Neural Network For Stability Analysis

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

NEURAL NETWORK FOR STABILITY ANALYSIS

OF RUBBLE-MoUND BREAKWATERS

By Hajime Mase,l Member, ASCE, Masanobu Sakamoto, 2 and Tetsuo Sakai, 3 Member, ASCE

ABSTRACT: This paper examines the applicability of a neural network to analyze model test data of the
stability of rubble-mound breakwaters. The neural network is an information-processing system, modeled on
the structure of the human brain, that is able to deal with information whose interrelation is not clear. Seven
parameters concerning the stability of rock slopes are used: the stability number, the damage level, the number
of attacking waves, the surf-similarity parameter, the permeability parameter, the dimensionless water depth
in front of the structure, and the spectral shape parameter. The damage levels predicted by the neural network,
calibrated by using a part of Van der Meer's 1988 experimental data, agree satisfactorily well with the measured
No other uses without permission. Copyright (c) 2012. American Society of Civil Engineers. All rights reserved.

damage levels of another part of the data source by Van der Meer 1988 and by Smith et al.'s 1992 data. The
Downloaded from ascelibrary.org by National Institute Technology on 07/26/12. For personal use only.

agreement between the predicted stability numbers by the neural network and the measured stability numbers
is also good.

INTRODUCTION parameters of the neural network are determined to minimize


the error between the output from the network and the right
Studies of stability of rubble-mound breakwaters (or rock answer (measured value) step by step. This procedure is called
slopes) under wave attack have been carried out for the safety
"learning." Programming in a digital computer corresponds
of design. The present design practices mainly depend on
to the learning by the neural network.
J. Waterway, Port, Coastal, Ocean Eng. 1995.121:294-299.

hydraulic model tests, and empirical formulas have been pro-


posed based on the experimental results. The formula pro-
ROCK-SLOPE STABILITY AND STABILITY FORMULA
posed by Hudson (1958) is used for estimating the required
weight of an individual armor unit in the cover layer of rubble Van der Meer (1988) performed a large number of tests
structures. The formula is applicable to regular waves; how- on the stability of rock slopes by changing the slope angle,
ever, the effects of wave period, permeability of the structure, the permeability of the breakwater, the stone size distribu-
and randomness of waves are neglected. Since sea waves are tion, the mass density of stones, the incident-wave spectral
random, it is necessary to study the behavior of rubble-mound shape, the significant wave height, and the period. In this
breakwaters under irregular-wave attack. paper, we use the experimental data of small-scale tests on
Van der Meer (1988) proposed a rock-slope stability for- static stability of rock slopes and large-scale tests on scale
mula based on a large number of irregular-wave experiments. effects by Van der Meer (1988) (listed in Appendix I of his
Kaku (1990) and Kaku et al. (1991) proposed an empirical report) except for the data of a low-crested structure. Static
formula to improve the degree of agreement between the stability means that the damaged slope profile is not sub-
measured damage levels by Van der Meer (1988) and pre- stantially large compared to the initial profile. In the data set
dicted damage levels, beyond the applicable range of static listed, there are data out of applicable range of static stability.
stability. Smith et al. (1992) performed hydraulic model tests Van der Meer (1988) showed that dominant parameters
on the stability of rock slopes by irregular waves, and com- concerning the stability of rubble-mound breakwaters are the
pared the observations of damage level with the predictions significant wave height H,; the mean wave period obtained
by Kaku et al.'s (1991) empirical stability formula; however, from moments of wave spectrum T",; the relative density of
the predictions did not agree well with the observations. In stone 6. = p,lPw - 1; the nominal diameter of stone D n ,,,;
order to get good agreement, the coefficients in the formula the permeability parameter of breakwater P; the number of
had to be changed. It is desirable to make an effective meth- waves acting on the breakwater N; and the slope angle of the
odology to represent or arrange the experimental data. breakwater 6. The damage level, S, was introduced as a pa-
This paper examines the applicability of a neural network rameter that indicates the degree of damage:
to assess the stability of rubble-mound breakwaters. A neural
network is modeled on the structure of the brain, which is S = A/D~5" (I)
said to be effective in dealing with information not having where A = eroded area of the cross section of the breakwater.
clear relations between causes and effects. For example, when The stability formula proposed by Van der Meer (1988) is
we classify data by a digital computer, we have to consider as follows:
all of the patterns. However, if we do not know the relations
between causes and effects beforehand, it is impossible to
N, = s )".2JI \.IE; ~m <
[6.2P"lH ( VN ~, (2)
program the classification. A neural network consists of a
collection of processing elements connected by passways. The
'Assoc. Prof., Dept. of Civ. Engrg., Kyoto Univ., Sakyo-ku, Kyoto N, = 1.0p- o n ( VN
S )".2 Vcot 6 ~;,,; ~'" 2: ~, (3)
606. Japan.
'EngL. Shimizu Corp., Sibaura 1-2-3, Minato-ku, Tokyo 105-07, Ja- where N s = H.J6.D nso = stability number; and ~m
pan. tan 6/V2'TrHs /gT;" is the surf-similarity parameter. The for-
'Prof., Dept. of Civ. Engrg., Kyoto Univ., Sakyo-ku, Kyoto 606,
Japan. mula is for the static condition of S = 1 - 20. The critical
Note. Discussion open until May I, 1996. To extend the closing date value ~c in (2) and (3) is expressed by
one month, a written request must be filed with the ASCE Manager of
Journals. The manuscript for this paper was submitted for review and (4)
possible publication on March 7, 1994. This paper is part of the Journal
of Waterway, Port, Coastal, and Ocean Engineering, Vol. 121, No.6, The stability number N, is calculated for given values of ~m'
NovemberlDecember, 1995. ©ASCE, ISSN 0733-950X/95/0006-0294- S, N, P, and cot 6. The comparison between the measured
0299/$2.00 + $.25 per page. Paper No. 7963. N s (horizontal axis) and the predicted N, (vertical axis) is

