively were
gression (NLMR) is a technique to achieve a non-linear relation- obtained for LMR models. As a result, it is found that there is no
ship between these parameters. Many studies highlighted the use salient difference among the developed models. Further, NLMR
of multiple regression analysis to predict UCS of the rock as equations were also proposed to predict UCS considering the same
mentioned earlier. In these models, different rock index properties datasets used for obtaining the LMR equations. In order to ex-
such as porosity, density, P-wave velocity, Rn and Is(50) were con- amine the non-linear multiple relations between the relevant
sidered as predictors to predict the UCS of rock (e.g.16,17,19,65). parameters and the UCS, five different datasets were randomly
The simple relationship shows that the best relations between selected for training and testing to develop NLMR models to
the UCS and both Vp and and Rn is a power law; however, the evaluate the capability of the purposed model for predicting the
relation of the UCS with Is(50) is linear. Due to that, both linear and UCS. As result, the developed equations and obtained coefficient of
nonlinear multiple regression analyses are conducted herein by correlations for both training and testing is given in Table 2.
using the statistical package, SPSSv.11.5.66 As result, five different It is concluded that, the R2 values range from 0.602 to 0.651 for
datasets were selected randomly to train and test for developing training and 0.445–0.684 for testing of the NLMR models. As a
180 D. Jahed Armaghani et al. / International Journal of Rock Mechanics & Mining Sciences 85 (2016) 174–186

Table 4
Effects of various Ncountry on the performance of ICA–ANN models.

Model no. No., of Country Network result Ranking Total rank

Train Test Train Test

R2 RMSE R2 RMSE R2 RMSE R2 RMSE

1 25 0.607 0.170 0.661 0.155 3 4 10 9 26


2 50 0.657 0.159 0.536 0.182 9 9 7 5 30
3 75 0.658 0.162 0.361 0.185 10 8 2 4 24
4 100 0.656 0.164 0.48 0.169 8 6 4 7 25
5 150 0.656 0.151 0.532 0.216 8 11 6 1 26
6 200 0.693 0.149 0.396 0.21 12 12 3 2 29
7 250 0.612 0.168 0.667 0.131 5 5 11 11 32
8 300 0.665 0.154 0.465 0.208 11 10 3 3 27
9 350 0.626 0.163 0.632 0.171 6 7 8 6 27
10 400 0.652 0.162 0.51 0.166 7 8 5 8 28
11 450 0.588 0.179 0.834 0.098 2 3 12 12 29
12 500 0.611 0.168 0.641 0.136 4 5 9 10 28

result, it is found that there is not much difference among the were suggested by,75,81–85 respectively, in order to determine op-
obtained results using NLMR method as well. As seen from Table 2, timum number of hidden nodes based on the number of input and
both the LMR and NLMR indicate similar performance capacity for output parameters.
predicting the UCS of rock.
≤2 × Ni + 1 (27)

4.3. ANN Model


(Ni + N0)/2 (28)
The implemented process of the developed ANN model is
highlighted in this section. The datasets used for developing the
multiple regression analyses were also utilized for developing the
ANN herein. As a first step of ANN modeling, all datasets were 2+N0 × Ni+0. 5N0 × (N02 + Ni )−3
normalized using the following equation: Ni + N0 (29)

X norm=(X − X min)/(X max − X min) (26)

where X is the measured value, Xnorm represents the normalized 2Ni/3 (30)
value of the measured parameter, Xmin and Xmax are the minimum
and maximum values of the measured parameters in the dataset.
As reported by several researchers (e.g. 72,73), selecting the Ni × N0 (31)
appropriate training algorithm and also number of hidden neurons
is the most important part of ANN design. Many investigations
highlighted the successful application of Levenberg–Marquardt 2Ni (32)
(LM) training algorithm (e.g.15,74). Note that, hidden layer equal to
one can estimate almost all complicated problems.75–77 This was where Ni and N0 are the number of input and output neurons,
also successfully highlighted in the studies conducted by several respectively.
researchers.67,78 In this study, based on the presented the ANN architectures in
The determining neuron number(s) in the hidden layer is the Eq. (27)–(32), the number of hidden nodes can vary from 1 to 7. So,
most critical task of the ANN architecture.67,79,80 Eqs. (27)–(32) a trial-and-error method is required to determine the number of

Fig. 9. The effects of N decade on the performance of ICA–ANN models.


D. Jahed Armaghani et al. / International Journal of Rock Mechanics & Mining Sciences 85 (2016) 174–186 181

Table 5
Effects of different Nimp on the performance of ICA–ANN models.

