Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1398e1412

Contents lists available at ScienceDirect

Journal of Rock Mechanics and

Geotechnical Engineering
journal homepage: www.jrmge.cn

Full Length Article

Hybrid ensemble soft computing approach for predicting penetration rate

of tunnel boring machine in a rock environment
Abidhan Bardhan a, *, Navid Kardani b, Anasua GuhaRay c, Avijit Burman a, Pijush Samui a,
Yanmei Zhang d, **
a
National Institute of Technology Patna, Patna, Bihar, India
b
Civil and Infrastructure Engineering Discipline, School of Engineering, Royal Melbourne Institute of Technology (RMIT), Victoria, 3001, Australia
c
Department of Civil Engineering, BITS-Pilani Hyderabad Campus, Hyderabad, Telangana, India
d
College of Aerospace Engineering, Chongqing University, Chongqing, 400044, China

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

Article history: This study implements a hybrid ensemble machine learning method for forecasting the rate of pene-
Received 21 April 2021 tration (ROP) of tunnel boring machine (TBM), which is becoming a prerequisite for reliable cost
Received in revised form assessment and project scheduling in tunnelling and underground projects in a rock environment. For
24 June 2021
this purpose, a sum of 185 datasets was collected from the literature and used to predict the ROP of TBM.
Accepted 24 June 2021
Available online 24 October 2021
Initially, the main dataset was utilised to construct and validate four conventional soft computing (CSC)
models, i.e. minimax probability machine regression, relevance vector machine, extreme learning ma-
chine, and functional network. Consequently, the estimated outputs of CSC models were united and
Keywords:
Tunnel boring machine (TBM)
trained using an artificial neural network (ANN) to construct a hybrid ensemble model (HENSM). The
Rate of penetration (ROP) outcomes of the proposed HENSM are superior to other CSC models employed in this study. Based on the
Artificial intelligence experimental results (training RMSE ¼ 0.0283 and testing RMSE ¼ 0.0418), the newly proposed HENSM is
Artificial neural network (ANN) potential to assist engineers in predicting ROP of TBM in the design phase of tunnelling and underground
Ensemble modelling projects.

2021 Institute of Rock and Soil Mechanics, Chinese Academy of Sciences. Production and hosting by
Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/
licenses/by-nc-nd/4.0/).

1. Introduction method is a well-developed technology, professionals are imple-

menting mechanical excavation in underground construction
In the field of civil engineering, there is an endless growing (Armaghani et al., 2017). A tunnel boring machine (TBM) is an
demand due to rapid industrialisation. The construction industry essential tool used for tunnelling in underground projects. TBMs
plays a vital role in the development of the country in the field of offer copious advantages over the drilling and blasting method
transportation, hydropower, mining, and underwater construction. (Javad and Narges, 2010). It should be noted that TBMs are gigantic
In several areas such as rail, metro, and hydropower, construction machines and during the course of excavation, TBMs are extremely
projects are underway across the world. In recent decades, rapid sensitive to conditions of rock mass throughout the path (Zhou
industrialisation is ongoing on a fast-track basis to reduce the et al., 2021). Though TBM is the best option for tunnelling in a
overall development costs and timely completion of projects. In rocky medium, insufficient information and unknown rock mass
most heavy and civil engineering projects, tunnelling works are conditions can lead to a decline in the safety and efficiency of the
now being carried out to avoid open excavation and blasting tunnelling operation. Therefore, proper planning and accurate
(Mahdevari et al., 2014). Even though the drilling and blasting estimation are required to adopt the best construction techniques
and to predict the performance of TBMs in a rock environment
(Zhou et al., 2021).
Prediction of TBM performance is an important factor to esti-
* Corresponding author.
mate project duration, project cost, project schedule, and pro-
** Corresponding author.
E-mail addresses: [email protected] (A. Bardhan), [email protected]
curement schedule of bulk materials (Yagiz et al., 2009).
(Y. Zhang). Theoretically, the performance of a TBM is measured in terms of
Peer review under responsibility of Institute of Rock and Soil Mechanics, Chi- rate of penetration (ROP), which is the ratio of excavation distance
nese Academy of Sciences.

https://doi.org/10.1016/j.jrmge.2021.06.015
1674-7755
2021 Institute of Rock and Soil Mechanics, Chinese Academy of Sciences. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-
NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
 A. Bardhan et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1398e1412 1399

to continuous operation time of TBM (Mahdevari et al., 2014). 2013; Armaghani et al., 2017; Koopialipoor et al., 2019; Xu et al.,
Several factors, such as intact rock properties, rock mass properties, 2019), ANFIS (Salimi et al., 2016), SVM/SVR (Martins and Miranda,
cutter geometry, and specification of TBMs, influence the perfor- 2013; Mahdevari et al., 2014; Salimi et al., 2016; Xu et al., 2019),
mance of TBMs (Zeng et al., 2021). Proper application of TBM to any least-squares support vector machine (LS-SVM) (Zhang et al.,
tunnelling projects depends on the parameters mentioned above, 2020a), GEP (Armaghani et al., 2018), RVM/RVR (Fattahi, 2019;
and hence a detailed assessment is required to compute the ROP of Zhang et al., 2020a), and DNN (Koopialipoor et al., 2019). Xu et al.
TBMs. (2019) also used KNN, CHAID, and CART to predict the ROP of
Predicting the performance of TBMs in a rocky medium is a TBM and attained a significant accuracy level (based on the deter-
complex task compared to that in a soil medium. Thus, researchers mination coefficient (R2) value) in many cases. These CSC models
investigated the inter-relationship between the rock properties and have also been successfully employed to solve problems in different
ROP of TBMs (Graham, 1976; Farmer and Glossop, 1980; Cassinelli engineering fields (Zhang and Goh, 2013, 2016; Chen et al., 2020;
et al., 1982). Based on the findings, it is clear that there is neither Goh et al., 2020; Ghani and Kumari, 2021; Ghani et al., 2021a;
an exclusive way nor a fixed technique for predicting the ROP of Kardani et al., 2021a; Kumar et al., 2021c). Despite the good per-
TBM in practice (Zhang et al., 2020a). Methods implemented by the formance of ANN, SVM, ANFIS, and many of the other conventional
researchers can be broadly categorised into three clusters: (i) methods, they are considered black-box models (Alavi and
theoretical and empirical, (ii) statistical, and (iii) computational. Gandomi, 2011; Mohammadzadeh et al., 2014; Bui et al., 2018).
These methods are developed from laboratory experiments and Also, the overfitting and local minima trapping issues limit their
field observation, a set of mathematical rules, and several soft applicability for further analysis, which are considered as the main
computing techniques, respectively. As mentioned by Gao et al. disadvantages of CSC algorithms (Armaghani et al., 2014; Roy and
(2020), TBM’s ROP estimation using theoretical and empirical Singh, 2020; Bardhan et al., 2021a).
models give relatively low accuracy, and hence new methods need Therefore, to circumvent the limitations of the CSC models,
to be developed. On the other hand, a comprehensive review of recent studies have resorted to hybrid soft computing model as
previous studies reveals that the performance of statistical models potential alternatives to predict the desired output (Bui et al., 2018;
(i.e. simple and multiple linear regression models) is relatively low Bardhan et al., 2021b). An amalgamation of meta-heuristic opti-
compared to the theoretical and empirical models (Koopialipoor misation algorithms and CSC models yields high-performance
et al., 2019; Shahrour and Zhang, 2021; Zhang et al., 2021a). Thus, computational models, which balance the exploration and exploi-
recent studies have resorted to soft computing techniques as po- tation processes during the course of optimisation, and offers a
tential alternatives to improve the reliability and exactness of the bendy and powerful method for solving high-dimensional and
existing empirical and statistical models (Grima et al., 2000; complex problems (Mohammadzadeh et al., 2014; Bui et al., 2018).
Benardos and Kaliampakos, 2004; Yagiz et al., 2009; Javad and For predicting the performance of TBM, Armaghani et al. (2017)
Narges, 2010; Yagiz and Karahan, 2011; Martins and Miranda, used two ANN-based hybrid models, i.e. particle swarm optimisa-
2013; Mobarra et al., 2013; Ghasemi et al., 2014; Mahdevari et al., tion (PSO)-ANN (ANN optimised with PSO) and imperialism
2014; Salimi et al., 2016; Armaghani et al., 2017, 2018; Fattahi, competitive algorithm (ICA)-ANN (ANN optimised with ICA), Zhang
2019; Koopialipoor et al., 2019; Xu et al., 2019; Gao et al., 2020; et al. (2020a) proposed a hybrid method of RVM and PSO; Zeng
Zhang et al., 2020a, b; Li et al., 2021; Zhou et al., 2021). et al. (2021) used six hybrid PSO-ELM techniques; Zhou et al.
Soft computing algorithms, with their ability in nonlinear (2021) used three hybrid SVM models, i.e. grey wolf optimisation
modelling, provide feasible tools for understanding and simulating (GWO)-SVM, whale optimisation algorithm (WOA)-SVM, and moth
diverse complicated problems (Chen et al., 2020; Goh et al., 2020; flame optimisation (MFO)-SVM, and found notable results.
Zhang and Goh, 2013, 2016), including predictions of the perfor- Armaghani et al. (2017) also compared the performance of PSO-
mance of TBMs (Goh et al., 2017, 2018; Xiang et al., 2018). In the ANN and ICA-ANN models with that of the conventional ANN,
recent past, several soft computing algorithms, such as neural and reported that the hybrid intelligent models are superior to the
network (NN), artificial NN (ANN), adaptive neuro-fuzzy inference simple ANN method. On the other hand, Zeng et al. (2021)
system (ANFIS), support vector regression (SVR), support vector concluded that the hybridisation of ELM and PSO results in
machine (SVM), gene expression programming (GEP), relevance higher accuracy than the single ELM model, and Zhou et al. (2021)
vector regression (RVR), relevance vector machine (RVM), deep NN showed that the hybrid SVM model (i.e. MFO-SVM) is more
(DNN), long-short-term memory (LSTM), classification and powerful than standard SVM in addressing TBM’s performance
regression trees (CART), k-nearest neighbour (KNN), and chi- with a high accuracy level. A thorough examination of these studies
squared automatic interaction detection (CHAID) (Grima et al., reveals that most of the authors consider single CSC models and
2000; Benardos and Kaliampakos, 2004; Mobarra et al., 2013; hybrid models in predicting the performance of TBM. However, the
Ghasemi et al., 2014; Mahdevari et al., 2014; Salimi et al., 2016; performance enhancement of the CSC models through ensemble
Armaghani et al., 2018; Koopialipoor et al., 2019; Xu et al., 2019; modelling was not addressed in earlier studies.
Gao et al., 2020; Zhang et al., 2020a; Li et al., 2021), have been used Recently, some researchers have focused on ensemble machine
successfully for predicting the ROP of TBMs. Successful application learning models for predicting the desired outputs in different
of these methods can also be found in the literature. Besides, many engineering disciplines (Basaran et al., 2019; Cai et al., 2020; Wang
other conventional soft computing (CSC) models, such as minimax et al., 2020; Asteris et al., 2021; Kardani et al., 2021b, c; Zhang et al.,
probability machine regression (MPMR), functional network (FN), 2021a, b). It is pertinent to say that ensemble models work on an
Gaussian process regression, and extreme learning machine (ELM), advanced level to augment the performance of “unstable” models
have been used successfully in different engineering disciplines for (Breiman, 1999; Qiu et al., 2014; Chatterjee et al., 2015). It must be
predicting the desired output(s) (Rajasekaran, 2004; Liu et al., 2015; noted that an ensemble model constructed with two or more in-
Jagan et al., 2019; Kumar and Samui, 2019; Samui, 2019; Ahmed dividual soft computing models exhibits better performance
et al., 2020; Mahmoud et al., 2020; Raja and Shukla, 2020, 2021a; compared to a single model (Qiu et al., 2014; Chatterjee et al., 2015;
Kumar et al., 2021a, b). Cai et al., 2020). Hence, taking these points as a reference, this study
For estimating the performance of TBM, several CSC models aims to implement a recently proposed approach, named hybrid
have been used, including ANN/NN (Benardos and Kaliampakos, ensemble technique (Asteris et al., 2021), for estimating the ROP of
2004; Yagiz et al., 2009; Javad and Narges, 2010; Mobarra et al., TBM. To this end, four common CSC techniques, i.e. MPMR, RVM,
1400 A. Bardhan et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1398e1412