294/ JOURNAL OF WATERWAY, PORT, COASTAL, AND OCEAN ENGINEERING / NOVEMBER/DECEMBER 1995
5

4
Stability Number
Van dar Meer's Formula
C.C.=O.92
50

40
Damage Level
Kaku et at.'s Formula
C.C.=O.70
L
o Yen. .om
z· (f) • Smhh ..... Oata
3 't:l 30
't:l G>
~ u
i 2 i 20
0..
0..

10 .
van der Meer's Data
0
0 2 3 4 5 10 20 30 40 50
Measured N. MeasuredS

FIG. 1. Comparison between Measured Stability Numbers and FIG. 3. Comparison between Measured Damage Levels of Van der
No other uses without permission. Copyright (c) 2012. American Society of Civil Engineers. All rights reserved.

those Predicted by Van der Meer's Formula Meer and those Predicted by Kaku et al.'s Formula
Downloaded from ascelibrary.org by National Institute Technology on 07/26/12. For personal use only.

50 XI
Damage Level o Van der
Van der Meer's Formula Meeo's
Dsts
0 C.C.=O.70 • SmIIh X2
40 0
00 Mal.'.
Dsts
00
0 0

..
30 y=f(P)
Xi
J. Waterway, Port, Coastal, Ocean Eng. 1995.121:294-299.