Model no. No. of Imperialist Network result Ranking Total rank

Train Test Train Test

R2 RMSE R2 RMSE R2 RMSE R2 RMSE

1 5 0.614 0.167 0.714 0.144 1 1 2 2 6


2 10 0.663 0.157 0.723 0.132 2 2 3 4 11
3 15 0.775 0.113 0.761 0.102 3 3 5 7 18
4 20 0.852 0.096 0.844 0.092 7 7 7 8 29
5 25 0.937 0.068 0.921 0.067 13 13 12 11 49
6 30 0.902 0.082 0.922 0.064 10 11 13 13 47
7 35 0.917 0.085 0.892 0.089 11 10 8 9 38
8 40 0.931 0.072 0.908 0.069 12 12 10 10 44
9 45 0.883 0.091 0.911 0.066 9 9 11 12 41
10 50 0.854 0.094 0.896 0.092 8 8 9 8 33
11 55 0.829 0.104 0.735 0.128 6 5 4 5 20
12 60 0.811 0.103 0.703 0.139 5 6 1 3 15
13 65 0.786 0.110 0.802 0.107 4 4 6 6 20

hidden node(s). In this regard, a series of analyses were performed generating the previous models including SR, LMR, NLMR and ANN
and their results were obtained based on R2 and root mean square was also utilized for developing the hybrid model by means of
error (RMSE) as listed in Table 3. As indicated in this table, each ICA–ANN herein. The main step of ICA–ANN model can be given as
model was iterated five times. According to obtained average R2 ICA parameter, number of country, number of decade, number of
and RMSE values for both training and testing datasets, Model No. imperialist and network architecture. So, the procedure of ICA–
6 with hidden neurons of 6 outperforms the other constructed ANN is described deeply in this section.
models. So, the architecture of this model (3 ×6 ×1) was chosen as
the proposed architecture of ANN model. Discussion about se- 4.4.1. ICA parameter
lecting the best ANN models will be given later. The most influential parameters on ICA are number of country
(Ncountry), number of imperialism (Nimp), number of decade
4.4. ICA–ANN model (Ndecade), β, θ and ζ. In the case of β, θ and ζ, a brief review of
previous studies is needed to determine them. A value of 2 for β
As mentioned earlier, ANN error can be minimized by ICA by was suggested in Ref. [46] and it was successfully implemented in
adjusting weights and biases of ANN. Due to that, a hybrid pre- the literature.49,51,86 So, Niknam et al.87 recommended a value
dictive model (ICA–ANN) was proposed to predict the UCS of rocks greater than 1 for β at the end of their research. Atashpaz-Gargari
to get rid of those disadvantages of the ANN. The dataset used for and Lucas46 introduced a value of π/4 for θ in the initial version of

Table 6
Results of performance prediction of the developed models together with their total ranks.

Method Model no. Network results Ranking Total rank

Train Test Train Test

R2 RMSE VAF R2 RMSE VAF R2 RMSE VAF R2 RMSE VAF

LMR 1 0.608 7.494 60.787 0.681 6.351 67.753 2 2 2 5 4 5 20


2 0.604 6.973 60.421 0.659 8.313 64.006 1 5 1 4 1 4 16
3 0.628 7.074 62.771 0.602 6.807 59.375 4 4 4 2 3 2 19
4 0.622 7.607 62.239 0.625 5.716 60.036 3 1 3 3 5 3 18
5 0.650 7.196 65.028 0.452 7.547 36.686 5 3 5 1 2 1 17

NLMR 1 0.608 7.494 60.191 0.684 6.321 60.068 2 2 1 5 4 4 18


2 0.602 6.992 60.205 0.646 8.447 62.809 1 5 2 4 1 5 18
3 0.628 7.071 62.804 0.609 7.943 59.884 4 4 4 2 2 3 19
4 0.623 7.599 62.315 0.621 5.814 58.633 3 1 3 3 5 2 17
5 0.651 7.189 65.101 0.445 7.551 36.516 5 3 5 1 3 1 18

ANN 1 0.821 5.063 82.114 0.767 5.714 73.795 2 2 2 2 4 1 13


2 0.839 4.659 83.925 0.757 6.792 74.810 4 4 4 1 1 2 16
3 0.850 4.492 85.001 0.769 6.093 76.386 5 5 5 3 2 3 23
4 0.830 4.750 82.675 0.808 5.993 79.446 3 3 3 4 3 4 20
5 0.812 5.386 80.939 0.858 3.784 84.344 1 1 1 5 5 5 18

ICA–ANN 1 0.940 2.919 93.899 0.930 2.911 92.942 3 2 2 2 3 2 14


2 0.945 2.649 94.501 0.948 3.245 94.815 4 3 3 4 1 4 19
3 0.949 2.602 94.769 0.940 2.997 93.915 5 5 4 3 2 3 22
4 0.938 3.203 92.698 0.951 2.802 95.090 2 1 1 5 4 5 18
5 0.949 2.628 94.904 0.929 2.473 92.518 5 4 5 1 5 1 21
182 D. Jahed Armaghani et al. / International Journal of Rock Mechanics & Mining Sciences 85 (2016) 174–186

Fig. 10. Predicted UCS values by LMR against the measured ones.