ELM, and FN, are used, analysed, and discussed. Among them, the Based on kernel formulation, the MPMR model is depicted
former two models are categorised as regression-based models, below, which shows the relationship between the independent
while ELM and FN are NN-based models. However, the working variable (x) and response variable (y):
principles of ELM and FN are different. As stated above, these CSC
algorithms have also been successfully utilised in many engineer- X
n
y ¼ f ðxÞ ¼ bi Kðxi ; xÞ þ b (2)
ing disciplines. Nonetheless, the application of MPMR, ELM, and FN
i¼1
in estimating TBM’s performance is relatively scarce.
In the subsequent step, a hybrid ensemble model (HENSM) was where Kðxi ; xÞ represents a kernel function, b and bi are the out-
used for aggregating the outputs of the aforementioned CSC models puts, and n is the total number of observations of the input variable.
via an ANN model. In particular, the ANN was utilised to combine In general, the radial basis function (RBF), i.e. Kðxi ;xÞ ¼ exp½  ðxi ;
the CSC model outputs, while the HENSM was formulated to pre- xÞðxi ; xÞT =ð2s2 Þ, is used, where s is the width of the RBF. To use the
dict the ROP of the TBM. The prediction performance of the HENSM above MPMR formulation for regression, original data with d inputs
and CSC models was investigated with respect to a number of are used to create two classes of points (gi ; hi Þ as follows:
performance indices followed by uncertainty analysis and statisti-
cal testing. The empirical results confirmed the applicability and gi ¼ ðyi þ ε; xi1 ; xi2 ; .:: ; xid Þ; hi ¼ ðyi  ε; xi1 ; xi2 ; .:: ; xid Þ
benefits of the introduced HENSM in TBM performance estimation (3)
in a rock environment.
The remainder of this study is organised into the following The minimum probability machine classification boundary be-
sections, including the introduction section presented above. Sec- tween points gi and hi is shown below:
tion 2 presents the theoretical details of the utilised models,
including the approach of HENSM, while Section 3 explains the X
2n
gi K c ðzi ; zÞ þ bc ¼ 0 (4)
details of the collected dataset, descriptive statistics, statistical i¼1
analysis, and performance parameters. This is followed by a
detailed discussion on the performance of the employed models in The above classification boundary between gi and hi corre-
Section 4. Finally, in Section 5, the concluding remarks are sponds to a regression surface, termed as MPMR model. Eq. (4) can
provided. formulate using any binary classification algorithm, and it is not
limited to MPMCA. Precisely, if a binary classifier derived from Eq.
(1) has the above form where c ¼ 1 for the first class and c ¼ 1 for
2. Theoretical background of the employed models
the second class, then a crossing point needs to be found first. For
the crossing point, the classifier separates these classes for a certain
This section discusses the theoretical background of the
input (X ¼ ðx1 ; x2 ; .:: ; xd ÞÞ which is equivalent to finding the
employed models in predicting the ROP of TBM. It begins with
output of the regression model for a given input combination. For
presenting the details of CSC models, i.e. MPMR, RVM, ELM, and FN
detailed methodology, the work of Strohmann and Grudic (2003)
employed. Finally, the details of the ANN and the methodological
can be referred to.
development of the proposed HENSM are presented.
2.2. Relevance vector machine (RVM)
2.1. Minimax probability machine regression (MPMR)
RVM, introduced by Tipping (1999), is a supervised machine
MPMR is a regression framework (Strohmann and Grudic, learning technique that uses Bayesian interpretation to find
2003), which is established based on the minimax probability regression solutions and probabilistic classification. It has the same
machine classification algorithm (MPMCA). A regression model can functional formulation as SVM, but provides probabilistic inter-
be formulated as maximising the minimum probability of predicted pretation. In supervised machine learning, inputs (x1, x2, ., xn)
values where the prediction of the output lies within some bound come allied with their corresponding outputs ðt1 ; t2 ; .::; tn Þ. The
of the true function. The regression formulation closely shadows outputs might be real values for regression and class labels for
the classification formulation (Marshall and Olkin, 1960; Huang classifications. For a given training dataset, machine learns a model
et al., 2006), which can be given as based on input and output values. SVM is also a supervised machine
X n o learning technique that makes predictions based on a function
1
supEðzÞ ¼ zCovðzÞ ¼ Pr aT z  b ¼ u2 given by
z 1 þ u2
¼ infaT z  bðz  zÞT S1 T X
N
z ðz  zÞ (1) y ¼ f ðxÞ ¼ wn Kðx ; xn Þ þ w0 (5)
n¼1
where a and b are two constants, z is a random vector, and
supremum (sup) is taken over the distribution having mean z and where wn is the model weight. RVM uses training data for building
covariance matrix Sz . The above hypothesis assumes linear regression model in the form of
boundaries, but its nonlinear version can be obtained by using
Mercer kernels (Strohmann and Grudic, 2003). The expression ti ¼ yðxi Þ þ εi (6)
given in Eq. (1) is used to construct a nonlinear regression model,
which maximises the minimum probability that the future pre- where εi is the noise element of the predicted vector with the mean
dictions will lies within ε to the true regression function of 0 and variance of s2 , which is denoted as Nð0; s2 Þ. In the case of a
(Strohmann and Grudic, 2003). Implementation of MRMR by con- linear model, the prediction function is a linear combination of
structing a nonlinear classifier which divides two sets of points, basis functions ai ðxÞ and depicted below:
obtained by transferring the regression data (between þε and eε)
along the response variable axis. The classification boundary be- X
M
tween the two sets is categorised as a regression surface, and it is yðxÞ ¼ wi ai ðxÞ (7)
termed as MPMR model. i¼1
 A. Bardhan et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1398e1412 1401

The above equation can be rewritten in the form of y ¼Fw þ ε,

where F is a matrix of N  M dimensions. Its ith column is formed
with the values of the basis function ai ðxÞ at all training points. The
noise vector is ε ¼½ε1 ; ε2 ; .::; εN . For a given dataset of indepen-
dent and dependent pairs fxn ; yn gN i¼1 , standard formulation is
normally followed and assumes  that pðyjxÞ is the multivariate
Gaussian distribution NðyyðxÞ; s2 Þ. As defined in Eq. (6), the mean
of this distribution is modelled by y(x) for a given x. The likelihood
of the dataset can be given by