10

o FIG. 4. Model of Neuron


o 10 20 30 40 50
MeasuredS der Meer's data only. The correlation coefficients between
FIG. 2. Comparison between Measured Damage Levels and those the predictions and the observations of Van der Meer are the
Predicted by Van der Meer's Formula same as 0.7 both in Figs. 2 and 3. Hence no remarkable
improvement can be seen in Kaku et al.'s formula.
shown in Fig. 1, in which the correlation coefficient (cq is Smith et al. (1992) performed stability tests of the rubble-
0.92. We can see from the figure that both agree well. mound breakwater and checked the validity of the stability
Eqs. (2) and (3) can be rewritten as follows: formulae of Van der Meer (1988) and Kaku et al. (1991). It
should be noted that the data set by Smith et al. (1992) con-
_(Ns~)S ~m < l
S = yiN 6.2p018 ; (5)
tains the data outside the range of applicability of static sta-
bility such as S > 20. The comparison between the measured
damage levels of Smith et al. (1992) and the predicted ones
Npol3 )S by Van der Meer's formula is shown in Fig. 2 by solid dia-
S = vN ( ~;"Vcot e ; ~m 2:: ~c (6)
monds, for which the permeability coefficient of P = 0.4 is
employed. The predictions are considerably smaller than the
The damage level S is calculated for given values of ~m, N" observations, due to the inclusion of the data outside of ap-
N, P and cot 6. It is important to estimate the damage level plicable range.
when N s = H,//1D"so as well as other parameters are given The comparison between the measured damage levels of
or to predict whether the damage is within an allowable level Smith et al. (1992) and the predicted ones by Kaku et al. 's
if H, exceeds the design wave height. Fig. 2 shows the com- formula (1991) is shown in Fig. 3 by solid diamonds. The
parison between the observations and the predictions for the predictions are also smaller than the observations. If the orig-
damage level S; the open circle denotes Van der Meer's data inal coefficients in Kaku et al.'s formula are changed, the
and the solid diamond denotes Smith et al.'s data explained agreement becomes good [see Fig. 6(b) of Smith et al. (1992)].
later. The CC in the figure is for Van der Meer's data only. In turn, it means that the empirical formula is not universal.
It can be seen from the figure that the agreement for S < 20
is fairly good, but not for the data out of the range of appli- INFORMATION PROCESSING BY NEURAL NETWORK
cability of S > 20. This is logical as (5) and (6) were not based
on data in this range. Outline of Neural Network
Kaku (1990) and Kaku et al. (1991) proposed a stability A neural network is an information-processing system
formula. The formula was also introduced in the paper of modeled on the structure of the human brain (Nakano et al.
Smith et al. (1992). The formula assumes that the damage 1989; Ichikawa 1993; Nielsen 1988). Its biggest merit is its
level approaches the equilibrium value, Seo expressed as ability to deal with fuzzy information whose interrelation is
(7) ambiguous or whose functional relation is not clear.
An artificial system of human neurons is called a "neuro,"
The shape of this formula was also given in Van der Meer and a model of a neuron is called a "unit." A model unit,
(1988). The parameters of Se and K were determined sepa- shown in Fig. 4, usually has many input signals and only one
rately for 665 data, including the data of S > 20 [see Kaku output signal. The unit is connected by passways correspond-
(1990), Kaku et al. (1991), or Smith et al. (1992)]. ing to synaptic connections, and signals are transmitted in one
Fig. 3 shows the comparison between the observations of direction. Input signals, Xi' from the incoming connections
Van der Meer (1988), denoted by open circles, and the pre- are summed up by being multiplied by the weights of the
dictions by the empirical formula of Kaku et al. (1991) for connections, Wi> and then the threshold value, 'Yk, is sub-
the damage level. The CC in the figure is calculated for Van tracted from the summed signal. The modified signal is fed
JOURNAL OF WATERWAY, PORT, COASTAL, AND OCEAN ENGINEERING / NOVEMBER/DECEMBER 1995/295
into the unit. The unit gives an output signal, y, through a weights and thresholds and systematically changing them. A
nonlinear transfer function,f(x). The output becomes an in- mathematical approach to minimize E p is a gradient-descent
put to other units. A common transfer function is the sigmoid procedure, which changes the weights in proportion to aEp /
function, expressed by avkj . The differential of (11) with respect to V kj is

f(x) = 1/{I + exp ( - ~)} = ~ {I + tanh (~,)} (8) aEp


aV
kj
aE" aO k aSk
= aO k aSk aVkj = - (Tk -
,
Ok)f (Sk)H j (12)

where Un = a constant. The sigmoid function has a charac- Eq. (12) can be rewritten by using (9) as
teristic that its differential is expressed by the sigmoid function
as follows: aEp 2
- = - - EkOk(l - Ok)Hj (13)
aVkj Uo
rex) = (2/u o)f(x){1 - f(x)} (9)
k
where E = T k - Ok' Therefore, the weights are changed by
In this study we adopt a three-layered network which con-
No other uses without permission. Copyright (c) 2012. American Society of Civil Engineers. All rights reserved.

