Professional Documents
Culture Documents
Hoang 2017
Hoang 2017
Hoang 2017
Vector Machines (LSSVM), and the Extreme Machine Learning (ELM). Furthermore, two data sets
with actual slope collapse events have been collected for this study.
The reasons for selecting the aforementioned three approaches are as follows: RBFNN [10,11],
LSSVM [12–15], and ELM [16–19] have been illustrated to be capable pattern classifiers; however,
performances of these three approaches in slope assessment have rarely been discussed in the literature.
In addition, based on a recent comparative work [20], LSSVM has shown superior prediction accuracies
in slope classification; therefore, comparison among the three models can provide helpful information
for readers including both academic researchers and practicing engineers.
The remaining part of this chapter is organized as follows. The second section reviews pertinent
works in the literature. The research framework is described in the third section, followed by the ex-
perimental results. Conclusions of this chapter are stated in the final section.
where cj denotes the coordination of the j th centroid, x represents the input data, x − cj denotes
the norm between the data and the centroid.
Since the task at hand is binary pattern recognition, the output of neuron in the output layer is
converted into binary values through the sigmoid function as follows:
0 if S(u) < t
f (u) = (18.2)
1 if S(u) ≥ t
where S(.) represents the sigmoid function, t denotes a threshold value of 0.2 used to convert the real
value input into binary outputs.
The formula of the sigmoid function is written as follows:
1
S(u) = (18.3)
1 + e−u
It is noted that in RBFNN, the weights between the hidden and output layers wj , the centroid
location cj , and the number of centroids M are determined so that prediction error of the model is
minimized. Thus, a least squares objective function for this learning problem can be defined as follows:
D
2
E= T (i) − y(i) (18.4)
i=1
where T (i) denotes the design output, D is number of training data, and y(i) denotes the network
output.
The network output is computed through a sum product of the network’s weight and the input vector
and it can be expressed in the following form:
M
y= wj zj (x) (18.5)
j =1
336 CHAPTER 18 SLOPE STABILITY EVALUATION USING RADIAL BASIS
n is the number of dimensions of the input data, the corresponding class of labels is denoted as yk ∈
{−1, +1}, the LS-SVM for classification is formulated as follows:
1 2
N
1
Min. Jp (w, e) = w T w + γ ek (18.6)
2 2
k=1
Subject to yk w T φ(xk ) + b = 1 − ek , k = 1, . . . , N (18.7)
where w ∈ R n is the normal vector to the classification hyperplane and b ∈ R is the bias, ek ∈ R are
error variables, and γ > 0 denotes a regularization constant.
The Lagrangian is given by:
N
L(w, b, e, a) = Jp (w, e) − αk yk w T φ(xk ) + b − 1 + ek (18.8)
k=1
where αk are Lagrange multipliers, φ(xk ) represents a kernel function. Applying the KKT conditions
of optimality, the above optimization problem is equivalent to this linear system after the elimination
of e and w:
0 yT b 0
−1
= (18.9)
y ω+γ I α 1 v
where αk and b denote the solution to the linear system (Eq. (18.9)). The kernel function that is com-
monly used is Radial Basis Function (RBF) kernel. The RBF kernel is described as follows:
xk − x1 2
K(xb , x1 ) = exp − (18.11)
2σ 2
where σ is the kernel function parameter.
assigned and remain constant during training and predicting phases. On the contrary, the weights that
connect hidden nodes to outputs can be trained very fast. Experimental studies in the literature [16,37,
38] showed that ELMs can produce acceptable predictive performance and their computational cost is
much lower than networks trained by the back-propagation algorithm.
The task at hand is to construct a classification model from a data set X = {xt ∈ R p }, t = 1, . . . , n,
with n samples and p input features. Given a network with p input units, q hidden neurons, and c
outputs, the ELM model’s output is written in the following formula [39]:
where mi ∈ R q , i ∈ {1, . . . , c} denotes the weight vector that connects the hidden neurons to the ith
output neuron. h(t) ∈ R q represents the vector of outputs of hidden neurons for a certain input pattern
x(t) ∈ R p . Then h(t) can be written in the following form:
h(t) = f w1T x(t) + b1 , f w2T x(t) + b2 , . . . , f wqT x(t) + bq (18.13)
where bk (k = 1, 2, . . . , q) denotes the bias of the kth hidden neuron, wk ∈ R p represents the weight
vector of the kth hidden neuron, and f (.) denotes a sigmoidal activation function. It is worth to notice
that the weight vectors wk as well as the bias bk are generated from a Gaussian distribution in a random
manner.