    N=2  
 1  0 2
p t w; s2 ¼ 2ps2 exp t  F w (8)
2ps2

where t ¼ ðt1 ; t2 ; .::; tN Þ, w ¼ ðw0 ; w1 ; .::; wN Þ; and F0 is a

design matrix of N  (N þ 1) dimensions with F0nm ¼ Kðxn ; xm1 Þ
and Fn1 0 ¼ 1. Maximum likelihood estimation of w and s2 often
results in overfitting. As suggested by Tipping (1999), the imposi-
tion of some prior constraints on w by adding a ‘complexity’ pen-
alty term to the likelihood or error function can overcome the
overfitting. This adjusts the simplification ability of the learning
process. Generally, new parameters (high level) are used to compel
an unambiguous zero-mean Gaussian prior probability distribution Fig. 1. A typical network of an extreme learning machine.
over the weights. Using Bayes’ rule, the posterior distribution of
weights is given by
parameters, the output of an ELM network with p inputs, q hidden
   pt w; s2  pwa
 neurons, and o output neurons can be written in the form of
p wt; a; s ¼
2    (9)
p t a ; s 2
oi ðtÞ ¼ mTi hðtÞ (12)
where a is a vector of N þ 1 hyperparameters.
The resulting posterior distribution over the weights is the where mi ˛ Rq ð i˛f1; 2; .; og Þ is the weight vector, and it con-
multivariable Gaussian distribution: nects the hidden neurons to the ith output neuron; and hðtÞ˛ Rq is
the output vector of hidden neurons, which is given by
    X T   1 h      i

p wt; a; s2 ¼ N s2 F t ; s2 FT F þ A (10) hðtÞ ¼ f wT1 xðtÞ þ b1 ; f wT2 xðtÞ þ b2 ; .; f wTq xðtÞ þ bq
(13)
 A ¼ diagða0 ; a1 ; .::; aN Þ. For uniform hyperpriors, the term
where
pðt a; s2 Þ is required to maximise: where bk ðk ¼ 1; 2; /; q) is the bias of kth hidden neuron, wk ˛Rp
   Z    is the weight vector of kth hidden neuron, and f ð ,Þ denotes the
 
p t a; s2 ¼ p t w; s2 pðwjaÞdw activation function (sigmoidal function). It is worth noting that the
bias (bk ) and weight (wk Þ vectors are produced in a random manner
ð2pÞN=2 1   from which the hidden layer output matrix (H) can be formed. The
¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi exp  yT s2 þ FA1 FT 1 y
2 weight matrix is then calculated using the ‘MooreePenrose pseudo
js2 þ FA1 FT j
inverse’ method, which is given by
(11)

 1
Maximisation of pðt a; s2 Þ is called as the Type II maximum M ¼ H  HT H  DT (14)
likelihood method. Hyperparameter estimation is carried out using
gradient descent on the objective function. The non-zero weights of where D ¼ ½dð1Þ; dð2Þ; .::; dðnÞ represents a matrix of o  n di-
w will be considered as relevant vectors. mensions, and its tth column is the real largest vector dðtÞ˛Ro. The
class label for the new input parameters can be calculated as

2.3. Extreme learning machine (ELM) Y ¼ arg maxfOi g ði ¼ 1; 2; .; OÞ (15)

ELM, introduced by Huang et al. (2006), is a single hidden layer where Y is the predicted class label. Detailed working principle of
feed-forward NN used for solving regression, clustering, and ELM can be referred to the literature (Zeng et al., 2021).
feature learning problems. During the course of the training and
predicting stages, weights and bias are assigned randomly between 2.4. Functional network (FN)
inputs and hidden nodes which remain constant throughout the
operation. On the other hand, weights between the hidden and FN, introduced by Castillo et al. (2000), is an extension of NNs.
output layers can be learned quickly. In addition, the computational The main benefit of NNs is to learn from records with the help of
cost of ELM is on the lower side compared to another back- various learning methods such as parametric and structural
propagation algorithm network. A typical structure of an ELM is learning methods. In NN, the error function is minimised for the
shown in Fig. 1. learning process. The main disadvantage is that the solution ob-
For the construction of a classification model from a set of data tained from NN cannot be guaranteed as the best one due to the
(X ¼ xt ˛Rp ðt ¼ 1; 2; . ; nÞ) having n samples and p input limited ability to develop the theory and uncertainties involved in
1402 A. Bardhan et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1398e1412

the problem. While solving a complicated problem, such limita-

tions can be overcome by using FN. FN is a more comprehensive
version of NNs, and it combines domain knowledge and data. FN
uses structural learning and parametric learning to develop a
model. In structural learning, functional equations are used for
simplification, whereas in parametric learning, neuron functions
are estimated based on the combination of shape functions.
Note that FN is a combination of storing units, layers of the Fig. 3. Flowchart showing the working steps of functional network.
computing unit, and a set of links. Storing unit consists of input
storing units, intermediate layer units, and output units. Neurons of
the computing unit assess the set of input values approaching from
the following layer. Functions that are not arbitrary but can be
determined by the network structure are termed as links. Besides
data, other informational and functional properties such as
invariance, commutativity, and associativity are also considered
during the construction of networks. In FN, weights of the functions
are merged and learned. In some cases, the learning method leads
to a global minimum quite easily. A graphical representation of an
FN is shown in Fig. 2. In addition, the steps of FN working are Fig. 4. Associativity of functional network.
depicted in Fig. 3.
The learning method of FN comprises neural functions obtained
from a set of dependent and independent variables, and the
X
4
learning process is based on minimising the Euclidean norm of the f4 ðx4 Þ ¼ a4i q4i (19)
error function (Rajasekaran, 2004): i¼1

From the input function, the below function can be constructed

1X n
as
E¼ ðO  FðiÞÞ2 (16)
2 i¼1 i
b
f 4 ðx4 Þ ¼ f1 ðx1 Þ þ f2 ðx2 Þ þ f3 ðx3 Þ (20)
and the approximate neural function can be organised as
and the error in the jth datum is given by
       
X
m ej ¼ f1 x1j þ f2 x2j þ f3 x3j  f4 x4j (21)
fi ðxÞ ¼ aij qij ðXÞ (17)
j¼1 The above equation can also be rewritten or reproduced in the
form of
where q is the shape function, and it can be associated with
different functions such as algebraic expressions, trigonometric ~g
ej ¼ bj fa (22)
functions, and exponential functions. These associative optimisa-
The sum of squares of the error for all data can be given by
tion functions lead the formulation to a system of linear or
nonlinear equations. For a system with three inputs ðx1 ; x2 ; x3 Þ and X
n
one output ðx4 Þ, the association of the FN can be constructed, as E ¼ ak ¼ aki qki ðxk0 Þ (23)
shown in Fig. 4, the mathematical formulation of which is given by i¼1

where xk0 is the initial value. For the uniqueness of the above so-
X
ms
lution, the following equation must be satisfied:
fs ðxs Þ ¼ asi qsi ðs ¼ 1; 2; .Þ (18)
i¼1
0 1
X
mk X
n
fk ðxk0 Þ ¼ ~ i@
eTj ej ¼ ha bj bj A fa
~ g ¼ ha
~i½Afa
~g (24)
where ms can be of any order; and qsi can be polynomial of any
i¼1 j¼1
degree, trigonometric, trigonometric polynomial, exponential, or
any other mathematical functions, termed as shape functions. For a To achieve more generalised functions, it is required to guar-
polynomial expression, the function can be expressed as antee the individuality of the illustration of the network. Some-
times different neuron functions lead to a same output for a similar
input which leads to an ill-conditioned estimation. To solve these
problems, a generalised method is used for all types of FNs (Castillo
and Gutiérrez, 1998).

2.5. Artificial neural network (ANN)

ANNs, inspired by the biological neural system, consist of many

single units and weighted artificial neurons (Asteris and Mokos,
2020). An ANN is also recognised as a processing element (PE)
since it processes information. Each PE contains one or more
weighted inputs, an output, and a transfer function. Basically, a PE is
Fig. 2. Architecture of functional network. an equation that balances the inputs and output. An ANN is also
 A. Bardhan et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1398e1412 1403