sists of an input layer, a hidden layer, and an output layer, aE" 2


8Vkj = -0: -V = - O:E k O k (l - Ok)Hj (14)
shown in Fig. 5. A set of data of an input layer is called an a kj U o
Downloaded from ascelibrary.org by National Institute Technology on 07/26/12. For personal use only.

"input pattern." The data are transmitted to a hidden layer,


and transformed in the hidden units. The outputs from the where 0: = a proportional coefficient. Introducing the error,
Eb defined by
hidden units are transmitted to the units of an output layer,
and transformed. The outputs from the output units are the
aE" aE" aO 2
predicted values. The network is supplied not only with input Ek = - - = - - - k = - EkO k(l - Ok) (15)
patterns but also with teach signals. To understand the neural- aSk aO aSk U k o
network method applied to the stability analysis of rubble- we can rewrite (14) as follows:
mound breakwater easily, let's consider the case of deter-
8Vkj = O:EkHj
J. Waterway, Port, Coastal, Ocean Eng. 1995.121:294-299.

mining N s (one output unit) from S, N, P, and ~m (four input (16)


units), see Fig. 5. A set of S, N, P, and ~m is an input pattern, Modification for the threshold is carried out as
the measured value of N, is a teach signal, and the output of
N s is a predicted value. 8'Yk = I3E k (17)
The weights and thresholds are modified step by step to
minimize the mean square error between the measured values where 13 = a proportional coefficient. Using (14) and (17),
(teach signals) of N s and the predicted values (outputs of the V kj and 'Yk are renewed as follows:
output unit) of N s . This procedure is called "learning." (18a,b)
Achievements by learning are kept in the weights of the pass-
ways and in the thresholds of the units. Back-propagation is Concerning modification of the weights between input units
one of the self-learning methods to give desired answers (Nak- and hidden units, Wji, and the thresholds of the hidden units,
ano et al. 1989; Ichikawa 1993; Nielsen 1988). aj, we also employ the gradient descent procedure. Let I; be
the output from an input unit #i. The output from a hidden
Learning Procedure unit, Hj , is expressed by

Let Hj be the output from a hidden unit #j, let Ok be the Hj = f(V); vj = 2: wj;'!; - aj (19)
output from an output unit #k, let Vkj be the weight between
the hidden unit #j and the output unit #k, and let 'Yk be the
threshold of the output unit #k (see Fig. 5); the relations
among these valuables are given by
Ok = f(Sk); Sk = 2: V kj ' H j -
j
'Yk (lOa,b)

The sum of the square errors, E p , between the output Ok and


the teach signal T k is described by
E" = (1/2) 2: (Tk -
k
Ok)2 (11)

The E" can be minimized by starting with an arbitrary set of


we can rewrite (20) as follows:

-
aE" = - 111 (22)
N Measured aWji J'
S Value
t
Therefore, the weights W j ; and the thresholds aj are changed
: Minimizing
by
o of Predicted
Ns Value (23, 24)

Since the error is fed backed to lower layers, the learning


procedure is called back-propagation. There are several var-
iations of back-propagation. In this study we employ a mod-
ified momentum method, which requires less time for learn-
ing. First, we calculate the modification values of the weights
and the thresholds, by (16), (17), (23), and (24), for each
296/ JOURNAL OF WATERWAY, PORT, COASTAL, AND OCEAN ENGINEERING / NOVEMBER/DECEMBER 1995
40 Damage Level
NlMnberofLeaming
N",.,=5OOO
30 C.C.=O.94
(J)

al
20

~ 10

Van de< Meer's 100 Data


0
0 10 20 30 40
(a) Measured S

40
No other uses without permission. Copyright (c) 2012. American Society of Civil Engineers. All rights reserved.

Damage Level
Num~ng
Downloaded from ascelibrary.org by National Institute Technology on 07/26/12. For personal use only.