Fig. 11. Predicted UCS values by NLMR against the measured ones.

Fig. 12. Predicted UCS values by ANN against the measured ones.
D. Jahed Armaghani et al. / International Journal of Rock Mechanics & Mining Sciences 85 (2016) 174–186 183

Fig. 13. Predicted UCS values by ICA–ANN against the measured ones.

the ICA. Later, Marto et al.50 and Ebrahimi et al.88 reported that a of sensitivity analyses (RMSE) on the Ndecade for various Ncountry
successful implementation of using this value. Moreover, a range ranging from 25 to 500. According to Fig. 9, significant changes
of 0–1 for θ was suggested in Ref. [89] In the case of ζ, several happened in the first decades for all Ncountry whereas, the changes
values have been recommended such as 0.1, 0.05 and 0.02 were moderate up to Ndecade ¼400. It is clear that no changes can
by,46,49,51 respectively. Based on the discussed literature, con- be seen in results of ICA–ANN network after Ndecade ¼400, there-
ducting a series of sensitivity analyses, for β, θ and ζ, values of 2, π/ fore, 400 was set as optimum Ndecade. It should be mentioned that
4 and 0.02 were chosen to be implemented, respectively. in determining number of decade, Nimp was set as 5 for all pur-
posed ICA–ANN models.
4.4.2. Number of country
In the next step of model procedure, number of country 4.4.4. Number of imperialists
(Ncountry) was determined. Various values of Ncountry have used to To determine the proper Nimp, several Nimp values in a range of
solve the engineering problems by several researchers. The value (5–65) were used. Performance indices including the R2 and RMSE
of 40 was recommended in Ref. [86] to estimate oil flow rate of the of ICA–ANN models with different Nimp values for training and
reservoir. Marto et al.50 and Hajihassani et al.51 suggested testing datasets are shown in Table 5. Similar to previous section, a
Ncountry ¼ 56 and Ncountry ¼135 for prediction of flyrock and ground simple ranking method was used to select the best Nimp. Based on
vibration induced by blasting, respectively. Based on results of values obtained for total ranks, it is found that imperialist number
these studies, it seems that a parametric study is needed to obtain of model 5 (Nimp ¼25) shows better network performance capacity
the proper Ncountry for solving the problem herein. In this regard, in comparison with the other number of imperialists. So, value of
12 ICA–ANN models with different Ncountry ranging from 25 to 500 25 was selected as optimum Nimp. In other word, the best perfor-
were constructed as shown in Table 4. In these models, Ndecade and mance of ICA–ANN models observed when Nimp to Ncountry ratio is
Nimp were set as 200 and 5, respectively. Moreover, suggested ANN 0.1.
architecture (3  6  1) was used to train the hybrid ICA–ANN
system. In Table 4, the obtained performance indices of the net- 4.4.5. Network architecture
work namely R2 and RMSE are presented for each model. Using the suggested ANN architecture (3  6  1) and obtained
As know that, the model is excellent if the R2 is one and RMSE ICA parameters (such as; Ncountry ¼ 250, Ndecade ¼ 400, Nimp ¼ 25),
is zero. In order to obtain the optimum Ncountry, a simple ranking five ICA–ANN models were trained as a last step of ICA–ANN
method proposed in Ref. [38] was applied herein. For example, model. Furthermore, testing datasets were also utilized in order to
there were 5 models developed and performance indices were evaluate capacity of the network performance. It is important to
obtained as herein, those indices were categorized according to express that all models of ANN and ICA–ANN in this study were
the values obtained for each index from the lowest (1) to highest constructed using MatLab version 7.14.0.739.90 The relevant results
(5). Then, the obtained ranking number of performance indices for of five trained ICA–ANN models were discussed and evaluated
each model was summed; and then the model having the highest according to their accuracy and applicability in practice in next
summed ranking values is considered as the best one among the section.
models (Table 4). This procedure was repeated to testing and
training data and the obtained values compared for each model
and total ranking values is given at the end of analysis as given in 5. Comparing the performance of the models
the last column of the Table 4. Results showed that Ncountry ¼250
with total rank of 32 is the best among all number of countries. This section presents evaluation of capacity performance of the
Hence, this value was chosen and utilized as optimum Ncountry. developed models for prediction of the UCS. By performing simple
regression, it was found that there is a need to develop UCS pre-
4.4.3. Number of decade dictive models with higher degree of accuracy using multi-input
Another parametric study is needed to investigate the effect of parameters. Hence, various models namely LMR, NLMR, ANN, and
various Ndecade on the performance of network. The analysis was ICA–ANN were proposed to predict UCS of sandstone samples. In
conducted by setting fixed Ndecade ¼1000. Fig. 9 displays the results these models, results of laboratory study i.e., Rn, Vp and Is(50), were
184 D. Jahed Armaghani et al. / International Journal of Rock Mechanics & Mining Sciences 85 (2016) 174–186