known as a connectionist model since the connection weights (1) The standalone machine learning models are trained via
stand for the system’s memory. input data X and output data Y.
An ANN involves three main sets of layers with neurons. The (2) The predicted values are generated.
first and last layers of an ANN are known as the input and output (3) The predicted values of the individual machine learning
layers, respectively. They contain many neurons as the input and models are amalgamated.
output variables. Also, hidden layers exist between these two (4) New input data Xnew are formed, and an ANN is trained using
layers. Signals are transmitted through the input layer. The hidden the real predicted values Y.
layers serve as the computation engine, and predictions are per- (5) An HENSM is constructed.
formed within the output layer. The weights and biases are crucial
parameters of an ANN. The weights represent the interconnections Fig. 5 shows the overall diagram of the proposed HENSM
between a layer’s neurons, while the biases determine the degree method used in this study for predicting the ROP of TBM.
of freedom. Each of the nodes, except for the input nodes, utilises a
nonlinear activation function to produce the output. The output is 3. Data description and computational analysis
then utilised as the input for the next node. This process continues
until a proper solution to a given problem is found. To calculate the 3.1. Descriptive details of the collected dataset
error by comparing the actual result (i.e. the objective of the
problem) to the prediction (i.e. the outcome of the network), back- A sum of 185 datasets was collected from the literature (Javad
propagation is utilised. Then, the error propagates to a layer back at and Narges, 2010) and was used for predicting the ROP of TBM.
a time throughout the ANN structure, changing the weights on the The collected dataset comprised 150 data of Queens water tunnel
basis of their error contributions. project of USA, 8 data of KarajeTehran tunnel project of Iran, and 27
data of Gilgel Gibe II hydroelectric project of Ethiopia. The main
dataset consists of three input parameters, i.e. uniaxial compressive
2.6. Implementation procedure of hybrid ensemble model (HENSM) strength (UCS), rock quality designation (RQD), and distance be-
tween planes of weakness (DPW). These three parameters were
Ensemble learning is defined as a machine learning process that used to predict the ROP (in m/h), i.e. the output variable. Table 1
is used to achieve superior results through the strategical combi- represents descriptive statistics of the collected dataset with the
nation of several learning algorithms (Qiu et al., 2014; Chatterjee minimum, the maximum, the mean, and so on, of each variable. In
et al., 2015), i.e. the principle of ensemble modelling. An addition, Figs. 6 and 7 plot the correlation matrix (based on Pearson
ensemble model merges several standalone/individual models to correlation coefficient) and frequency histograms, respectively. The
deliver better prediction accuracy (Dietterich, 2000). Specifically, frequency histograms are presented in the form of a comparative
ensemble models have three main advantages: (i) the ability to histogram between the input and predicted variables. This diagram
decrease the risk of choosing the worst-performance model by the is useful to compare the input variables against the output. How-
aggregation of multiple candidate models; (ii) efficiency in dealing ever, for a fair comparison, all the parameters were scaled in a pre-
with locally optimal solutions; and (iii) a representational advan- defined range between 0 and 1. From Table 1, it is noted that the
tage (in numerous cases, one model cannot represent a true func- UCS of rocks lies in the range of 30e199.7 MPa, RQD has the min-
tion by a single hypothesis or formulation) (Qiu et al., 2014). It is imum and maximum values of 40.6% and 99.88%, respectively,
worth noting that, for regression and time series forecasting while the DPW locates between 0.05 m and 2 m. The output vari-
problems, different machine learning models that use identical able, ROP, is distributed in the range of 1.27e14.43 m/h.
data produce different prediction results. Thus, all the outputs
created by several machine learning models may be combined, 3.2. Statistical analysis
providing a more robust prediction compared with those by
separate machine learning models. An analysis of the relation be- From the descriptive analysis presented above, it was under-
tween the actual and predicted outputs makes it promising to stood that the present database comprises a wide range of data, and
create a more effective model via different ensemble methods, such hence statistical analysis was performed to assess the linear asso-
as randomness injection, voting, boosting, stacking, and bagging ciation between the input parameters and ROP, the details of which
(Bar et al., 2013; Qiu et al., 2014). are furnished in Fig. 8aec. The ROP values were correlated with the
Nonetheless, the present study uses an approach proposed by UCS, RQD, and DPW and presented in terms of determination co-
Asteria et al. (2021) instead of standalone ensemble techniques efficient (R2) value. As can be seen, the values of R2 scatter in the
mentioned above. The presented HENSM is made up of four indi-
vidual machine learning models (i.e. MPMR, RVM, ELM, and FN)
and an ANN. In particular, the ANN was employed for aggregating Table 1
Descriptive statistics of UCS, RQD, DPW, and ROP.
the outputs of the four individual machine learning models to
formulate the HENSM. The process details are as follows: Parameters UCS (MPa) RQD (%) DPW (m) ROP (m/h)

Minimum 30 40.6 0.05 1.27

1st Quartile 129.2 90.98 0.3 1.85
Mean 143 91.05 0.87 2.33
3rd Quartile 160.3 99.81 1.6 1.43
Maximum 199.7 99.88 2 14.43
Trimmed 143.53 94.51 0.83 2.13
Standard error 2.05 1.13 0.05 0.1
Standard deviation 27.93 15.38 0.67 1.32
Variance 780.32 15.38 0.67 1.32
Skewness 0.4 1.76 0.41 6.18
Kurtosis 4.32 4.84 1.67 49.06
Geometric mean 139.72 89.37 0.55 2.18
Harmonic mean 135.23 87.19 0.29 2.1
Fig. 5. Schematic diagram of the proposed hybrid ensemble model.
1404 A. Bardhan et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1398e1412

dimensional effects. Therefore, before the development of the

models, the main dataset was normalised in the range of 0e1.
Subsequently, the normalised dataset was partitioned into training
and testing subsets. For this purpose, 70% of the main dataset (i.e.
130 samples) was selected at random for the training subset, and
the left-over dataset, i.e. 30% of the main dataset (i.e. 55 samples),
was used as the testing subset. The entire steps of developing the
predictive models of ROP of TBM can be described as: (i) normal-
isation of the main dataset, (ii) selection of training and testing
subsets, (iii) construction of models using training subset, and (iv)
validation of the developed models using the testing subset.
Consequently, to evaluate the models’ performance, several
widely used statistical indices, i.e. determination coefficient (R2),
performance index (PI), variance account for (VAF), Willmott’s in-
dex of agreement (WI), mean absolute error (MAE), mean bias error
(MBE), root mean square error (RMSE), and weighted mean abso-
lute percentage error (WMAPE), were determined (Kardani et al.,
2020, 2021a, d; Bardhan et al., 2021a, b; Ghani et al., 2021a,b;
Kumar et al., 2021c; Ray et al., 2021). The mathematical expressions
Fig. 6. Correlation matrix.
of the above-mentioned indices are given in the following
equations:

range of 0.0944e0.2133, which in turn indicates that the parame- Pn P

i¼1 ðyi  ymean Þ2  ni¼1 ðyi  b
y i Þ2
ters are marginally correlated with ROP. To visualise the nature of R2 ¼ Pn 2
(25)
the input parameters, the collected dataset is presented in Fig. 8aec i¼1 ðyi  ymean Þ
through scatter plots of each input parameter.
PI ¼ adj:R2 þ 0:01VAF  RMSE (26)
3.3. Artificial intelligence-based analysis
varðyi  byiÞ
In the soft computing field, data normalisation is a pre- VAF ¼ 1   100% (27)
varðyi Þ
processing task, which is performed to cancel out the multi-

9 9 9
(a) UCS (b) RQD (c) DPW
8 ROP 8 ROP 8 ROP
7 7 7
6 6 6
Density

Density

5
Density

5 5
4 4 4
3 3 3
2 2 2
1 1 1
0 0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
UCS and ROP RQD and ROP DPW and ROP

Fig. 7. Frequency histograms.

Fig. 8. Liner relationship between the inputs and ROP.

 A. Bardhan et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1398e1412 1405

Pn
i¼1 ðyi b
y i Þ2
WI ¼ 1  Pn (28)
b  ymean j þ jyi  ymean jÞ2
i¼1 ðj y i

1X n
MAE ¼ jb
y  yi j (29)
n i¼1 i

1X n
MBE ¼ ðb
y  yi Þ (30)
n i¼1 i

vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u n
u1 X
RMSE ¼ t ðy  b y i Þ2 (31)
n i¼1 i Fig. 9. Illustration of RVM weights corresponding to the training dataset.

 
Pn yi by i  rigorously checked within a predefined range of 0.5e5 for s and
i¼1  yi yi
0.001e1 for ε, respectively. Following the trial-and-error approach,
WMAPE ¼ Pn (32)
i¼1 yi the best performance was obtained when the values of s and ε were
set to 1.1 and 0.003, respectively, which are the designed values of s
where yi and b y i represent the actual and estimated ith values, and ε for the developed MPMR model.
respectively; n is the number of samples for the dataset under For modelling the RVM, the widely used Gaussian kernel was
consideration; ymean is the mean of the actual values; and adj:R2 is used (Samui et al., 2011) for the same training dataset. The value of
the adjusted R2. Note that for an ideal model, the values of these s was selected through trial-and-error runs and the best perfor-
indices should be equal to their ideal values, as given in Table 2. mance was obtained at s ¼ 0.4. Therefore, the denominator term
Furthermore, score analysis was conducted to evaluate the (2s2) of the RBF, i.e. exp½  ðxi ;xÞðxi ; xÞT =ð2s2 Þ, can be calculated as
developed models’ overall performance (Zorlu et al., 2008). It may 2  0.42, i.e. 0.32. The details of the weight vector against the
be noted that score analysis is an easy and effective method for training dataset are shown in Fig. 9. The following equation rep-
comparing the models’ performances. In this technique, the model resents the developed RVM model for calculating the ROP of TBM:
with the least value for each performance parameter is assigned a " #
score of 1, and the model with the best value for the same perfor- X
130
ðx  xi Þðx  xi ÞT
mance parameter is assigned a score of m (the total number of ROP ¼ wi exp  (33)
i¼1
0:32
considered models) for training and testing results separately. In
the present work, m ¼ 5, i.e. the considered models are MPMR,
where xi is the training dataset of 130 observations, and x is the
RVM, ELM, FN, and HENSM. The subsequent step involves calcu-
input for which ROP is to be determined.
lating the total score by adding up their separate scores. The ulti-
In contrast, to determine the ELM structure, the number of
mate model score was computed by adding the total scores
hidden neurons varying from 2 to 30 was examined. In this study,
belonging to the training and testing phases.
sigmoid activation function was used to select the best possible
ELM model. Through trial-and-error runs, the most appropriate
value of hidden neurons was obtained as 12.
4. Results and discussion
The fourth CSC model, i.e. FN, was also constructed using the
same training dataset as those used in MPMR, RVM, and ELM
4.1. Realisations of conventional soft computing models
modelling. A trade-off was made in the present study, and FN with
degree 3 and polynomial basis function was adopted. The equation
As stated earlier, a sum of 185 datasets was collected and par-
for the prediction of the ROP of TBM was obtained as follows:
titioned into training and testing subsets, of which the training
subset was employed to construct the models, while the testing
X
n X
m0  
subset was utilised to validate the developed models. Subsequently, y ¼ aþ aij fi xj (34)
different indices were determined to measure the prediction ca- i¼1 j¼1
pabilities of the developed models, the details of which are pre-
sented and discussed below. Note that the same training dataset where m0 is the degree of variable.
was used in all cases. However, before discussing the model’s per-
formance, the parametric configurations of each model are pre-
4.2. Realisation of hybrid ensemble model (HENSM)
sented below.
In MPMR modelling, two hyper-parameters, i.e. s (width of the
This sub-section discusses the realisation of the HENSM. Note
RBF) and ε (noise component of the measurement) play an
that a model that achieves better accuracy in the testing phase is
important role, and hence the values of these parameters were
usually considered to be a robust model and accepted with higher
conviction. Therefore, one may create an even more robust model
Table 2
by improving the testing phase performance. With this consider-
Ideal values of performance indices.
ation, the superiority of the ensemble method was leveraged in the
Index Ideal value Index Ideal value work to construct an HENSM. This will be more advantageous with
R2 1 MAE 0 respect to the variational inaccuracies of the utilised models.
PI 2 MBE 0 It is worth mentioning that in many cases, computational
VAF (%) 100 RMSE 0 models produce different results with different accuracies when
WI 1 WMAPE 0
there are insufficient samples. Hence, ensemble techniques can
1406 A. Bardhan et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1398e1412