30 C.C.-o.95
(J)

al 20
i
I l.
10
FIG. 6. Flowchart of Calculation Van de< Maefs 100 Dala
0
0 10 20 30 40
J. Waterway, Port, Coastal, Ocean Eng. 1995.121:294-299.

input pattern. Each modification value is summed up for all (6) Measured S
sets of input patterns, respectively. Finally, each modification
value, say OWi, is calculated by the moment average of the FIG. 7. Comparison between Measured Damage Levels and Pre-
present time-step value, oW, and the previous time step value, dicted Ones by Neural Network against Randomly selected 100
OWi-l, as Experimental Data of Van der Meer: (a) after 5,000 Learning Itera-
tions; (b) after 50,000 Learning Iterations
OWi = (1 - m)8W + m8W- 1 (25)
1.1
where m = a momentum coefficient increasing with time Damage Level
c
... •._--_
step. Fig. 6 shows a flowchart of calculation. Van der Maar's 100 Data
.~ 1.0
The initial value of m is set to 0.6, and increased by 0.02
in one step, up to 0.9. The coefficients, a and ~ in (16), (17), ==
_ ..
(23), and (24) are set to 0.1. The initial values of the weights
and the thresholds are given by uniform random numbers
80.9
c:
o
.--..-~::::==~~==.o-=.-.-.-.-.-.- . --.-:=~
from - 1 to 1. Inputs signals are normalized to be from 0 ~
to 1. ~ 0.8 -0- Ntr 4
Ntr12
It has been found from the existing studies (Ichikawa 1993) .-0-.

that: (1) The required number of units depends on an amount


8 ..•. Ntr20

of input and teach signals; (2) a large number of units is not O.7 .....--r---r~--r-1['""T""'f.,.,...---.---.,,...-,,........
necessarily required; (3) there can be a phenomenon called 1000 10000
"overlearning" so as to also learn the subtle structure of a Number of Learning
given data set.
FIG. 8. Correlation Coefficient between Measured and Predicted
Damage Levels against Randomly Selected 100 Experimental Data
STABILITY ASSESSMENT BY NEURAL NETWORK of Van der Meer
As a preliminary examination, we check the characteristics
of the neural network. The number of hidden units is changed and that learning more than 5,000 iterations does not improve
by 4, 8, 12, 16 and 20; the number of learning iterations is the CC so much. When the number of hidden units, N H' is
changed by 1,000,5,000, 10,000,30,000, and 50,000. 8, which is double the number of input units, the CC becomes
We employ four input units corresponding to the perme- larger than that in the case of N H = 12; the reason is not
ability coefficient P, the number of waves N, the stability clear. From the simplicity of the network, the small number
number H,It::..D n50 and the surf-similarity parameter ~m to pre- of N H is preferable.
dict the damage level S. We adopt 100 experimental data set We tried to construct the network by using 579 experi-
(input patterns) randomly selected from the experimental data mental data set of Van der Meer (1988). But it was impossible
set of Van der Meer (1988). The results of the comparisons to construct a proper network. Then we constructed the net-
between the measured and predicted damage levels are shown work by using the randomly selected 100 experimental data
in Fig. 7, where the number of hidden units is 16, and the set and applied the network to 579 experimental data set.
number of learning iterations is 5,000 [in Fig. 7(a)] and 50,000 Fig. 9 shows the Cc. It can be seen from the figure that the
[in Fig. 7(b)]. Both figures indicate that the agreement be- CC becomes smaller when the number of learning iterations
tween the observations and the predictions is good and that exceeds 5,000; this tendency is opposite to that seen in Fig.
the larger number of learning iterations gives better agree- 8. This phenomenon is attributed to overlearning; that is, the
ment. network is so fit to the structure of the 100 input patterns
The value of the CC, when changed the numbers of hidden with increase in the number of learning iterations that the
units and learning iterations, is shown in Fig. 8. It is seen network does not fit well to the other input patterns. Al-
from the figure that the CC is larger than 0.92 when the though Fig. 9 shows that the CC is largest in the case of four
number of learning iterations is 5,000 and N H is 8, 16, 20, hidden units, it is necessary to choose the number of hidden
JOURNAL OF WATERWAY, PORT, COASTAL, AND OCEAN ENGINEERING / NOVEMBER/DECEMBER 1995/297
0.9 Damage level
50
Damage level
"E Van dar Maar's 579 Data
40
Neural NetworI<
.~ 0.8 C.C.=O.81
IE p-_~::=::::~. rn
,~ ':;•.•.•.•. ~ ~.. 0 30
....,
c: 0.7 -0- NtF 4
,
"
::::::---...
•.•.•.•.•.•.• :a::-.c.
•.••• '0
~
o -6- Nu= 8 "c. .......• !2O
1il
- ··0·· n
NtF12 ••••• _ /
0 0..

