79

Parameter Optimization using GA in

SVM to Predict Damage Level of Non-
Reshaped Berm Breakwater
N. Harish1, Lokesha2, S. Mandal3, Subba Rao4 And S.G. Patil5
1Assistant Professor, CET, Jain University, Jakksandra Post, Ramanagara District
562112, Karnataka, India,
E-mail:harishnnitk@gmail.com
2Reseach Scholar, Department of Ocean Engineering, Indian Institute of Technology

Madras, Chennai, Tamil Nadu, 600036, INDIA, E-mail:lokesha.iitm@gmail.com

3Chief Scientist, Ocean Engineering Division, National Institute of Oceanography, Dona

Paula, 403004, Goa, India, E-mail:smandal@nio.org

4Professor,Department of Applied Mechanics and Hydraulics, National Institute of

Technology, Surathkal, Karnataka, 575025, India, E-mail:surakrec@gmail.com

5Professor, Department of Built and Natural Environment, Caledonian College of

Engineering, PO Box: 2322, CPO Seeb, PC 111, Sultanate of Oman, E-mail:

Sanras5@gmail.com

Received: March 17, 2014; Accepted: May 19, 2014

Abstract
In the present study, Support Vector Machines (SVM) and hybrid of Genetic Algorithm
(GA) with SVM models are developed to predict the damage level of non-reshaped berm
breakwaters. Optimal kernel parameters of SVM are determined by using GA algorithm.
The models are trained and tested on the data set obtained from the experiments which
were carried out at Marine Structures Laboratory, Department of Applied Mechanics and
Hydraulics, National Institute of Technology Karnataka, Surathkal, India. The results of
SVM and GA-SVM models are compared in terms of statistical measures like correlation
coefficient, root mean square error and scatter index. The results on SVM and GA-SVM
models reveals that the performance of GA-SVM is better compared to SVM models in
predicting the damage level of non-reshaped berm breakwater.

Keywords: Hybrid model, Polynomial kernel function, Statistical parameters.

1. INTRODUCTION
The berm breakwater is a rubble mound structure with the presence of horizontal berm at or above still
water level (SWL). During the wave attack, non-reshaped structures i.e., statically stable structures
under goes no or minor damage. According to Van der Meer (1988), the term ‘’Damage level’’ is
defined as the displacement of armor units. The breakwaters are constructed parallel to the shore in
order to protect the coast and harbors against wave action. They are also used for dual purposes like
dissipating wave energy and providing loading and unloading facilities for cargo and passengers. In the
past, several researchers (Priest et al 1964, Van der Meer 1988, 1992, Tørum et al 2003, Subba Rao et
al 2004, 2008) experimented and developed physical models which are time consuming and expensive
in terms of cost. Apart from this, they also failed to give a simple mathematical model to predict the
damage level considering all the boundary conditions. This may be due to vagueness and complexity
associated with many design parameters of berm breakwater. To minimize the cost, time and
complexity in designing physical models, soft computing tools, such as Artificial Neural Network
(ANN), Support Vector Machine (SVM), Adaptive Neuro Fuzzy Inference System (ANFIS), etc., are

Volume 5 · Number 2 · 2014

80 Parameter Optimization using GA in SVM to Predict Damage Level
of Non-Reshaped Berm Breakwater

successfully used in different fields (Kazperkiewiecz et al 1995, Voga and Belchior 2006, Dong et al
2005). Also in coastal fields some works have been carried out using soft computing tools.
Mase et al (1995) applied neural network technique to predict the stability and damage level of
rubble mound breakwater. They found that the predicted damage levels are in agreement with the
measured damage level which was experimentally carried out by Van der Meer (1988) and Smith et al
(1992). Yagci et al (2005) used neural network technique to predict damage ratio of breakwater.
Physical model estimated damage ratio values matched the predicted values by neural network method.
Further they revealed that the ANN application facilitates the consideration of wave period, wave
steepness, slope of the breakwater and wave height in estimating damage ratio. Kim and Park (2005)
applied ANN method to design the rubble mound breakwater. According to them the neural network
technique gave more accurate results than the conventional empirical model and the extent of accuracy
was affected by the structure of neural network.
Mahjoobi and Mosabbeb (2009) used SVM for the prediction of significant wave height. Their result
shows that the SVM can be successfully used for the prediction of significant wave height. Kim et al
(2010) used SVM to predict the stability number of armour blocks of breakwaters. The proposed
method proves to be an effective tool for designers to support their decision process and to improve
design efficiency of rubble mound breakwaters. Balas et al (2010) applied hybrid model for the
preliminary design of rubble mound breakwater. Their hybrid model has shown better agreement
between the predicted and measured values when compared to ANN and stability equations obtained
by Van der Meer (1988). Patil et al (2011) developed ANFIS model for predicting wave transmission
coefficient of horizontally interlaced multilayer moored floating pipe breakwater and showed that
ANFIS models outperformed ANN models for predicting transmitting waves.
However, it is observed that there are hardly any applications of hybrid GA-SVM model in
predicting the damage level for non-reshaped berm breakwaters. In the present paper, GA is used to
optimize the SVM and kernel parameters. Results of GA-SVM models are compared with that of SVM
models.