Fig. 10. Utilisation of standalone models for the development of hybrid ensemble model. TRe Training; TS e Testing.

decrease the risk of choosing the less accurate model via the ag- section, the details of which are listed in Tables 3 and 4 for the
gregation of multiple individual models. The present study at- training and testing datasets, respectively. Furthermore, the scatter
tempts hybrid ensemble modelling via an HENSM combining four plots between the actual and predicted outputs are presented in
CSC models that are used to estimate the ROP of TBM. As mentioned Figs. 11 and 12. As can be seen, overall, all CSC models were able to
in Section 2.6, the ANN was used for constructing the HENSM. This predict the ROP with significant accuracy that can be evident by
model combines the outputs from all four individual models as high R2 value (ranging between 0.7214 for FN and 0.911 for ELM)
inputs and the real output, i.e. actual ROP of TBM. In Fig. 10, the and low RMSE (ranging between 0.0286 for ELM and 0.0505 for FN)
process of implementing the HENSM is presented, from which the in the training phase. For the testing dataset, the values of R2 were
amalgamation procedure of CSC models can be visualised. It must obtained in the range of 0.7767 (for MPMR) to 0.8518 (for RVM),
be noted that a trial-and-error approach was followed to finalise while the values of RMSE were obtained in the range of 0.0453 (for
the structure of ANN in HENSM modelling. In this case, the RVM) to 0.0532 (for FN). Based on the score analysis, the best
Levenberg-Marquardt back-propagation function was selected as standalone model is ELM (total score of 30) in the training phase
the training function. The ultimate ANN structure is composed of and RVM (total score of 30) in the testing phase. Contrarily, the
three input neurons, two hidden neurons, and one output neuron. constructed HENSM attained the most accurate prediction with the
The activation function used in input to hidden and output layers highest R2 (0.9126 for training and 0.865 for testing dataset) and
are tan-sigmoid and purelin, respectively. Also, a total of 10% of the lowest RMSE value (0.0283 for training and 0.0418 for testing
combined training dataset was utilised as the validation dataset to dataset) in both cases. The outcomes of score analysis also show
stop overfitting. that the constructed HENSM outperformed the best performing
CSC model and attained a final score of 36 and 38 in the training
and testing phases, respectively. The results of score analysis are
4.3. Performance indices and evaluation
presented in Fig. 13 (for training results) and Fig. 14 (for testing
results) in the form of stacked bar plots. The comparisons of score
A comprehensive assessment of the CSC and HENSM models
analyses presented in Table 5 indicate that RVM and ELM are the
based on multiple performance indices is presented in this sub-

Table 3
Outcomes of the training dataset.

Index Particular R2 PI VAF WI MAE MBE RMSE WMAPE Total score

MPMR Value 0.8956 1.7578 89.5593% 0.972 0.0227 0 0.0309 0.2852

Score 2 2 2 2 2 5 2 2 19
RVM Value 0.9099 1.789 90.9938% 0.9761 0.0205 0.0007 0.0287 0.2575
Score 3 3 3 3 4 2 3 3 24
ELM Value 0.911 1.7913 91.0966% 0.9764 0.0205 0 0.0286 0.2574
Score 4 4 4 4 3 3 4 4 30
FN Value 0.7214 1.3856 72.1372% 0.9139 0.0332 0 0.0505 0.4198
Score 1 1 1 1 1 4 1 1 11
HENSM Value 0.9126 1.7947 91.2556% 0.9766 0.0204 0.001 0.0283 0.2554
Score 5 5 5 5 5 1 5 5 36
 A. Bardhan et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1398e1412 1407

Table 4
Outcomes of the testing dataset.

Index R2 PI VAF WI MAE MBE RMSE WMAPE Total score

MPMR Value 0.7767 1.4872 77.5715% 0.9358 0.0344 0.0027 0.0521 0.3873
Score 1 1 2 2 2 4 2 2 16
RVM Value 0.8518 1.6289 83.1133% 0.959 0.03 0.0043 0.0453 0.3346
Score 4 4 4 4 4 2 4 4 30
ELM Value 0.8267 1.5909 82.1114% 0.9525 0.0301 0.0046 0.0467 0.3367
Score 3 3 3 3 3 1 3 3 22
FN Value 0.8039 1.505 76.5798% 0.9164 0.0363 0.0021 0.0532 0.4115
Score 2 2 1 1 1 5 1 1 14
HENSM Value 0.865 1.6715 85.6303% 0.9636 0.0282 0.0043 0.0418 0.3142
Score 5 5 5 5 5 3 5 5 38

15.00 15.00 15.00

MPMR (TR) RVM (TR) ELM (TR)

10.00 10.00 10.00

Predicted

Predicted
Predicted

5.00 5.00 5.00

0.00 0.00 0.00

0.00 5.00 10.00 15.00 0.00 5.00 10.00 15.00 0.00 5.00 10.00 15.00
Actual Actual Actual
(a) (b) (c)
15.00 15.00
FN (TR) HENSM (TR)

10.00 10.00
Predicted

Predicted

5.00 5.00

0.00 0.00
0.00 5.00 10.00 15.00 0.00 5.00 10.00 15.00
Actual Actual
(d) (e)
Fig. 11. Scatter plots between actual and estimated ROP values for the training (TR) dataset.

top two standalone models (attained overall scores of 54 and 52, measure the model’s predictive performance for the training and
respectively); however, when the outputs of each model are com- testing datasets. Fig. 15 displays the accuracy matrix for the per-
bined with the ANN, they exhibit better performance than the best formance indices determined in this study. It indicates the accuracy
standalone model in predicting the ROP of TBM. of the performance parameters (in percentage) by comparing them
In addition, the objective function (OBJ) criterion was used to with their corresponding ideal values. For example, the ideal value
assess the predictive performance of all the employed models. The of MAE is 0 and, in the present work, the value of MAE for the
details of this parameter can be found in the literature (Gandomi training subset was determined as 0.0227 for the MPMR model (see
et al., 2013; Samadi et al., 2020; Raja et al., 2021; Raja and Shukla, Table 3). Thus, it can be estimated that the MPMR model attained
2021b). However, the expression presented by Samadi et al. 97.73% ((1e0.0227)  100%) accuracy in terms of MAE. On the other
(2020) was used. It should be noted that the best model has a hand, the values of R2 and PI were obtained as 0.865 and 1.6715 in
minimum OBJ value (ideal value ¼ 0). The values of OBJ for MPMR, the testing phase, respectively, for the HENSM (see Table 4), which
RVM, ELM, FN, and HENSM were calculated as 0.0366, 0.03, 0.0307, shows that HENSM attained 86.5% (0.865/1  100%) and 83.575%
0.0455, and 0.0284, respectively. These results indicate that HENSM (1.6715/2  100%) accuracy in terms of R2 and PI, respectively.
attained the most desired prediction with the lowest OBJ value, and Similar procedure was followed for the other parameters as well.
hence the proposed hybrid model is most accurate from this However, it may be noted that the parameters such as VAF, which
viewpoint. are determined in percentage terms, should be converted in their
To better demonstrate the values of performance indices, a decimal form before implementing the above procedure. The su-
recently proposed heat map-shaped graphical assessment, called premacy of the developed HENSM based on uncertainty analysis
accuracy matrix, is tested to visualise the model efficiency (Kardani (UA) and statistical testing is presented in the following sub-
et al., 2021a). This matrix displays multiple statistical parameters to section.
1408 A. Bardhan et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1398e1412

15.00 15.00 15.00

MPMR (TS) RVM (TS) ELM (TS)

10.00 10.00 10.00

Predicted

Predicted
Predicted

5.00 5.00 5.00

0.00 0.00 0.00

0.00 5.00 10.00 15.00 0.00 5.00 10.00 15.00 0.00 5.00 10.00 15.00
Actual Actual Actual
(a) (b) (c)
15.00 15.00
FN (TS) HENSM (TS)

10.00 10.00
Predicted

Predicted
5.00 5.00

0.00 0.00
0.00 5.00 10.00 15.00 0.00 5.00 10.00 15.00
Actual Actual
(d) (e)
Fig. 12. Scatter plots between actual and estimated ROP values for the testing (TS) dataset.