used to predict UCS of the rock. Furthermore, all data (108 data- have relationship with coefficients of correlation ranges from 0.43
sets) were randomly selected to five different datasets for devel- to 0.52. However, reliability of obtained results is not good enough
opment of the UCS predictive models. Performance indices of for predicting the UCS of rock itself. Since each input parameter
RMSE, R2 and variance account for (VAF) were calculated to eval- has good relationship with the UCS of rock, LMR and NLMR models
uate the performance prediction of developed models: were also conducted to achieve the better result. Furthermore,
N
conventional ANN and hybrid ICA–ANN model were constructed to
∑i = 1 (y − y′)2 predict the UCS of rock. In order to develop the hybrid model and
R2=1− N
∑i = 1 (y −ỹ )2 (33) for optimization of the weights and biases of the network con-
nection for training, ICA algorithm is integrated with ANN.
Further, five different datasets to train and test were estab-
var(y − y′) lished randomly for obtaining the best models. After developing
VAF = [1− ] × 100 the models, various performance indices including R2, VAF and
var(y) (34)
RMSE were utilized to evaluate the accuracy and efficiency of
these models. Further, based on simple ranking method, total rank
N
values of all predictive models were computed and compared to
1 each other to select the best prediction model.
RMSE= ∑ (y − y′)2
N i=1 (35) It is found that the highest correlation coefficient (R2 ¼ 0.949)
was achieved by implementing the ICA–ANN model, while the
where y, y′ and y~ are the measured, predicted and mean of the y other developed models including ANN, NLMR and LMR shows R2
values, respectively, N is the total number of data and P is the as 0.85, 0.628 and 0.608 in order. Moreover, the same trend was
number of predictors. Theoretically, a predictive model with R2 of observed for testing datasets of the developed models.
unity, VAF of 100, and RMSE of zero, is considered to be excellent. Apart from R2, it is also found that the ICA–ANN model has the
Table 6 presents the results of performance prediction (VAF, highest VAF and RMSE indices as comparing the others. As con-
RMSE and R2) for all developed predictive models in approx- sidering the total ranking value of the models, it is concluded that
imating UCS of the rock. In Table 6, as a result of developed model, the ANN, and ICA–ANN techniques having Model 3 shows better
the obtained performance indices are very similar; hence, select- performance in comparison with other techniques and models.
ing the best models among them is not easy. It is quite compli- The performance indices obtained by the ICA–ANN model re-
cated and difficult to decide the best model by using only R2, RMSE veal the high reliability of the new predictive model for estimating
or VAF. As mentioned earlier, a simple method suggested in Ref. UCS of sandstone. Although, all predictive models can be used for
[38] was applied in order to select the best models. The procedure UCS prediction, nevertheless, the most accurate results can be
of this method in detail can be found in the original study.38 It is achieved by developing hybrid ICA–ANN model. It is important to
important to note that in order to have a better comparison be- mention that the proposed hybrid model is designed based on the
tween results, the obtained normalized results of ANN and ICA– compressive strength of sandstone rock samples; hence, the de-
ANN models were converted to the pre-normalized data. veloped model should be used for similar rock type with caution.
According to Table 6, it is found that the model 1, 3, 3 and
3 with total rank values of 20, 19, 23 and 22 exhibited the best
result for estimating the UCS for LMR, NLMR, ANN and ICA–ANN Acknowledgment
techniques, respectively. Based on obtained results, the prediction
performances of the hybrid ICA–ANN models are significantly The data on rock properties highlighted in this paper is part of
higher than those of other developed models. The selected LMR the study on assessing excavatability of clastic sedimentary rocks
and NLMR equations in predicting the UCS are shown in (Eqs. (36) in Malaysia. The study was funded by Construction Research In-
and 37), respectively: stitute of Malaysia (CREAM) and managed by Research Manage-
UCS=0. 422×R n+6. 357 × Vp+7. 224×Is(50) + 0. 997 ment Centre (RMC), Universiti Teknologi Malaysia (UTM) that the
(36)
authors were appreciated.