2. EXPERIMENTAL DATA
The experimental work was carried out in Marine Structures Laboratory, Department of Applied
Mechanics and Hydraulics, National Institute of Technology Karnataka (NITK), Surathkal, India. The
wave flume is 50m long, 0.71m wide, 1.1m deep, and has 42m long smooth concrete bed. Figure.1
shows a sketch of the wave flume.
Four sets of experiments were carried out for 3000 waves (Subba Rao et al 2004). In the first set of
experiments, stability for different wave periods and height on conventional breakwater model with
trapezoidal cross section with armour stone weight W50 = 74gms was tested. In the second set of
experiments, statically stable non-reshaped berm breakwater model was tested with the armour stones
weight W50 = 52gms which is about 30% less than 74gms. They studied the influence of berm width
on the stability of the breakwater, run-up and rundown. The third set of experiments was tested with
armour stones weight W50 = 58.6gms which is about 20% less than 74gms. The influence of tidal effect
and stability was studied by changing the depth of water in front of the breakwater model. In the fourth
set of experiments the influence of location of the berm and stability was studied by keeping the armour
stones weight W50 = 52gms, the weight used in conventional breakwater (Subba Rao et al 2008). Range
of experimental variables is shown in Table 1. Many problems involving fluid motions are quite
complex in nature. In the present case the complex flow phenomenon responsible for energy dissipation
cannot be easily represented by mathematical equations and one has to rely on experimental
investigations. The results of such investigations are more useful when expressed in the form of
dimensionless relation. To arrive at such dimensionless relation of different variables, dimensional
analysis is carried out. After conducting the dimensional analysis using Buckingham’s-II theorem the
dimensionless parameters, namely wave steepness (H/L0), surf similarity(ζ), relative berm position by
water depth (hB/d), armour stone weight (W50/W50max), relative berm width (B/ L0) and relative berm
location (hB/L0) are obtained.

International Journal of Ocean and Climate Systems

N. Harish, Lokesha, S. Mandal, Subba Rao And S.G. Patil 81

Figure 1. Experimental set-up

Table 1. Range of Experimental Variables

Variable Range
Wave height H (m) 0.10, 0.12, 0.14, 0.16
Wave period T (Sec) 1.6, 2.0, 2.6
Water depth above the bed level d (m) 0.25, 0.30, 0.35, 0.40
Water depth above or below the berm dB (m) +0.08,+0.03, -0.02,-0.07
Armour stone weight W50 (gm) 52 to 74
Crest height above the seabed (m) 0.7
Berm width B (m) 0.6
Berm position above seabed hB (m) 0.32

For the present damage level analysis, experimental data are divided into two sets, one for training
(86 data set) and another for testing (24 data set). The input parameters that influence the damage level
(S) of non-reshaped berm breakwater are H/L0, ζ, hB/d, W50/W50max, B/ L0 and hB/L0 which are used to
train SVM and GA-SVM models.