Providing wk and bk , the next step is to establish a matrix of hidden layer output H . It is noted
that H is q × n matrix; its t th column is the vector of a hidden layer output h(t). Accordingly, the
weight matrix M = [m1 , m2 , . . . , mc ]can be calculated via the Moore–Penrose pseudo inverse method
as follows:
−1
M = H × HT H × DT (18.14)
where D = [d(1), d(2), . . . , d(n)] denotes a c × n matrix whose t th column is the actual target vector
d(t) ∈ R c .
With the network’s parameter being fully specified, the class label for a new input pattern is deter-
mined as follows:
Y = arg max{oi } (18.15)
i=1,...,c
Table 18.1 Slope Influencing Factors and Their Statistical Descriptions of Data Set 1
Factors Notation Definition Max Average Std. Min
X1 γ Unit weight (kN/m3 ) 31.30 21.76 4.13 12.00
X2 C Soil cohesion (kPa) 300.00 34.12 45.82 0.00
X3 ϕ Internal friction angle (°) 45.00 28.72 10.58 0.00
X4 β Slope angle (°) 59.00 36.10 10.22 16.00
X5 H Slope height (m) 511.00 104.19 132.68 3.60
X6 Ru Pore pressure ratio 45.00 0.48 3.45 0.00
Table 18.2 Slope Influencing Factors and Their Statistical Descriptions of Data Set 2
Factors Definition Max Average Std. Min
X1 Slope direction (°) 345.00 171.33 93.95 0.00
X2 Slope angle (°) 90.00 62.20 11.74 30.00
X3 Slope height (m) 100.00 22.39 14.86 5.00
X4 Road curvature (1/m) 0.03 0.00 0.02 −0.05
X5 Strata type 5.00 4.40 1.06 1.00
X6 Thickness of canopy cover (m) 4.50 2.07 1.05 0.50
X7 Catchment area (m2 ) 255743.00 19706.81 39956.69 406.00
X8 Height of toe cutting (m) 50.00 7.00 6.52 2.00
X9 Change of slope grade (°) 35.00 9.08 10.36 0.00
X10 Peak ground acceleration (gal) 391.90 251.00 115.71 0.00
Table 18.1 provides the information of the influencing factors and their statistical descriptions.
Table 18.2 provides the information of the influencing factors and their statistical descriptions of Data
Set 2. The data sets are provided in Appendices 1 and 2, within which the output of −1 indicates
a non-collapsed slope and the output of +1 denotes a collapsed slope. For more detailed explanation
regarding the influencing factors of slope, the readers are guided to previous works of Hoang and Pham
[20] for Data Set 1 and Ching et al. [32] for Data Set 2. Furthermore, scatter plots of all input variables
of the two data sets with class label distinction are plotted in Fig. 18.1 and Fig. 18.2. A preliminary
observation from Fig. 18.1 and Fig. 18.2 is that there is a high degree of overlapping regions within
each input feature of the two data sets.
FIGURE 18.1
Data distribution of Data Set 1.
FIGURE 18.2
Data distribution of Data Set 2.
340 CHAPTER 18 SLOPE STABILITY EVALUATION USING RADIAL BASIS
In case of RBFNN, to determine the parameters M and cj , the method of Orthogonal Least Squares
(OLS) is employed [40,41]. In OLS, each data point is initially set as a possible location of the centroid
(cj ); the assessment includes computing the network error for each data point in the training set as the
new centroid of the corresponding cluster. The data point that can reduce the RBFNN error the most
is the new centroid [10]. This assessment procedure is repeated until the network error reaches an
acceptable value. In our study, the threshold value of the RBFNN training phase is set to be 90%. That
is, the training process of the RBFNN will terminate when the classification accuracy rate reaches 0.9.
Our observation is that setting a lower value of threshold results in under-trained models; meanwhile,
setting a higher value causes the problem of overfitting. In this study, the RBFNN model is coded in
Matlab by the authors.