UCS=0. 368R n0.957+13. 974Vp0.755+6. 50×Is(50) –5. 563 (37)


References
The developed relationships between the predicted UCS and
the measured one using LMR, NLMR, ANN and ICA–ANN models
1. Aufmuth RE. A systematic determination of engineering criteria for rocks. Bull
are given in Figs. 10–13, respectively. Based on these conclusions, Assoc Eng Geol. 1973;11:235–245.
the best UCS predictive model is the ICA–ANN method with R2 2. Singh RN, Hassani FP, Elkington PAS. The application of strength and de-
values of 0.949 and 0.940 for training and testing data, respec- formation index testing to the stability assessment of coal measures excava-
tions. In: Proceedings of the 24th US Rock Mechanics Symposium. Texas A&M
tively compared to others including LMR, NLMR and ANN models. University, AEG; 1983:599–609.
3. Cargill JS, Shakoor A. Evaluation of empirical methods for measuring the uni-
axial compressive strength of rock. Int J Rock Mech Min Sci. 1990;27:495–503.
4. Tugrul A, Zarif IH. Correlation of mineralogical and textural characteristics with
6. Conclusions engineering properties of selected granitic rocks from Turkey. Eng Geol.
1999;51:303–317.
In this research, a series of laboratory tests including Schmidt 5. Sulukcu S, Ulusay R. Evaluation of the block punch index test with particular
reference to the size effect, failure mechanism and its effectiveness in pre-
hammer porosity, P-wave velocity, point load index and UCS were dicting rock strength. Int J Rock Mech Min Sci. 2001;38:1091–1111.
conducted on sandstone collected from Dengkil site in Malaysia. 6. Kahraman S. Evaluation of simple methods for assessing the uniaxial com-
Total 108 samples were prepared and tested according to the pressive strength of rock. Int J Rock Mech Min Sci. 2001;38:981–994.
7. Yasar E, Erdogan Y. Estimation of rock physiomechanical properties using
standards. Data composed of rock properties tests including Rn, Vp hardness methods. Eng Geol. 2004;71:281–288.
and Is(50) was used for development of linear and non-linear 8. Basu A, Aydin A. Predicting uniaxial compressive strength by point load test:
models for predicting the UCS of sandstone. significance of cone penetration. Rock Mech Rock Eng. 2006;39:483–490.
9. Sharma PK, Singh TN. A correlation between P-wave velocity, impact strength
Firstly, various simple linear regression analyses were per- index, slake durability index and uniaxial compressive strength. Bull Eng Geol
formed; it was found that the UCS and relevant rock properties Environ. 2008;67:17–22.
D. Jahed Armaghani et al. / International Journal of Rock Mechanics & Mining Sciences 85 (2016) 174–186 185