3. GENETIC ALGORITHM FOR PARAMETER OPTIMIZATION IN SUPPORT

VECTOR MACHINE
3.1 Support Vector Machines
The foundation of SVM has been developed by Vapnik (1995) and is gaining popularity due to many
attractive features and promising empirical performance. SVMs were developed to solve the
classification problem, but recently they have been extended to the domain of regression problems
[Vapnik et al., 1996]. The basic concept of support vector regression is to map nonlinearly the original
data into a higher dimensional feature space and solve a linear regression problem in this feature space.
By considering a training data set g = {( x1 , y1 ) , ( x2 , y2 ) ,......( x p , y p )} , such that xi ∈ v N is a vector of input

Volume 5 · Number 2 · 2014

82 Parameter Optimization using GA in SVM to Predict Damage Level
of Non-Reshaped Berm Breakwater

variables and xi ∈ v is the corresponding scalar output (target) value. Here, the modeling objective is
to find a regression function, y = f ( x ) , such that it accurately predicts the outputs corresponding to a
new set of input-output examples, {(x, y)}. Support vector regression (SVR) considers the following
linear equation:
f ( x ) = (w ⋅ x ) + b (1)

Where, w and b are weight and bias. f(x) denotes a function termed feature, and (w · x) represents the
dot product in the feature space, l, such that ϕ : x → l, w ∈ l. The regression problem is equivalent to
minimize the following regularized risk function:

1 n 1
R( f ) = ∑ L ( f ( xi ) − yi ) + 2 w
2
(2)
n i

Where,
 f ( x ) − y − ε , for f ( x ) − y x ≥ ε 
L ( f ( x ) − y) =   (3)
 0, otherwise 

eqn [3] is also called-insensitive loss function. This function defines a ε -tube. If the predicted value is
within the ε -tube, the loss is zero or else the loss is equal to the magnitude of the difference between
the radius ε of the tube and the predicted value. The radius of the tube located around the regression
function is represented by a precision parameter ε and the ‘’ε -intensive zone’’ is the region enclosed
by the tube.
The SVM algorithm attempts to position the tube around the data as shown in Figure 2. By
substituting the ε-insensitive loss function into eqn [2], the optimization object becomes:

w C + ∑ (ξ i + ξ i∗ )
1 2 n
Minimize (4)
2 i

With the constraints,

 yi − ( w ⋅ x ) − b ≤ ε + ξ i 
 
Subjecyd  ( w ⋅ x ) + b − yi ≤ ε + ξ i∗  (5)
 
 ξ i , ξ i∗ ≥ 0 

Where, the constant C > 0 stands for the penalty degree of the sample with error exceeding ε. The
distance from actual values to the corresponding boundary values of ε -tube is represented by the two
positive slack variables.
The SVM fits f(x) to the data in a manner such that: Minimizing the slack variables i.e., ξ i , ξ i∗ the
training error is minimized and, to increase the flatness of f(x) or to penalize over complexity of the
fitting is w 2 minimized.
The performance of SVM regression depends on the good setting of SVM and kernel parameters. In
the present study, quadratic loss function is used. The main idea of using this loss function is to ignore
the errors, which are situated within the certain distance of the true value. The kernel function used in
present study is shown in Table 2.

International Journal of Ocean and Climate Systems

N. Harish, Lokesha, S. Mandal, Subba Rao And S.G. Patil 83

Figure 2. A schematic diagram of support vector regression using e-insensitive loss function

Table 2. Kernel Functions

3.2 Genetic algorithm based feature selection and parameters optimization for
support vector machines
Genetic algorithms (GA) are well suited for searching global optimal values in complex search space,
coupled with the fact that they work with raw objectives when compared with conventional techniques
(Holland 1975, Goldberg 1989). The algorithm is based on the Darwin’s theory, in which the genetic
information is carried and retained from generation to generation. The main purpose of using GA based
parameter selection procedure is to reduce the input space of the model and thus improves the
performance of the model. In the presesnt study, GA is used to find the optimal values of SVM (C, ε)
and kernel parameters (d, γ), which improves the performance of the model.

The working procedure of GA-SVM is described as below:

Population size: In the present study, a size of 30 intial population with the randomly generated
chromosomes is used. The SVM parameters C, ε and kernel parameters d, γ are directly coded to
generate chromosome.

Fitness definition: The fitness value of each chromosome is calculated using a negative normalised
mean square given in eqns [6 and 7].

Volume 5 · Number 2 · 2014

84 Parameter Optimization using GA in SVM to Predict Damage Level
of Non-Reshaped Berm Breakwater

1
∑
N
( xi − yi )
2
Fitness Function = (6)
−σ N
2 i =1

Where,
1 N
(
∑ xi − x i )
2
σ2 = (7)
N i =1

N is the total number of data in the test set, x–1 is the mean of the actual value, xi is the actual value, and
yi is the predicted value.