~ 0.6 -0- NtF16 -'0/ 10


8 ..•. NtF20 Van de< Meef's Data
o
0.5 'r---r---y---r-r-r..,...,..,...----y--.,..-...,........ o 10 20 30 40 50
1000 Measured S
FIG. 10. Comparison between Measured Damage Levels of Van
No other uses without permission. Copyright (c) 2012. American Society of Civil Engineers. All rights reserved.

der Meer and Predicted Ones by Neural Network


FIG. 9. Correlation Coefficient between Measured and Predicted
Damage Levels against 579 Experimental Data of Van der Meer
Downloaded from ascelibrary.org by National Institute Technology on 07/26/12. For personal use only.

40 Damage level
N8II'al Networ1c
units in order to get good agreement at least against the input C.C.=O.82
30
patterns used to construct the network. rn 00 00 00
From the foregoing preliminary examination it was con- 0

" oil"

~
cluded that the 100 experimental data set is enough to con- 20
II
~"'o
struct a network and the number of learning iterations 5,000 0
0
0.. 0
is sufficient. With these conditions we compared the mea- 10
sured damage levels with the predicted ones; however, the
J. Waterway, Port, Coastal, Ocean Eng. 1995.121:294-299.

degree of agreement is the same as the prediction by Kaku smith eI al.'s Data
et al. 's (1991) formula, shown in Fig. 3. Therefore, we should o"1T----,.--.-----r--"""T
consider other possibilities in order to improve the agree- o 10 20 30 40
Measured S
ment.
Until now, we employed four input units. The dimension- FIG. 11. Comparison between Measured Damage Levels of Smith
et al. and those Predicted by Neural Network
less water depth, hlHso was neglected as the input parameter.
It is sure that the significant wave height in front of the break-
5
water includes some effect of the water depth. However, even Stability Number
if the significant wave heights are the same, other represent- Neural Network
ative wave heights or wave-height distributions are affected 4 C.C.=O.91
by the water depth. In other words, even if the significant
wave heights are the same, the damage level of the break- Z
.
3
water would change according to the occurrence probability '0