4.4. Uncertainty analysis and statistical testing efficiency of the models. The smaller the value of WCB, the better
the model certainty, meaning that the model will suffer from less
This sub-section presents the quantitative evaluation of the error, and it will predict more precise expected output. Based on
developed models through UA in the prediction of the ROP of TBM. the values of WCB, all the five models were ranked. It is seen that
For this purpose, the testing dataset, which contains 55 real-time the proposed HENSM attained the lowest WCB value and secured
observational data points, was used. Thus, it can be useful to logi- the first rank, rendering it the most reliable predictive model. The
cally compare predictive outputs for evaluating the reliability of the RVM model scored the second rank, which is the top performing
predictive models (Gholami et al., 2019). To carry out the UA, the model among the CSC models, followed by ELM, FN, and MPMR.
absolute error (εi) between observational (yi ) and predicted ( b yi ) However, the MPMR model exhibits the highest WCB value and the
values, mean of absolute error (MOE), and standard deviation (SD) lowest uncertainty. Fig. 16aec provides UA graphical representation
of error were determined using the following expressions: as a bar chart displaying MOE, WCB, and ME values for a better
comparison.
εi ¼ jyi  b
yij (35) Furthermore, a one-tailed t-test was executed to estimate the
significant difference of the proposed HENSM’s performance to the
PN CSC models in predicting the ROP of TBM in a rock environment.
i¼1 εi
MOE ¼ (36) This statistical test was executed on the MAE values with the
N
hypothesised mean difference HMD ¼ 0. At HMD ¼ 0 and confi-
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi dence interval of 95% (i.e. a ¼ 0.05), the hypotheses of one-tailed t-
PN 2
i¼1 ðεi  MOEÞ test are H0: MAEHENSM e MAECSC models ¼ 0 and HA: MAEHENSM e
SD ¼ (37) MAECSC-models < 0, where H0 and HA represent null hypothesis and
N1
alternate hypothesis, respectively. The outcomes are presented in
where N is the number of observations (here N ¼ 55). Initially, MOE Table 7. Note that the rejection (failed to accept) of H0 (i.e. t-stat < t-
and SD were determined for the testing datasets using the above critical) shows that the proposed HENSM considerably out-
equations. Subsequently, the width of confidence bound (WCB) performed all four CSC models in reducing the MAE value in both
value was determined by calculating the margin of error (ME) at a phases.
95% confidence interval. The standard error (SE), lower and upper
bounds (LB and UB) were calculated as follows:
4.5. Sensitivity analysis
SD
SE ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffi; LB ¼ MOE  ME; UB ¼ MOE þ ME; To evaluate the relative influence of the input variables on the
N1
output, sensitivity analysis was performed. Note that the most
WCB ¼ UB  LB (38)
widely used method, i.e. the cosine amplitude method, was used in
Note that in UA, WCB is an error range where about 95% of the this study (Asteris and Mokos, 2020; Kardani et al., 2021d). To
entire data are located. Table 6 provides UA details, including the perform the sensitivity analysis, the strength of relation (Ryn ;m ) of
MOE, SD, SE, ME, LB, UB, and WCB. These indices help to assess the input parameters (UCS, RQD, and DPW) in predicting the ROP of
 A. Bardhan et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1398e1412 1409

4.6. Discussion

From the experimental results presented in the above sub-

sections, it is observed that RVM and ELM are the top two con-
ventional models. However, detailed scrutiny reveals that RVM
attained the most accurate prediction in the testing phase only.
Although overall analysis shows that RVM outperformed other CSC
models with R2 ¼ 0.8518 and RMSE ¼ 0.0453, a lower performance
in the testing phase (R2 ¼ 0.9099 and RMSE ¼ 0.0287) against the
performance of ELM (R2 ¼ 0.911 and RMSE ¼ 0.0286) encourages
the authors to analyse the prediction capability in another way.
Therefore, a hybrid ensemble method was implemented to
augment the CSC models’ performance, especially for the testing
dataset. The performance of the proposed HENSM was then
Fig. 13. Bar plot of score analysis (training results). assessed and compared for predicting the ROP of TBM in a rock
environment. Overall, the proposed HENSM attained the maximum
final score of 74 (36 (training) þ 38 (testing)) by far, followed by
RVM (54), ELM (52), MPMR (35), and FN (25). These results indicate
that RVM and ELM are the top two models among the 4 conven-
tional models (i.e. MPMR, RVM, ELM, and FN) employed in this
study. However, the FN is the worst performing model with the
lowest final score of 25.
Analogous to CSC models, the proposed HENSM was developed
utilising the outputs of the CSC models in a MATLAB (2015a
version) environment. The computation costs were recorded as
0.311139 s for MPMR, 0.823362 s for RVM, 0.325061 s for ELM,
7.864904 s for FN, and 8.1252 s for HENSM. All the programmes run
in MATLAB with i3-8130U CPU @ 2.20 GHz, 12 GB RAM, which in-
dicates that the constructed HENSM attained higher prediction
accuracy with very low computing cost. This is one of the main
advantages of the proposed approach due to the fact that this
Fig. 14. Bar plot of score analysis (testing results).
approach can be implemented quickly. Also, the present technique
can reduce the risk of picking the least performing model by
combining individual models.
Table 5
Overall score analysis for the developed models. 5. Concluding remarks
Phase MPMR RVM ELM FN HENSM
Based on the data mining from the literature, 185 data records of
Training 19 24 30 11 36
Testing 16 30 22 14 38 ROP of TBM were collected, which cover three rock properties, i.e.
Final score 35 54 52 25 74 UCS, RQD, and DPW. These parameters were used as the input
parameters to predict the ROP of TBM in a rock environment. The
collected dataset was used to rigorously train and validate the
TBM was determined for the actual data and employed models employed models to enable accurate predictions of ROP of TBM.
using the following expression: Four standalone models, i.e. MPMR, RVM, ELM, and FN, and an
HENSM were employed in this study. The accuracies of these
models in predicting the ROP of TBM were analysed and compared
Pk from different aspects. Subsequently, score analysis was performed
i¼1 xm;i yn;i to select the best-performing model. The UA and statistical testing
Ryn ;m ¼ rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi (39)
Pk   Pk   were also performed to estimate the reliability of the proposed
2 2
i¼1 xm;i i¼1 yn;i HENSM in predicting the ROP of TBM. The experimental results
reveal that the constructed HENSM attained higher performance
where xm;i represents the ith value of mth independent variable; k compared to the standalone models and can be used effectively to
and m are the number of total observations and independent var- estimate the ROP of TBM in the design phase of engineering
iables, respectively; yn;i denotes the ith value of nth dependent projects.
variable; Ryn ;m denotes the strength of relation of mth independent To sum up, a hybrid ensemble approach is successfully applied
variable to the nth dependent variable; and n is the number of in this study to predict the ROP of TBM in a rock environment. It
dependent variables. In this study, m ¼ 3, n ¼ 1, and k ¼ 185. The appears to be an accurate yet computationally efficient method due
strength of relation between the ROP and the input parameters, i.e. to the following reasons: (i) easy implementation procedure; (ii)
UCS, RQD, and DPW, is shown in Fig. 17. For actual and employed higher prediction accuracy; (iii) very low computational cost; (iv)
models, the strength of relation of each variable is presented in the produced greater performance compared to CSC models, and (v)
figure itself. As can be seen, the UCS is the most influential being representational. Experimental results exhibit that the
parameter on the ROP, followed by RQD and DPW, in call cases. It is HENSM outperformed other models at all levels (in terms of per-
also worth noting that the employed models, especially HENSM formance indices, score analysis, UA, and statistical testing), which
and RVM, almost modelled the actual output in predicting the ROP in turn indicates that the proposed model has high generalisation
of TBM. capability in handling overfitting related issues of CSC techniques.
1410 A. Bardhan et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1398e1412

Fig. 15. Accuracy matrix. TR e Training; TE e Testing.

Table 6 Table 7
Results of uncertainty analysis. Results of one tailed t-test.

Model MOE SD SE ME LB UB WCB Rank Phase Model Sample HMD Degree of freedom t-stat t- H0
number (dof) critical
MPMR 0.0344 0.0391 0.0053 0.0106 0.0239 0.045 0.0211 5
RVM 0.03 0.034 0.0046 0.0092 0.0208 0.0391 0.0184 2 Training MPMR 130 0 129 2.1526 1.6568 Reject
ELM 0.0301 0.0357 0.0048 0.0096 0.0204 0.0397 0.0193 3 RVM 130 0 129 0.1596 1.6568 Reject
FN 0.0363 0.0388 0.0052 0.0105 0.0258 0.0468 0.021 4 ELM 130 0 129 0.2053 1.6568 Reject
HENSM 0.0282 0.0309 0.0042 0.0084 0.0198 0.0365 0.0167 1 FN 130 0 129 4.3162 1.6568 Reject
Testing MPMR 55 0 54 2.5739 1.6736 Reject
RVM 55 0 54 1.3706 1.6736 Reject
ELM 55 0 54 1.6799 1.6736 Reject
Despite these advantages, the proposed hybrid ensemble technique FN 55 0 54 1.4713 1.6736 Reject
can be improved in the future as follows: (i) incorporation of large
datasets to predict the desired output(s) more accurately; (ii) a
detailed assessment of results of the testing dataset through cross- adding expert endorsed data to cover much broader diversities; (iv)
validation of different conventional machine learning models; (iii) extending the application of the proposed approach in different

Fig. 16. Bar plots of UA showing: (a) MOE, (b) WCB, and (c) ME values.
 A. Bardhan et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1398e1412 1411

Fig. 17. Bar plots of sensitivity analysis for the employed models.