Selection: The selection of better parent chromosome from the current population for reproduction is
based on roulette wheel selection principle (Holland, 1975).
Crossover: The binary crossover is applied to the randomly paired chromosomes to form new
offsprings. The intermediate crossover principle in producing offsprings is given by eqn [8].

Offspring = parent1 ± α ( parent 2 − parent1) (8)

is the scaling factor within the range (-d, 1+d) and where, d is taken as 0.25 for the present study.

Mutation: The mutation operation is followed by crossover operation which helps in preventing the
population hibernation at any local optima. The randomly selected string contains variable which
undergoes mutation using incremental operator. Crossover and mutation occurs during evolution
depending on a user defined probabilities, which are taken as 0.8 and 0.5 for crossover and mutation
respectively for the present study.

Accept and Replace: Offspring replaces the old population using the diversity replacement startegy and
generates the new population used for further run of the algorithm.

Stopping criteria: If the stopping criteria is satisfied, the process ends and return the best solution in
current population; otherwise proceed with next generation.
In the present study, SVM is combined with genetic algorithm using Matlab to find the optimal values
of SVM and kernel parameters simultaneously for better generalization of the proposed model. In the
first stage, training input, training target, kernel functions, range of kernel and SVM parameters are fed
to the system. GA generates the initial population that would be used to find optimum factors of kernel
functions and SVMs. In the second stage, the system performs typical SVM process using assigned
value of the factors in the chromosomes, and calculates the performance of each chromosome using
fitness function given in Equation 6. If the calculated fitness value satisfies the terminal condition in
GA, then the optimal parameters are selected; otherwise, in the third stage, new population is generated
by applying the genetic operators such as selection, crossover and mutation. After the production of
new generation, the training process with calculation of fitness value is performed again. From this
point, stage two and three are iterated until stopping condition is satisfied. Once the stopping condition
is satisfied, the genetic search finishes and the chromosome that shows the best performance in the last
population is selected as the final result. In the fourth and final stage, optimized parameters obtained
by GA are tested with the testing data. The final decision about the optimum models is not based on the
training data, but on the testing data.

International Journal of Ocean and Climate Systems

N. Harish, Lokesha, S. Mandal, Subba Rao And S.G. Patil 85

4. RESULTS
In the present paper, SVM and GA-SVM models are developed for prediction of damage level of non-
reshaped berm breakwater. The performances of these models are studied based on the statistical
measures namely correlation coefficient (CC), root-mean-square error (RMSE) and scatter index, which
are defined as:

∑ (O − O )( P − P )
n
i i i i
CC = i =1
(9)
∑ (O − O ) ( P − P )
n 2 2

i =1 i i i i

1 n
∑ (Oi − Pi ) × 100%
2
RMSE = (10)
n i=1

RMSE
SI = (11)
Oi

Where, Oi and Pi are observed and predicted damage level respectively, n is the number of data set used
– and P– are average observed and predicted damage level respectively.
and O i i
The better selection of SVM and kernel parameters decides the performance of these models. The
number of support vectors (nsv), kernel parameters (d, γ) and SVM parameters (C, ε) are used in SVM
and GA-SVM as shown in Table 3. The results obtained during training and testing processes for input
parameters are evaluated using statistical measures like CC, RMSE and SI values as shown in Table 4
and Table 5 for SVM and GA-SVM models respectively. Among all kernel functions polynomial gives
better results. The optimal value of d in case of polynomial kernel function obtained by manual search
is 3. The SVM model with polynomial kernel function gives generalization performance with CC =
0.9081 and 0.8883, RMSE= 2.8811 and 3.2357, SI = 0.2542 and 0.2674 for training and testing
respectively. The results on GA-SVM model with polynomial kernel function show improved
generalization performance with CC = 0.9100 and 0.9074, RMSE= 2.6949 and 2.4828, SI = 0.2378 and
0.2349 for training and testing respectively. The comparison of predicted and observed damage level
for SVM and GA-SVM model with polynomial kernel is shown in Figure 3 and 4 respectively.

a b

Figure 3. Comparison of predicted and observed damage level for SVM model with polynomial kernel
for (a) Training &(b) Testing

Volume 5 · Number 2 · 2014

86 Parameter Optimization using GA in SVM to Predict Damage Level
of Non-Reshaped Berm Breakwater

a b

Figure 4. Comparison of predicted and observed damage level for GA-SVM model with polynomial
kernel for (a) Training & (b)Testing