10. Kilic A, Teymen A. Determination of mechanical properties of rocks using Comput. 2011;11:2587–2594.
simple methods. Bull Eng Geol Environ. 2008;67(2):237–244. 40. Yagiz S, Sezer EA, Gokceoglu C. Artificial neural networks and nonlinear re-
11. Yagiz S. P-wave velocity test for the assessment of some geotechnical properties gression techniques to assess the influence of slake durability cycles on the
of rock materials. Bull Mat Sci. 2009;34:947–953. prediction of uniaxial compressive strength and modulus of elasticity for car-
12. Yilmaz I, Yuksek G. Prediction of the strength and elasticity modulus of gypsum bonate rocks. Int J Numer Anal Met. 2012;36:1636–1650.
using multiple regression, ANN, and ANFIS models. Int J Rock Mech Min Sci. 41. Mishra DA, Basu A. Estimation of uniaxial compressive strength of rock mate-
2009;46(4):803–810. rials by index tests using regression analysis and fuzzy inference system. Eng
13. Kahraman S. The determination of uniaxial compressive strength from point Geol. 2013;160:54–68.
load strength for pyroclastic rocks. Eng Geol. 2014;170:33–42. 42. Momeni E, Jahed Armaghani D, Hajihassani M, Amin MFM. Prediction of uni-
14. Jahed Armaghani D, Tonnizam Mohamad E, Momeni E, Narayanasamy MS, axial compressive strength of rock samples using hybrid particle swarm opti-
Mohd Amin MF. An adaptive neuro-fuzzy inference system for predicting un- mization-based artificial neural networks. Measurement. 2015;60:50–63.
confined compressive strength and Young's modulus: a study on Main Range 43. Ceryan N, Okkan U, Kesimal A. Prediction of unconfined compressive strength
granite. Bull Eng Geol Environ. 2014, http://dx.doi.org/10.1007/ of carbonate rocks using artificial neural networks. Environ Earth Sci. 2012;68
s10064-014-0687-4. (3):807–819.
15. Meulenkamp F, Alvarez Grima M. Application of neural networks for the pre- 44. Beiki M, Majdi A, Givshad AD. Application of genetic programming to predict
diction of the unconfined compressive strength (UCS) from Equotip hardness. the uniaxial compressive strength and elastic modulus of carbonate rocks. Int J
Int J Rock Mech Min Sci. 1999;36(1):29–39. Rock Mech Min Sci. 2013;63:159–169.
16. Gokceoglu C, Zorlu K. A fuzzy model to predict the unconfined compressive 45. Jahed Armaghani D, Mohamad ET, Hajihassani M, Yagiz S, Motaghedi H. Ap-
strength and modulus of elasticity of a problematic rock. Eng Appl Artif Intell. plication of several non-linear prediction tools for estimating uniaxial com-
2004;17:61–72. pressive strength of granitic rocks and comparison of their performances. Eng
17. Yagiz S. Predicting uniaxial compressive strength, modulus of elasticity and Comput. 2015, http://dx.doi.org/10.1007/s00366-015-0410-5.
index properties of rocks using the Schmidt hammer. Bull Eng Geol Environ. 46. Atashpaz-Gargari E, Lucas C. Imperialist competitive algorithm: an algorithm
2009;68(1):55–63. for optimization inspired by imperialistic competition. In: Proceedings of the
18. Dehghan S, Sattari GH, Chehreh CS, Aliabadi MA. Prediction of unconfined IEEE Congr Evol Comput; 2007:4661–4667.
compressive strength and modulus of elasticity for Travertine samples using 47. Kaveh A, Talatahari S. Optimum design of skeletal structures using imperialist
regression and artificial neural. New Min Sci Technol. 2010;20:0041–0046. competitive algorithm. Comput Struct. 2010;88:1220–1229.
19. Yesiloglu-Gultekin N, Gokceoglu C, Sezer EA. Prediction of uniaxial compressive 48. Nazari-Shirkouhi S, Eivazy H, Ghodsi R, Rezaie K, Atashpaz-Gargari E. Solving
strength of granitic rocks by various nonlinear tools and comparison of their the integrated product mix-outsourcing problem using the imperialist com-
performances. Int J Rock Mech Min Sci. 2013;62:113–122. petitive algorithm. Expert Syst Appl. 2010;37(12):7615–7626.
20. Rezaei M, Majdi A, Monjezi M. An intelligent approach to predict unconfined 49. Taghavifar H, Mardani A, Taghavifar L. A hybridized artificial neural network
compressive strength of rock surrounding access tunnels in longwall coal and imperialist competitive algorithm optimization approach for prediction of
mining. Neural Comput Appl. 2014;24(1):233–241. soil compaction in soil bin facility. Measurement. 2013;46:2288–2299.
21. Sachpazis CI. Correlating Schmidt hardness with compressive strength and 50. Marto A, Hajihassani M, Jahed Armaghani D, Tonnizam Mohamad E, Makhtar
Young's modulus of carbonate rocks. Bull Int Assoc Eng Geol. 1990;42:75–83. AM. A novel approach for blast-induced flyrock prediction based on imperialist
22. Xu S, Grasso P, Mahtab A. Use of Schmidt hammer for estimating mechanical competitive algorithm and artificial neural network. Sci World J. 2014 Article ID
properties of weak rock. In: Proceeding of 6th International IAEG Congress. Rot- 643715.
terdam: Balkema; 1990:511–519. 51. Hajihassani M, Jahed Armaghani D, Marto A, Tonnizam Mohamad E. Ground
23. Tsiambaos G, Sabatakakis N. Considerations on strength of intact sedimentary vibration prediction in quarry blasting through an artificial neural network
rocks. Eng Geol. 2004;72:261–273. optimized by imperialist competitive algorithm. Bull Eng Geol Environ. 2014,
24. Entwisle DC, Hobbs RN, Jones LD, Gunn D, Raines MG. The relationship be- http://dx.doi.org/10.1007/s10064-014-0657-x.
tween effective porosity, uniaxial compressive strength and sonic velocity of 52. Hutchinson CS, Tan DNK. Geology of Peninsular Malaysia. University of Malaya &
intact Borrowdale volcanic group core samples from Sellafield. Geotech Geol The Geological Society of Malaysia; 2009:479.
Eng. 2005;23:793–809. 53. ISRM, The complete ISRM suggested methods for rock characterization, testing