fields; and (v) amalgamation of meta-heuristic optimisation algo- Breiman, L., 1999. Prediction games and arcing algorithms. Neural Comput. 11 (6),
1493e1517.
rithms and in-depth assessment of the proposed model over the
Bui, D.T., Nhu, V.H., Hoang, N.D., 2018. Prediction of soil compression coefficient for
hybrid ANN and meta-heuristics models. Besides, more input pa- urban housing project using novel integration machine learning approach of
rameters should be considered to predict the ROP of TBM more swarm intelligence and multi-layer perceptron neural network. Adv. Eng. Inf.
accurately, particularly in a rock environment. However, this study 38, 593e604.
Cai, R., Han, T., Liao, W., Huang, J., Li, D., Kumar, A., Ma, H., 2020. Prediction of
presents the application of hybrid ensemble techniques for pre- surface chloride concentration of marine concrete using ensemble machine
dicting the ROP of TBM in a rock environment for the first time. learning. Cement Concr. Res. 136, 106164.
Cassinelli, F., Cina, S., Innaurato, N., Mancini, R., Sampaolo, A., 1982. Power Con-
sumption and Metal Wear in Tunnel-Boring Machines: Analysis of Tunnel-
Declaration of competing interest Boring Operation in Hard Rock. In: Tunnelling ’82. Institution of Mining and
Metallurgy, London, UK, pp. 73e81.
Castillo, E., Cobo, A., Gutiérrez, J.M., Pruneda, E., 2000. Functional networks: a new
The authors declare that they have no known competing
network-based methodology. Comput. Aided Civ. Infrastruct. Eng. 15 (2), 90e
financial interests or personal relationships that could have 106.
appeared to influence the work reported in this paper. Castillo, E., Gutiérrez, J.M., 1998. Nonlinear time series modeling and prediction
using functional networks. Extracting information masked by chaos. Phys. Lett.
244 (1e3), 71e84.
References Chatterjee, S., Dash, A., Bandopadhyay, S., 2015. Ensemble support vector machine
algorithm for reliability estimation of a mining machine. Qual. Reliab. Eng. Int.
Ahmed, A., Elkatatny, S., Ali, A., Abdulraheem, A., 2020. Prediction of Lost Circula- 31 (8), 1503e1516.
tion Zones Using Artificial Neural Network and Functional Network. In: Abu Chen, L., Zhang, W., Gao, X., Wang, L., Li, Z., Böhlke, T., Perego, U., 2020. Design
Dhabi International Petroleum Exhibition & Conference. UAE, Abu Dhabi. charts for reliability assessment of rock bedding slopes stability against bi-
Alavi, A.H., Gandomi, A.H., 2011. A robust data mining approach for formulation of planar sliding: SRLEM and BPNN approaches. Georisk. https://doi.org/10.1080/
geotechnical engineering systems. Eng. Comput. Int. J. Comput.-Aided Eng. 28, 17499518.2020.1815215.
242e274. Dietterich, T.G., 2000. Ensemble Methods in Machine Learning. In: Zhou, Z.H.,
Armaghani, D.J., Faradonbeh, R.S., Momeni, E., Fahimifar, A., Tahir, M.M., 2018. Roli, F., Kittler, J. (Eds.), Multiple Classifier Systems: International Workshop on
Performance prediction of tunnel boring machine through developing a gene Multiple Classifier Systems (MCS 2013), vol. 7872, pp. 1e15. Springer, Berlin,
expression programming equation. Eng. Comput. 34, 129e141. Heidelberg, Germany.
Armaghani, D.J., Hajihassani, M., Mohamad, E.T., Marto, A., Noorani, S.A., 2014. Farmer, I.W., Glossop, N.H., 1980. Mechanics of disc cutter penetration. Tunn. Tunn.
Blasting-induced flyrock and ground vibration prediction through an expert 12 (6), 22e25.
artificial neural network based on particle swarm optimization. Arab. J. Geosci. Fattahi, H., 2019. Tunnel boring machine penetration rate. Int. J. Optim. Civil Eng. 9
7, 5383e5396. (2), 343e353.
Armaghani, D.J., Mohamad, E.T., Narayanasamy, M.S., Narita, N., Yagiz, S., 2017. Gandomi, A.H., Yun, G.J., Alavi, A.H., 2013. An evolutionary approach for modeling of
Development of hybrid intelligent models for predicting TBM penetration rate shear strength of RC deep beams. Mater. Struct. 46, 2109e2119.
in hard rock condition. Tunn. Undergr. Space Technol. 63, 29e43. Gao, B., Wang, R., Lin, C., Guo, X., Liu, B., Zhang, W., 2020. TBM penetration rate
Asteris, P.G., Mokos, V.G., 2020. Concrete compressive strength using artificial prediction based on the long short-term memory neural network. Undergr.
neural networks. Neural Comput. Appl. 32, 11807e11826. Space. https://doi.org/10.1016/j.undsp.2020.01.003 (in press).
Asteris, P.G., Skentou, A.D., Bardhan, A., Samui, P., Pilakoutas, K., 2021. Cement and Ghani, S., Kumari, S., 2021. Liquefaction study of fine-grained soil using computa-
concrete research predicting concrete compressive strength using hybrid tional model. Innov. Infrastruct. Solut. 6, 58.
ensembling of surrogate machine learning models. Cement Concr. Res. 145, Ghani, S., Kumari, S., Bardhan, A., 2021a. A novel liquefaction study for fine-grained
106449. soil using PCA-based hybrid soft computing models. S adhana 46, 113.
Bar, A., Rokach, L., Shani, G., Shapira, B., Schclar, A., 2013. Improving Simple Ghani, S., Kumari, S., Choudhary, A.K., Jha, J.N., 2021b. Experimental and compu-
Collaborative Filtering Models Using Ensemble Methods. In: Zhou, Z.H., Roli, F., tational response of strip footing resting on prestressed geotextile-reinforced
Kittler, J. (Eds.), Multiple Classifier Systems: International Workshop on Mul- industrial waste. Innov. Infrastruct. Solut. 6, 98.
tiple Classifier Systems (MCS 2013), vol. 7872. Springer, Berlin, Heidelberg, Ghasemi, E., Yagiz, S., Ataei, M., 2014. Predicting penetration rate of hard rock
Germany. https://doi.org/10.1007/978-3-642-38067-9_1. tunnel boring machine using fuzzy logic. Bull. Eng. Geol. Environ. 73, 23e35.
Bardhan, A., Gokceoglu, C., Burman, A., Samui, P., Asteris, P.G., 2021a. Efficient Gholami, A., Bonakdari, H., Samui, P., Mohammadian, M., Gharabaghi, B., 2019.
computational techniques for predicting the California bearing ratio of soil in Predicting stable alluvial channel profiles using emotional artificial neural
soaked conditions. Eng. Geol. 291, 106239. networks. Appl. Soft Comput. 78, 420e437.
Bardhan, A., Samui, P., Ghosh, K., Gandomi, A.H., Bhattacharyya, S., 2021b. ELM- Goh, A.T.C., Zhang, R.H., Wang, W., Wang, L., Liu, H.L., Zhang, W.G., 2020. Numerical
based adaptive neuro swarm intelligence techniques for predicting the Cali- study of the effects of groundwater drawdown on ground settlement for
fornia bearing ratio of soils in soaked conditions. Appl. Soft Comput. 110, excavation in residual soils. Acta Geotech 15, 1259e1272.
107595. Goh, A.T.C., Zhang, W., Zhang, Y., Xiao, Y., Xiang, Y., 2018. Determination of earth
Basaran, K., Özçift, A., Kılınç, D., 2019. A new approach for prediction of solar ra- pressure balance tunnel-related maximum surface settlement: a multivariate
diation with using ensemble learning algorithm. Arabian J. Sci. Eng. 44, 7159e adaptive regression splines approach. Bull. Eng. Geol. Environ. 77, 489e500.
7171. Goh, A.T.C., Zhang, Y., Zhang, R., Zhang, W., Xiao, Y., 2017. Evaluating stability of
Benardos, A.G., Kaliampakos, D.C., 2004. Modelling TBM performance with artificial underground entry-type excavations using multivariate adaptive regression
neural networks. Tunn. Undergr. Space Technol. 19 (6), 597e605. splines and logistic regression. Tunn. Undergr. Space Technol. 70, 148e154.
1412 A. Bardhan et al. / Journal of Rock Mechanics and Geotechnical Engineering 13 (2021) 1398e1412