To construct an LSSVM model, it is necessary to specify the regularization constant (γ ) and the
kernel function parameter (σ ). Previous works [20,42] point out that these two parameters affect the
learning performance of LSSVM considerably. Therefore, in this study, a grid search procedure is used
to appropriately set the values of the regularization constant (γ ) and the kernel function parameter (σ ).
These two parameters are allowed to be varied within the following set of values: [0.01, 0.05, 0.1, 0.5,
1, 5, 10, 50, 100, 500, 1000]. It is noted that the training set is further divided into two sets: Set 1
(90%) used for model construction and Set 2 (10%) used for computing the fitness of each pair of
hyper-parameters. It is noted that the LSSVM model is implemented via the LS-SVMlab Toolbox [43]
and the grid search procedure is coded by the authors.
In case of ELM, the sigmoid function is often employed as activation function. The only hyper-
parameter to be set for this model is the number of neurons in the hidden layer. Herein, the number of
neurons is allowed to vary from the number of input features to 100. For the purpose of model selection,
the training data is also split into two sets (Set 1 and Set 2) with a similar manner to the process used
in parameter setting of LSSVM. The number of neurons that maximize ELM performance is selected
for the testing phase. The ELM model is implemented in Matlab environment with the program codes
provided by Huang [44].
Furthermore, besides the classification accuracy rate (CAR), the following four metrics can be used
to measure the classification performance [15]: true positive rate TPR (the percentage of positive in-
stances correctly classified), true negative rate TNR (the percentage of negative instances correctly
classified), false positive rate FPR (the percentage of negative instances misclassified), and false nega-
tive rate FNR (the percentage of positive instances misclassified). The formulations for computing the
above four metrics are stated as follows:
TP
TPR = (18.16)
TP + FN
TN
TNR = (18.17)
TN + FP
FP
FPR = (18.18)
FP + TN
FN
FNR = (18.19)
TP + FN
where TP, TN, FP, and FN are the numbers of true positive, true negative, false positive, and false
negative, respectively.
18.5 EXPERIMENTAL RESULTS 341
FIGURE 18.3
True positive, true negative, false positive, false negative rates for Data Set 1.
FIGURE 18.4
True positive, true negative, false positive, false negative rates for Data Set 2.
18.6 CONCLUSION
This chapter has investigated the capabilities of RBFNN, LSSVM, and ELM in slope stability as-
sessment with two historical data sets. To accurately evaluate each model’s performance, a ten-fold
cross-validation process has been carried out. Results obtained from experiments demonstrate that
LSSVM and ELM are superior methods for tackling the problem at hand. The performance of RBFNN
is significantly worse than that of the other two models. Considering the CAR, LSSVM is the best
REFERENCES 343
method for Data Set 1; and ELM is the most desirable method for Data Set 2. Nevertheless, when
taking FNR into account, LSSVM deems to be the most suitable for slope evaluation since this method
has resulted in the lowest FNR in both data sets. Overall, LSSVM and ELM are highly recommended
as an intelligent tool to assist decision-making process in slope assessment. Further directions of the
current study may include: (1) Investigate other advanced machine learning approaches in slope eval-
uation, (2) Combine feature selection technique with LSSVM and ELM, and (3) Enhancing prediction
accuracy by ensemble and boosting methods.
REFERENCES
[1] H.B. Wang, W.Y. Xu, R.C. Xu, Slope stability evaluation using Back Propagation Neural Networks, Eng. Geol. 80 (2005)
302–315.
[2] M.-Y. Cheng, N.-D. Hoang, Slope collapse prediction using Bayesian framework with K-Nearest Neighbor density estima-
tion: case study in Taiwan, J. Comput. Civ. Eng. 30 (2016) 04014116.
[3] H.-M. Lin, S.-K. Chang, J.-H. Wu, C.H. Juang, Neural network-based model for assessing failure potential of highway
slopes in the Alishan, Taiwan Area: pre- and post-earthquake investigation, Eng. Geol. 104 (2009) 280–289.
[4] A. Manouchehrian, J. Gholamnejad, M. Sharifzadeh, Development of a model for analysis of slope stability for circular
mode failure using genetic algorithm, Environ. Earth Sci. 71 (2014) 1267–1277.
[5] N.-D. Hoang, D. Tien-Bui, A novel relevance vector machine classifier with cuckoo search optimization for spatial predic-
tion of landslides, J. Comput. Civ. Eng. 30 (2016) 04016001.