25. Kahraman S, Gunaydin O, Fener M. The effect of porosity on the relation be- and monitoring: 1974–2006. In: Ulusay and Hudson (eds.) Suggested Methods
tween uniaxial compressive strength and point load index. Int J Rock Mech Min Prepared by the Commission on Testing Methods, International Society for
Sci. 2005;42:584–589. Rock Mechanics; 2007.
26. Yilmaz I, Yuksek AG. An example of artificial neural network (ANN) application 54. Anon. Classification of rocks and soils for engineering geological mapping; part
for indirect estimation of rock parameters. Rock Mech Rock Eng. 2008;41 1-rock and soil materials. Bull Int Assoc Eng Geol. 1979;19:364–371.
(5):781–795. 55. Simpson PK. Artificial Neural System: Foundation, Paradigms, Applications and
27. Khandelwal M, Singh TN. Correlating static properties of coal measures rocks Implementations. New York: Pergamon; 1990.
with P-wave velocity. Int J Coal Geol. 2009;79(1):55–60. 56. Khandelwal M, Singh TN. Prediction of blast-induced ground vibration using
28. Moradian ZA, Behnia M. Predicting the uniaxial compressive strength and static artificial neural network. Int J Rock Mech Min Sci. 2009;46(7):1214–1222.
Young's modulus of intact sedimentary rocks using the ultrasonic test. Int J 57. Jahed Armaghani D, Momeni E, Abad SVANK, Khandelwal M. Feasibility of
Geomech. 2009;9:1–14. ANFIS model for prediction of ground vibrations resulting from quarry blasting.
29. Diamantis K, Gartzos E, Migiros G. Study on uniaxial compressive strength, Environ Earth Sci. 2015, http://dx.doi.org/10.1007/s12665-015-4305-y.
point load strength index, dynamic and physical properties of serpentinites 58. Monjezi M, Dehghani H. Evaluation of effect of blasting pattern parameters on
from Central Greece: test results and empirical relations. Eng Geol. back break using neural networks. Int J Rock Mech Min Sci. 2008;45(8):1446–
2009;108:199–207. 1453.
30. Mishra DA, Basu A. Use of the block punch test to predict the compressive and 59. Sepehri Rad H, Lucas C. Application of imperialistic competition algorithm in
tensile strengths of rocks. Int J Rock Mech Min Sci. 2012;51:119–127. recommender systems. In: Proceeding of the 13th International CSI computer
31. Kohno M, Maeda H. Relationship between point load strength index and uni- conference (CSICC’08), Kish, Iran; 2008.
axial compressive strength of hydrothermally altered soft rocks. Int J Rock Mech 60. Atashpaz-Gargari E, Hashemzadeh F, Rajabioun R, Lucas C. Colonial competitive
Min Sci. 2012;50:147–157. algorithm, a novel approach for PID controller design in MIMO distillation
32. Khandelwal M. Correlating P-wave velocity with the physico-mechanical column process. Int J Intell Comput Cybern. 2008;1:337–355.
properties of different rocks. Pure Appl Geophys. 2013;170:507–514. 61. Rajabioun E, Atashpaz-Gargari E, Lucas C. Colonial competitive algorithm as a
33. Minaeian B, Ahangari K. Estimation of uniaxial compressive strength based on tool for Nash equilibrium point achievement. In: Computational science and its
P-wave and Schmidt hammer rebound using statistical method. Arab J Geosci. applications-iccsa 2008, Springer; 2008:680–695.
2013;6(6):1925–1931. 62. Jahed Armaghani D, Hajihassani M, Marto A, Faradonbeh RS, Mohamad ET.
34. Tonnizam Mohamad E, Armaghani DJ, Momeni E, Alavi Nezhad, Khalil Abad SV. Prediction of blast-induced air overpressure: a hybrid AI-based predictive
Prediction of the unconfined compressive strength of soft rocks: a PSO-based model. Environ Monit Assess. 2015;187(11):1–13.
ANN approach. Bull Eng Geol Environ. 2014, http://dx.doi.org/10.1007/ 63. Jahed Armaghani D, Hasanipanah M, Mohamad ET. A combination of the ICA–
s10064-014-0638-0. ANN model to predict air overpressure resulting from blasting. Eng Comput.
35. Alvarez Grima M, Babuška R. Fuzzy model for the prediction of unconfined 2016;32:155–171.
compressive strength of rock samples. Int J Rock Mech Min Sci. 1999;36(3):339– 64. Bahrami A, Monjezi M, Goshtasbi K, Ghazvinian A. Prediction of rock frag-
349. mentation due to blasting using artificial neural network. Eng Comput. 2011;27
36. Singh VK, Singh D, Singh TN. Prediction of strength properties of some schis- (2):177–181.
tose rocks from petrographic properties using artificial neural networks. Int J 65. Monjezi M, Khoshalan HA, Razifard M. Aneuro-genetic network for predicting
Rock Mech Min Sci. 2001;38:269–284. uniaxial compressive strength of rocks. Geotech Geol Eng. 2012;30(4):1053–
37. Sonmez H, Tuncay E, Gokceoglu C. Models to predict the uniaxial compressive 1062.
strength and the modulus of elasticity for Ankara Agglomerate. Int J Rock Mech 66. SPSS Inc. SPSS for Windows (Version 16.0). Chicago: SPSS Inc; 2007.
Min Sci. 2004;41(5):717–729. 67. Yagiz S, Gokceoglu C, Sezer E, Iplikci S. Application of two non-linear prediction
38. Zorlu K, Gokceoglu C, Ocakoglu F, Nefeslioglu HA, Acikalin S. Prediction of tools to the estimation of tunnel boring machine performance. Eng Appl Artif
uniaxial compressive strength of sandstones using petrography-based models. Intell. 2009;22(4):808–814.
Eng Geol. 2008;96(3):141–158. 68. Jahed Armaghani D, Mohamad ET, Hajihassani M, Abad SANK, Marto A, Mo-
39. Cevik A, Sezer EA, Cabalar AF, Gokceoglu C. Modeling of the uniaxial com- ghaddam MR. Evaluation and prediction of flyrock resulting from blasting
pressive strength of some clay-bearing rocks using neural network. Appl Soft operations using empirical and computational methods. Eng Comput. 2015,
186 D. Jahed Armaghani et al. / International Journal of Rock Mechanics & Mining Sciences 85 (2016) 174–186