Graham, P.C., 1976. Rock Exploration for Machine Manufacturers. In: Exploration for Raja, M.N.A., Shukla, S.K., 2020. An extreme learning machine model for
Rock Engineering: Proceedings of the Symposium on Exploration for Rock geosynthetic-reinforced sandy soil foundations. Proc. Inst. Civil Eng. e Geotech.
Engineering, vol. 1. A. A. Balkema, Cape Town, South Africa, pp. 173e180. Eng. https://doi.org/10.1680/jgeen.19.00297.
Grima, M.A., Bruines, P.A., Verhoef, P.N.W., 2000. Modeling tunnel boring machine Raja, M.N.A., Shukla, S.K., Khan, M.U.A., 2021. An intelligent approach for predicting
performance by neuro-fuzzy methods. Tunn. Undergr. Space Technol. 15 (3), the strength of geosynthetic-reinforced subgrade soil. Int. J. Pavement Eng.
259e269. https://doi.org/10.1080/10298436.2021.1904237.
Huang, G.B., Zhu, Q.Y., Siew, C.K., 2006. Extreme learning machine: theory and Rajasekaran, S., 2004. Functional networks in structural engineering. J. Comput. Civ.
applications. Neurocomputing 70 (1e3), 489e501. Eng. 18 (2), 172e181.
Jagan, J., Samui, P., Kim, D., 2019. Reliability analysis of simply supported beam Ray, R., Kumar, D., Samui, P., Roy, L.B., Goh, A.T.C., Zhang, W., 2021. Application of
using GRNN, ELM and GPR. Struct. Eng. Mech. 71 (6), 739e749. soft computing techniques for shallow foundation reliability in geotechnical
Javad, G., Narges, T., 2010. Application of artificial neural networks to the prediction engineering. Geosci. Front. 12 (1), 375e383.
of tunnel boring machine penetration rate. Min. Sci. Technol. 20 (5), 727e733. Roy, B., Singh, M.P., 2020. An empirical-based rainfall-runoff modelling using
Kardani, N., Bardhan, A., Kim, D., Samui, P., Zhou, A., 2021a. Modelling the energy optimization technique. Int. J. River Basin Manag. 18 (1), 49e67.
performance of residential buildings using advanced computational frameworks Salimi, A., Rostami, J., Moormann, C., Delisio, A., 2016. Application of non-linear
based on RVM, GMDH, ANFIS-BBO and ANFIS-IPSO. J. Build. Eng. 35, 102105. regression analysis and artificial intelligence algorithms for performance pre-
Kardani, N., Bardhan, A., Samui, P., Nazem, M., Zhou, A., Armaghani, D.J., 2021d. diction of hard rock TBMs. Tunn. Undergr. Space Technol. 58, 236e246.
A novel technique based on the improved firefly algorithm coupled with Samadi, M., Sarkardeh, H., Jabbari, E., 2020. Explicit data-driven models for pre-
extreme learning machine (ELM-IFF) for predicting the thermal conductivity of diction of pressure fluctuations occur during turbulent flows on sloping chan-
soil. Eng. Comput. https://doi.org/10.1007/s00366-021-01329-3. nels. Stoch. Environ. Res. Risk Assess. 34, 691e707.
Kardani, N., Zhou, A., Nazem, M., Lin, X., 2021b. Modelling of municipal solid waste Samui, P., 2019. Determination of friction capacity of driven pile in clay using
gasification using an optimised ensemble soft computing model. Fuel 289, Gaussian process regression (GPR), and minimax probability machine regres-
119903. sion (MPMR). Geotech. Geol. Eng. 37, 4643e4647.
Kardani, N., Zhou, A., Nazem, M., Shen, S.L., 2021c. Improved prediction of slope Samui, P., Lansivaara, T., Kim, D., 2011. Utilization relevance vector machine for
stability using a hybrid stacking ensemble method based on finite element slope reliability analysis. Appl. Soft Comput. 11 (5), 4036e4040.
analysis and field data. J. Rock Mech. Geotech. Eng. 13 (1), 188e201. Shahrour, I., Zhang, W., 2021. Use of soft computing techniques for tunneling
Kardani, N., Zhou, A., Nazem, M., Shen, S.L., 2020. Estimation of bearing capacity of optimization of tunnel boring machines. Undergr. Space 6 (3), 233e239.
piles in cohesionless soil using optimised machine learning approaches. Geo- Strohmann, T., Grudic, G.Z., 2003. A Formulation for Minimax Probability Machine
tech. Geol. Eng. 38, 2271e2291. Regression. In: Dietterich, T.G., Becker, S., Ghahramani, Z. (Eds.), Advances in
Koopialipoor, M., Tootoonchi, H., Armaghani, D.J., Mohamad, E.T., Hedayat, A., 2019. Neural Information Processing Systems 14: Proceedings of the 2001 Conference.
Application of deep neural networks in predicting the penetration rate of MIT Press, Cambridge, USA.
tunnel boring machines. Bull. Eng. Geol. Environ. 78, 6347e6360. Tipping, M.E., 1999. The Relevance Vector Machine. In: Solla, S.A., Leen, T.K.,
Kumar, M., Bardhan, A., Samui, P., Hu, J.W., Kaloop, R.M., 2021c. Reliability analysis Müller, K. (Eds.), NIPS’99: Proceedings of the 12th International Conference on
of pile foundation using soft computing techniques: a comparative study. Neural Information Processing Systems. MIT Press, Cambridge, USA, pp. 652e
Processes 9 (3), 486. 658.
Kumar, M., Samui, P., 2019. Reliability analysis of pile foundation using ELM and Wang, L., Wu, C., Tang, L., Zhang, W., Lacasse, S., Liu, H., Gao, L., 2020. Efficient
MARS. Geotech. Geol. Eng. 37, 3447e3457. reliability analysis of earth dam slope stability using extreme gradient boosting
Kumar, R., Samui, P., Kumari, S., Dalkilic, Y.H., 2021a. Reliability analysis of circular footing method. Acta Geotech 15, 3135e3150.
by using GP and MPMR. Int. J. Appl. Metaheuristic Comput. (IJAMC) 12, 1e19. Xiang, Y., Liu, H., Zhang, W., Chu, J., Zhou, D., Xiao, Y., 2018. Application of trans-
Kumar, R., Samui, P., Kumari, S., Roy, S.S., 2021b. Determination of reliability index of parent soil model test and DEM simulation in study of tunnel failure mecha-
cantilever retaining wall by RVM, MPMR and MARS. Int. J. Adv. Intell. Paradigms nism. Tunn. Undergr. Space Technol. 74, 178e184.
18 (3), 316e336. Xu, H., Zhou, J., Asteris, P.G., Armaghani, D.J., Tahir, M.M., 2019. Supervised machine
Li, J., Li, P., Guo, D., Li, X., Chen, Z., 2021. Advanced prediction of tunnel boring learning techniques to the prediction of tunnel boring machine penetration
machine performance based on big data. Geosci. Front. 12 (1), 331e338. rate. Appl. Sci. 9 (18), 3715.
Liu, Z., Shao, J., Xu, W., Wu, Q., 2015. Indirect estimation of unconfined compressive Yagiz, S., Gokceoglu, C., Sezer, E., Iplikci, S., 2009. Application of two non-linear
strength of carbonate rocks using extreme learning machine. Acta Geotech 10, prediction tools to the estimation of tunnel boring machine performance.
651e663. Eng. Appl. Artif. Intell. 22 (4e5), 808e814.
Mahdevari, S., Shahriar, K., Yagiz, S., Shirazi, M.A., 2014. A support vector regression Yagiz, S., Karahan, H., 2011. Prediction of hard rock TBM penetration rate using
model for predicting tunnel boring machine penetration rates. Int. J. Rock Mech. particle swarm optimization. Int. J. Rock Mech. Min. Sci. 48 (3), 427e433.
Min. Sci. 72, 214e229. Zeng, J., Roy, B., Kumar, D., Mohammed, A.S., Armaghani, D.J., Zhou, J.,
Mahmoud, A.A., Elkatatny, S., Ali, A., Abdulraheem, A., Abouelresh, M., 2020. Esti- Mohamad, E.T., 2021. Proposing several hybrid PSO-extreme learning machine
mation of the total organic carbon using functional neural networks and sup- techniques to predict TBM performance. Eng. Comput. https://doi.org/10.1007/
port vector machine. In: International Petroleum Technology Conference, s00366-020-01225-2.
Dhahran, Kingdom of Saudi Arabia. https://doi.org/10.2523/IPTC-19659-MS. Zhang, R., Wu, C., Goh, A.T.C., Böhlke, T., Zhang, W., 2021a. Estimation of diaphragm
Marshall, A.W., Olkin, I., 1960. Multivariate Chebyshev inequalities. Ann. Math. Stat. wall deflections for deep braced excavation in anisotropic clays using ensemble
31 (4), 1001e1014. learning. Geosci. Front. 12 (1), 365e373.
Martins, F.F., Miranda, T.F.S., 2013. Prediction of Hard Rock TBM Penetration Rate Zhang, W., Goh, A.T.C., 2016. Multivariate adaptive regression splines and neural
Based on Data Mining Techniques. In: Proceedings of the 18th International network models for prediction of pile drivability. Geosci. Front. 7 (1), 45e52.
Conference on Soil Mechanics and Geotechnical Engineering. Presses des Ponts, Zhang, W., Li, Y., Wu, C., Li, H., Goh, A.T.C., Liu, H., 2020b. Prediction of lining
Paris, France, pp. 1751e1754. response for twin tunnels constructed in anisotropic clay using machine
Mobarra, Y., Hajian, A., Rahgozar, M., 2013. Application of Artificial Neural Networks learning techniques. Undergr. Space. https://doi.org/10.1016/j.undsp.2020.02.
to the Prediction of TBM Penetration Rate in TBM-Driven Golab Water Transfer 007.
Tunnel. In: International Conference on Civil Engineering Architecture & Urban Zhang, W., Wu, C., Zhong, H., Li, Y., Wang, L., 2021b. Prediction of undrained shear
Sustainable Development. Tabriz, Iran. strength using extreme gradient boosting and random forest based on Bayesian
Mohammadzadeh, S.D., Bolouri Bazaz, J., Alavi, A.H., 2014. An evolutionary optimization. Geosci. Front. 12 (1), 469e477.
computational approach for formulation of compression index of fine-grained Zhang, W.G., Goh, A.T.C., 2013. Multivariate adaptive regression splines for analysis
soils. Eng. Appl. Artif. Intell. 33, 58e68. of geotechnical engineering systems. Comput. Geotech. 48, 82e95.
Qiu, X., Zhang, L., Ren, Y., Suganthan, P.N., Amaratunga, G., 2014. Ensemble Deep Zhang, Y., Wei, M., Su, G., Li, Y., Zeng, J., Deng, X., 2020a. A novel intelligent method
Learning for Regression and Time Series Forecasting. In: 2014 IEEE Symposium for predicting the penetration rate of the tunnel boring machine in rocks. Math.
on Computational Intelligence in Ensemble Learning (CIEL). IEEE, New York, Probl Eng. 2020, 3268694.
USA. https://doi.org/10.1109/CIEL.2014.7015739. Zhou, J., Qiu, Y., Zhu, S., Armaghani, D.J., Li, C., Nguyen, H., Yagiz, S., 2021. Optimi-
Raja, M.N.A., Shukla, S.K., 2021a. Multivariate adaptive regression splines model for zation of support vector machine through the use of metaheuristic algorithms
reinforced soil foundations. Geosynth. Int. 28 (4), 368e390. in forecasting TBM advance rate. Eng. Appl. Artif. Intell. 97, 104015.
Raja, M.N.A., Shukla, S.K., 2021b. Predicting the settlement of geosynthetic- Zorlu, K., Gokceoglu, C., Ocakoglu, F., Nefeslioglu, H.A., Acikalin, S., 2008. Prediction
reinforced soil foundations using evolutionary artificial intelligence technique. of uniaxial compressive strength of sandstones using petrography-based
Geotext. Geomembranes 49 (5), 1280e1293. models. Eng. Geol. 96 (3e4), 141e158.