[6] C.-I. Wu, H.-Y. Kung, C.-H. Chen, L.-C. Kuo, An intelligent slope disaster prediction and monitoring system based on
WSN and ANP, Expert Syst. Appl. 41 (2014) 4554–4562.
[7] E. Salmi, S. Hosseinzadeh, Slope stability assessment using both empirical and numerical methods: a case study, Bull. Eng.
Geol. Environ. 74 (2015) 13–25.
[8] P. Luciano, L. Serge, Assessment of slope stability, in: Geotechnical Engineering State of the Art and Practice, 2012,
pp. 122–156.
[9] H. Zhao, S. Yin, Z. Ru, Relevance vector machine applied to slope stability analysis, Int. J. Numer. Anal. Meth. Geomech.
36 (2012) 643–652.
[10] K.-W. Liao, J.-C. Fan, C.-L. Huang, An artificial neural network for groutability prediction of permeation grouting with
microfine cement grouts, Comput. Geotech. 38 (2011) 978–986.
[11] D. Tien Bui, D.T. Quach, V.H. Pham, I. Revhaug, V.L. Ngo, T.H. Tran, et al., Spatial prediction of landslide hazard along
the National Road 32 of Vietnam: a comparison between Support Vector Machines, Radial Basis Function neural networks,
and their ensemble, in: Proceedings of the Thematic Session, 49th CCOP Annual Session, 22–23 October 2013, Sendai,
Japan, 2013.
[12] P. Samui, J. Karthikeyan, Determination of liquefaction susceptibility of soil: a least square support vector machine ap-
proach, Int. J. Numer. Anal. Meth. Geomech. 37 (2013) 1154–1161.
[13] P. Samui, D.P. Kothari, Utilization of a least square support vector machine (LSSVM) for slope stability analysis, Sci. Iran.
18 (2011) 53–58.
[14] M.-Y. Cheng, N.-D. Hoang, Groutability prediction of microfine cement based soil improvement using evolutionary LS-
SVM inference model, J. Civ. Eng. Manag. 20 (2014) 1–10.
[15] N.-D. Hoang, D. Tien Bui, Predicting earthquake-induced soil liquefaction based on a hybridization of kernel Fisher dis-
criminant analysis and a least squares support vector machine: a multi-dataset study, Bull. Eng. Geol. Environ. (2016)
1–14.
[16] G. Huang, G.-B. Huang, S. Song, K. You, Trends in extreme learning machines: a review, Neural Netw. 61 (2015) 32–48.
[17] D. Avci, A. Doğantekin, An expert diagnosis system for Parkinson disease based on genetic wavelet kernel extreme learning
machine, Parkinson’s Dis. 2016 (2016) 5264743.
[18] P. Samui, J. Jagan, R. Hariharan, An alternative method for determination of liquefaction susceptibility of soil, Geotech.
Geol. Eng. 34 (2016) 735–738.
[19] O. Anicic, S. Jović, H. Skrijelj, B. Nedić, Prediction of laser cutting heat affected zone by extreme learning machine, Opt.
Laser Eng. 88 (2017) 1–4.
344 CHAPTER 18 SLOPE STABILITY EVALUATION USING RADIAL BASIS
[20] N.-D. Hoang, A.-D. Pham, Hybrid artificial intelligence approach based on metaheuristic and machine learning for slope
stability assessment: a multinational data analysis, Expert Syst. Appl. 46 (2016) 60–68.
[21] F. Kang, J. Li, Artificial bee colony algorithm optimized support vector regression for system reliability analysis of slopes,
J. Comput. Civ. Eng. 30 (2015) 04015040.
[22] P. Lu, M.S. Rosenbaum, Artificial neural networks and grey systems for the prediction of slope stability, Nat. Hazards 30
(2003) 383–398.
[23] K.-p. Zhou, Z.-Q. Chen, Stability prediction of tailing dam slope based on neural network pattern recognition, in: Proc. of
the Second International Conference on Environmental and Computer Science, ICECS ’09, 28–30 Dec. 2009, Dubai, the
United Arab Emirates, 2009, pp. 380–383.
[24] J.-P. Jiang, BP neural networks for Prediction of the factor of safety of slope stability, in: Proc. of the International Confer-
ence on Computing, Control and Industrial Engineering, CCIE, 20–21 Aug. 2011, Wuhan, China, 2011.