Table 3. Optimal parameters for SVM and GA-SVM models for different kernel function

Table 4. Statistical Measures of SVM model for different kernel function

Training Data Testing Data

Kernel Training Testing
functions CC CC
RMSE SI RMSE SI

polynomial 0.9081 0.8883 2.8811 0.2542 3.2357 0.2674

rbf 0.8795 0.8665 3.1259 0.2758 3.4569 0.2857
erbf 0.7824 0.7697 4.5877 0.4048 4.9127 0.4060
spline 0.8983 0.8609 2.9335 0.2589 3.4851 0.2880

Table 5. Statistical Measures of GA-SVM model for different kernel function

Training Data Testing Data

Kernel Training Testing
functions CC CC
RMSE SI RMSE SI

polynomial 0.9100 0.9074 2.6949 0.2378 2.4828 0.2349

rbf 0.8784 0.8759 3.1383 0.2769 3.3428 0.2762
erbf 0.8291 0.7707 4.1309 0.3645 4.7208 0.3901
spline 0.9250 0.8789 2.0835 0.1838 3.2455 0.2682

International Journal of Ocean and Climate Systems

N. Harish, Lokesha, S. Mandal, Subba Rao And S.G. Patil 87

5. DISCUSSION
“The performance of SVM depends on the good setting of SVM and kernel parameters. In developing
SVM models, initially parameters are randomly selected by rough search (i.e. for C=100,200,
300….2000; ε = 0.5, 1…2; = 1,2,…6 and γ = 1,2,...6) and d = 1,2,...6) to identify the near optimal
values, and then a fine search (i.e. for C=1,10,20, 30….2000; ε = 0.000001,…2; γ = 0.01,0.02,…6 and
d =1,2....6) is done to identify the final optimal values. But there is non-linearity and complexity
associated in mapping input and output parameters of non-reshaped berm breakwater.
GA is an optimizing tool which is used to search better SVM (C, ε) and kernel (d, γ) parameters to
yield better results compared to SVM model. It is noticed that the performance of the model depends
on the better selection of SVM and kernel parameters. The kernel parameter (d=3) and SVM parameters
(C=137, ε= 0.000137) obtained by GA (Table 3) were used to train and test the data sets. By interfacing
the GA with SVM model, generalization performance of the GA-SVM model shows improvement in
terms of statistical measures such as CC, RMSE, and SI over SVM model as shown in Table 5.
Although GA-SVM with polynomial results in a marginal increase of training CC of 0.9100 with an
scatter index of 0.2378 and root mean square error 2.6949, but gives a considerable increase in testing
CC of 0.9074 with scatter index of 0.2349 and root mean square error of 2.4828 compared to SVM
model with polynomial function. The GA-SVM model with polynomial kernel function shows better
performance than the SVM model with polynomial kernel function in terms of statistical measures”

6. CONCLUSIONS
The performance of GA-SVM appears to be highly influenced by the choice of its kernel function, and
the good setting of kernel and SVM parameters. From the results obtained, it is clearly revealed that
GA-SVM model with polynomial kernel function gives a marginal increase in Training CC of 0.9100
and a noticeable increase in Testing CC of 0.9074 compared to SVM model with polynomial kernel
function. Comparing GA-SVM and SVM model, the GA-SVM model with polynomial kernel function
for testing yields improved results in terms of statistical measures like root mean square (2.4828),
correlation coefficient (0.9074) and scatter index (0.2349). As the input space of the GA-SVM is
smaller and the learning speed of GA-SVM is faster than the individual SVM model, the GA-SVM
model with small input space gives better performance compared to SVM model. Hence, GA-SVM
model is an effective soft computing tool in predicting the damage level of non-reshaped berm
breakwater. Therefore it can be used as an alternative tool to determine the damage level of non-
reshaped berm breakwater.

ACKNOWLEDGEMENT
The authors are grateful to the Director, National Institute of Oceanography, Goa, India; Director, NITK
Surathkal and Head, Department of Applied Mechanics and Hydraulics, NITK Surathkal, India for the
support and encouragement provided to them and for permission to publish.