of larger waves. Mase and Kobayashi (1990) showed that the ~


shape of wave-height distribution varies at the location or the ~2
water depth. Therefore, we had better adopt the dimension- c..
less water depth in front of the breakwater as a parameter
o Van der Mws Data
for the neural network. Van der Meer (1988) took into ac- • Smith eI al.'s Deta
count the effect of a non-Rayleigh distribution by using
H2':fIHs '
Another parameter is the wave groupiness. Van der Meer o 2 3 4 5
Measured Ns
(1988) mentioned that wave grouping does not have signifi-
cant effects on the stability of rock slopes. However, Ryu FIG. 12. Comparison between Predicted Stability Numbers by
and Sawaragi (1986) and Medina et al. (1990) pointed out Neural Network and Measured Ones of Van der Meer and Smith
that the effect of wave grouping on the stability of rubble et al.
structures should be taken into account. The wave grouping
roughly depends on the spectral shape. Van der Meer (1988) der Meer's data (1988), including the data of S > 20, and the
employed the mean period obtained from the spectral mo- predictions, where the correlation coefficient is 0.81. It can
ments in order to include the effect of the spectral shape be seen from this figure that the agreement between the mea-
indirectly. The neural network can deal with qualitative data sured and predicted damage levels is improved compared to
by assigning the values to them. Therefore, in this study, the those in Figs. 2 and 3, specially for large S.
spectral shapes of narrowband, medium-band, and wideband Fig. 11 shows the comparison between the measured dam-
are assigned to 1.0,0.5, and 0, respectively. Of course, if a age levels of Smith et al. (1992) and the predicted ones by
parameter representing spectral shape such as the spectral the aforementioned network, constructed by the Van der
peakedness parameter is given, we can use it. Meer's data, where P = 0.5 is used. The correlation coeffi-
As a summary, the dimensionless water depth in front of cient is 0.82. Fig. 11 indicates that the predictions by the
the breakwater, hlHso and the spectral shape parameter are network become better than those in Figs. 2 and 3 indepen-
added as the parameters hereafter. dent on data source, which means the neural network is useful
We construct a network with 12 hidden units (double the in arranging the stability data.
number of input units), by using the randomly selected 100 In the case of predicting the stability number N" the net-
experimental data set of Van der Meer (1988) and by using work with six input units (P, N, S, ~m, hlHso and the spectral
5,000 learning iterations. Employing the network, we predict parameter) is employed. The network is constructed by the
the damage level and compare it with the measured one. Fig. randomly selected 100 experimental data set of Van der Meer
10 shows the comparison between the observations of Van (1988) and by 5,000 learning iterations. Fig. 12 shows the
298/ JOURNAL OF WATERWAY, PORT, COASTAL, AND OCEAN ENGINEERING / NOVEMBER/DECEMBER 1995
comparison between the predicted stability numbers by the ing waves and its empirical numerical model considering surf beat."
network and the measured ones of Van der Meer (1988) and Proc., Coast. Sediments '91, ASCE, New York, N.Y., 688-702.
Medina, J. R., Fassardi, C., and Hudspeth, R. T. (1990). "Effects of
Smith et al. (1992), where the correlation coefficient is 0.91. wave groups on the stability of rubble mound breakwaters." Proc.,
We can see from the figure that the agreement of the stability 22nd Int. Conf. on Coast. Engrg., ASCE, New York, N.Y., 1552-
number is much better than that against the damage level. 1563.
Concerning N" the agreements between the predictions and Nakano, K., Inuma, K., and Kiritani, W. (1989). "Neurocomputer."
observations by Van der Meer's formula and by the neural Gijutu-Hyoronsha Co. Ltd. (in Japanese).
network are comparable. From (2) and (3), since the N s is a Nielsen, R. H. (1988). "Neural computing: picking the human brain."
IEEE Spectrum, 25(3), 36-41.
function of S to the 0.2 power, the variation of S becomes Ryu, C. R., and Sawaragi, T. (1986). "A new design method of rubble
small in the predicted values of N,. mound structures." Proc., 20th Int. Conf. on Coast. Engrg., ASCE,
In this paper, the dimensionless parameters basically pointed New York, N.Y., 2188-2202.
out by Van der Meer (1988) were employed. We can also use Smith, W. G., Kobayashi, N., and Kaku, S. (1992). "Profile changes of
raw parameters such as H" Tm , P" Pw' and so on. When there rock slopes by irregular waves." Proc., 23rd Int. Conf. on Coast.
are many parameters, we require larger number of input units Engrg., ASCE, New York, N.Y., 1559-1572.
No other uses without permission. Copyright (c) 2012. American Society of Civil Engineers. All rights reserved.