[25] S.K. Das, R.i. Biswal, N. Sivakugan, B. Das, Classification of slopes and prediction of factor of safety using differential
evolution neural networks, Environ. Earth Sci. 64 (2011) 201–210.
[26] P. Samui, Slope stability analysis: a support vector machine approach, Environ. Geol. 56 (2008) 255–267.
[27] J. Li, F. Wang, Study on the forecasting models of slope stability under data mining, in: Proc. of the Earth and Space
2012: Engineering, Science, Construction, and Operations in Challenging Environments, Honolulu, Hawaii, United States,
ASCE, 2010, pp. 765–776.
[28] J. Li, M. Dong, Method to predict slope safety factor using SVM, in: Proc. of the Earth and Space 2012: Engineering,
Science, Construction, and Operations in Challenging Environments, Pasadena, California, United States, ASCE, 2012,
pp. 888–899.
[29] D. Tien Bui, B. Pradhan, O. Lofman, I. Revhaug, Landslide susceptibility assessment in Vietnam using support vector
machines, decision tree, and naive Bayes models, Math. Probl. Eng. 2012 (2012) 26.
[30] M.-Y. Cheng, N.-D. Hoang, Typhoon-induced slope collapse assessment using a novel bee colony optimized support vector
classifier, Nat. Hazards 78 (2015) 1961–1978.
[31] A. Ahangar-Asr, A. Faramarzi, A.A. Javadi, A new approach for prediction of the stability of soil and rock slopes, Eng.
Comput. 27 (2010) 878–893.
[32] J. Ching, H.-J. Liao, J.-Y. Lee, Predicting rainfall-induced landslide potential along a mountain road in Taiwan, Geotech-
nique 61 (2011) 153–166.
[33] X. Yan, X. Li, Bayes discriminant analysis method for predicting the stability of open pit slope, in: Proc. of the International
Conference on Electric Technology and Civil Engineering, ICETCE, 22–24 April 2011, Lushan, China, 2011, pp. 147–150.
[34] M.-Y. Cheng, N.-D. Hoang, A Swarm-Optimized Fuzzy Instance-based Learning approach for predicting slope collapses
in mountain roads, Knowl.-Based Syst. 76 (2015) 256–263.
[35] S. Chen, C.F.N. Cowan, P.M. Grant, Orthogonal least squares learning algorithm for radial basis function networks, IEEE
Trans. Neural Netw. 2 (1991) 302–309.
[36] J. Suykens, J.V. Gestel, J.D. Brabanter, B.D. Moor, J. Vandewalle, Least Square Support Vector Machines, World Scientific
Publishing Co. Pte. Ltd., Singapore, 2002.
[37] G.-B. Huang, Q.-Y. Zhu, C.-K. Siew, Extreme learning machine: theory and applications, Neurocomputing 70 (2006)
489–501.
[38] G. Li, P. Niu, Y. Ma, H. Wang, W. Zhang, Tuning extreme learning machine by an improved artificial bee colony to model
and optimize the boiler efficiency, Knowl.-Based Syst. 67 (2014) 278–289.
[39] A.S.C. Alencar, A.R. Rocha Neto, J.P.P. Gomes, A new pruning method for extreme learning machines via genetic algo-
rithms, Appl. Soft Comput. 44 (2016) 101–107.
[40] F.M. Ham, I. Kostanic, Principles of Neurocomputing for Science and Engineering, McGraw-Hill, New York, United States,
2001.
[41] V. Kecman, Learning and Soft Computing: Support Vector Machines, Neural Networks, and Fuzzy Logic Models, The MIT
Press, Cambridge, 2001.
[42] G.S. Dos Santos, L.G.J. Luvizotto, V.C. Mariani, L. dos Santos Coelho, Least squares support vector machines with tuning
based on chaotic differential evolution approach applied to the identification of a thermal process, Expert Syst. Appl. 39
(2012) 4805–4812.
[43] K. De Brabanter, P. Karsmakers, F. Ojeda, C. Alzate, J. De Brabanter, K. Pelckmans, et al., LS-SVMlab Toolbox User’s
Guide Version 1.8, Internal Report 10-146, ESAT-SISTA, K.U. Leuven, Leuven, Belgium, 2010.
[44] G.-B. Huang, Basic ELM algorithms, http://www.ntu.edu.sg/home/egbhuang/elm_codes.html, 2016.