REFERENCES
Balas C.E., Koc M.L. and Tur R.(2010) ‘’Artificial neural networks based on principal component
analysis, fuzzy systems and fuzzy neural networks for preliminary design of rubble mound
breakwaters’’, Applied Ocean Research, 32, 425 – 433.
Dong B., Cao C. and Lee S.E.(2005) ‘’Applying support vector machines to predict building energy
consumption in tropical region’’, Energy and Buildings, 37(5), 545–553.
Goldberg D. E.(1989) ‘’Genetic Algorithms in Search Optimization and Machine Learning’’,
Addison–Wesley, New York.
Holland, J.H.(1975) “Adaptation in natural and artificial systems”, University of Michigan press, ANN
Arbor, Michigan.
Kazperkiewiecz J., Raez J. and Dubrawski A.(1995) ‘’HPC Strength prediction using artificial neural
networks’’, ASCE J. Computing in Civil Engineering, 9(4),279-284.

Volume 5 · Number 2 · 2014

88 Parameter Optimization using GA in SVM to Predict Damage Level
of Non-Reshaped Berm Breakwater

Kim D. H. and Park W. S.(2005) ‘’Neural network for design and reliability analysis of rubble mound
breakwaters’’. Ocean Engineering, 32, 1332-1349.
Kim D. K., Kim D. H., Chang S. K., Lee J. J. and Lee D. H.(2010). “Stability Number Prediction for
Breakwater Armor Blocks Using Support Vector Regression”, KSCE Journal of Civil
Engineering. 15(2), 225-230.
Mahjoobi J. and Mosabbeb E. A., (2009) ‘’Prediction of significant wave height using regressive
support vector machines’’, Ocean Engineering, 36, 339-347.
Math Works.(2011) ‘’MATLAB – The Language of Technical Computing”, URL
http://www.mathworks.com/.
Mase H., Sakamoto M. and Sakai T.(1995) ‘’Neural network stability of rubble mound breakwaters’’,
J. of Waterway, Port, Coastal and Ocean Engineering, ASCE, 121(6), 294-299.
Patil S. G., Mandal S., Hegde A.V., and Alavandar S.(2011) ‘’Neuro-Fuzzy based Approach for Wave
Transmission Prediction of Horizontally Interlaced Multilayer Moored Floating Pipe
Breakwater’’, Ocean Engineering, 38, 186-196.
Priest M. S., Pugh J. W. and Singh R.(1964) ‘’Seaward profile for rubble mound breakwaters’’. Proc.,
9th Int. Conf. on Coastal Engineering, ASCE, 553-559.
Smith W.G., Kobayashi N. and Kaku S.(1992) ‘’Profile changes of rock slopes by irregular waves’’,
Proceedings of 23th International Conference Coast Engineering ASCE, New York, NY, 1992,
1559–1572.
Subba Rao, Pramod C. H., and Rao B.K.(2004) ‘’Stability of berm breakwater with reduced armor
stone weight’’, Ocean Engineering, 31, 1577–1589.
Subba Rao, Subrahmanya K., Balakrishna K. Rao and Chandramohan V. R.(2008) ‘’Stability aspects
of nonreshaped berm breakwaters with reduced armour weight’’, J. Waterway, Port, Coastal, and
Ocean Engineering, 134(2), 81-87.
Tørum A., Franziska K. and Andreas Menze.(2003) ‘’On berm breakwaters stability, scour,
overtopping’’, Coastal Engineering, 49, 209-238.
Van der Meer J. W.(1988) ‘’Deterministic and probabilistic design of breakwater armor layer’’, J.
Waterways, Ports, Coastal and Ocean Engineering, 114, 66-80.
Van der Meer J. W.(1992) ‘’Stability of the seaward slope of berm breakwaters’’, Coastal Engineering,
16, 205-234.
Vapnik V.(1995) ‘’The Nature of Statistical Learning Theory’’, Springer Verlag, New York.
Vapnik V., Golowich S. and Smola A.(1996) “Support vector method for function approximation
regression estimation and signal processing”, Adv. in Neural Inform. Proces.Syst, 9, 281-287.
Voga G.P. and Belchior J.C.(2006) ‘’An approach for interpreting thermogravimetric profiles using
artificial intelligence’’, J.ThermochimicaActa, 452(2), 140–148.
Yagci O., Mercan D.E., Cigizoglu H.K. and Kabdasli M.S.(2005) ‘’Artificial intelligence methods in
breakwater damage ratio estimation’’, Ocean Engineering, 32, 2088-2106.

International Journal of Ocean and Climate Systems