Van der Meer, J. W. (1988). "Rock slopes and gravel beaches under
and longer central processing unit (CPU) time. Therefore, wave attack," PhD thesis, Delft Univ. of Techno!., Delft, The Neth-
the dimensionless parameters, estimated by a physical con-
Downloaded from ascelibrary.org by National Institute Technology on 07/26/12. For personal use only.

erlands.
sideration or by a dimensional analysis, are advantageous.
Using the determined dimensionless parameters, the neural APPENDIX II. NOTATION
network can be calibrated. After that, the parameters could
be reduced one by one, by checking the error between the The following symbols are used in this paper:
observations and predictions.
A eroded area of breakwater;
SUMMARY AND CONCLUSIONS CC correlation coefficient between observations and pre-
dictions;
A neural network, which is modeled on the structure of D n50 nominal diameter of stone;
J. Waterway, Port, Coastal, Ocean Eng. 1995.121:294-299.

the human brain, is effective in pattern-recognition problems. Ep sum of square error defined by (11);
A neural network consists of a collection of units connected g acceleration of gravity;
by passways. The neural network changes the strengths of H 2% 2% exceedance large-wave height;
connections so as to give the right answer through learning. Hj output from hidden unit;
This paper examines the applicability of a neural network Hs significant wave height;
in assessing the stability of rubble-mound breakwaters. For h/Hs dimensionless water depth in front of breakwater;
parameters concerning stability of rock slopes we use the Ii output from input unit;
seven parameters of stability number, damage level, number K coefficient in Kaku et al. 's (1991) stability formula;
m momentum coefficient in (25);
of attacking waves, surf-similarity parameter, permeability
N number of waves acting on breakwater;
parameter, the dimensionless water depth in front of the number of hidden units;
N/f
structure, and the spectral shape parameter. The neural net- Ns stability number;
work constructed here was three-layered one. The damage Ok output from output unit;
levels predicted by the neural network, calibrated by using a P permeability parameter of breakwater;
part of Van der Meer's experimental data (1988), agree sat- S damage level;
isfactorily well with the measured ones of different data sources Se coefficient in Kaku et al.'s (1991) stability formula;
by Van der Meer (1988) and Smith et al. (1992). The agree- Sk input into output unit;
ment between the predicted stability numbers by the neural Tk teach signal;
network and the measured ones is also good, but not better Tm mean wave period;
than the stability formula itself. Vj input into hidden unit;
uo constant in sigmoid function;
ACKNOWLEDGMENTS Vkj weight between hidden unit and output unit;
W ji weight between input unit and hidden unit;
This study was supported by the Grant-in-Aid for Developmental
Scientific Research (No. 06555148), The Ministry of Education, Science,
<X proportional coefficient of increment of weight;
Sports and Culture. Japan. f3 proportional coefficient of increment of threshold;
'Yk threshold;
A relative density of stone;
APPENDIX I. REFERENCES
Aj error defined by (21);
Hudson, R. Y. (1958). "Design of quarry stone cover layer for rubble oW increment of weight at present step;
mound breakwaters." Res. Rep. No. 2-2, Waterways Experiment Sta- OWi-! increment of weight at previous step;
tion, Coast. Engrg. Res. Ctr., Vicksburg, Miss. oW' increment of weight at next step;
Ichikawa, H. (1993). Layered neural network. Kyoritu-Shuppan Co., E
k
difference between teach signal and output value;
Ltd., 184 (in Japanese). Ek error defined by (15);
Kaku, S. (1990). "Hydraulic stability of rock slopes under irregular wave
attack," master thesis, Univ. of Delaware, Newark, De!.
e angle of slope of breakwater;
Kaku, S., Kobayashi, N., and Ryu, C. R. (1991). "Design formulas for e j threshold of hidden unit;
hydraulic stability of rock slopes under irregular wave attack." Proc., ~c critical surf similarity parameter;
38th Japanese Conf. on Coast. Engrg., Japan Soc. of Civ. Engrs. (JSCE), surf similarity parameter;
Tokyo, Japan, 661-665 (in Japanese). Ps density of stone; and
Mase, H., and Kobayashi, N. (1991). "Transformation ofrandom break- Pw density of water.

JOURNAL OF WATERWAY, PORT, COASTAL, AND OCEAN ENGINEERING / NOVEMBER/DECEMBER 1995/299

You might also like