http://dx.doi.org/10.1007/s00366-015-0402-5. 80. Sonmez H, Gokceoglu C. Discussion on the paper by H. Gullu and E. Ercelebi A
69. Swingler K. Applying Neural Networks: A Practical Guide. New York: Academic neural network approach for attenuation relationships: An application using
Press; 1996. strong ground motion data from Turkey. Eng Geol. 2008;97:91–93.
70. Looney CG. Advances in feed-forward neural networks: demystifying knowl- 81. Ripley BD. Statistical aspects of neural networks. In: Barndoff- Neilsen OE,
edge acquiring black boxes. IEEE Trans Knowl Data Eng. 1996;8(2):211–226. Jensen JL, Kendall WS, eds. Networks and Chaos-statistical and Probabilistic As-
71. Nelson M, Illingworth WT. A Practical Guide to Neural Nets. Addison-Wesley, pects. London: Chapman & Hall; 1993:40–123.
Reading Mass; 1990. 82. Paola JD. Neural network classification of multispectral imagery. MSc thesis.
72. Hush DR. Classification with neural networks: a performance analysis. In: University of Arizona; 1994.
Proceedings of the IEEE International Conference on Systems Engineering. Dayton, 83. Wang C. A theory of generalization in learning machines with neural application.
OH, USA, 1989:277–280. Ph.D. thesis. University of Pennsylvania; 1994.
73. Kanellopoulas I, Wilkinson GG. Strategies and best practice for neural network 84. Masters T. Practical Neural Network Recipes in Cþ þ . Boston MA: Academic
image classification. Int J Remote Sens. 1997;18:711–725. Press; 1994.
74. Atici U. Prediction of the strength of mineral admixture concrete using multi- 85. Kaastra I, Boyd M. Designing a neural network for forecasting financial and
variable regression analysis and an artificial neural network. Expert Sys Appl. economic time series. Neurocomputing. 1996;10:215–236.
2011;38:9609–9618. 86. Ahmadi MA, Ebadi M, Shokrollahi A, Majidi SMJ. Evolving artificial neural
75. Hecht-Nielsen R. Kolmogorov's mapping neural network existence theorem. In: network and imperialist competitive algorithm for prediction oil flow rate of
Proceedings of the First IEEE International Conference on Neural Networks. San the reservoir. Appl Soft Comput. 2013;13(2):1085–1098.
Diego, CA, USA, 1987:11–14. 87. Niknam T, Taherian Fard E, Pourjafarian N, Rousta A. An efficient hybrid algo-
76. Hornik K, Stinchcombe M, White H. Multilayer feedforward networks are rithm based on modified imperialist competitive algorithm and K-means for
universal Approximators. Neural Netw. 1989;2:359–366. data clustering. Eng Appl Artif Intell. 2011;24(2):306–317.
77. Baheer I. Selection of methodology for modeling hysteresis behavior of soils 88. Ebrahimi E, Mollazade K, Babaei S. Toward an automatic wheat purity mea-
using neural networks. J Comput Aid Civil Infrastruct Eng. 2000;5(6):445–463. suring device: a machine vision-based neural networks-assisted imperialist
78. Gordan B, Jahed Armaghani D, Hajihassani M, Monjezi M. Prediction of seismic competitive algorithm approach. Measurement. 2014;55:196–205.
slope stability through combination of particle swarm optimization and neural 89. Abdechiri M, Faez K, Bahrami H. Neural network learning based on chaotic
network. Eng Comput. 2015, http://dx.doi.org/10.1007/s00366-015-0400-7. imperialist competitive algorithm. In: Proceedings of the 2nd International
79. Sonmez H, Gokceoglu C, Nefeslioglu HA, Kayabasi A. Estimation of rock mod- Workshop on Intelligent Systems and Applications (ISA), 2010. IEEE; 2010:1–5.
ulus: for intact rocks with an artificial neural network and for rock masses with 90. Demuth H, Beale M, Hagan M. MATLAB Version 7.14.0.739; Neural Network
a new empirical equation. Int J Rock Mech Min Sci. 2006;43:224–235. Toolbox for Use with Matlab. The Mathworks. 2009.

Potrebbero piacerti anche