nhess-10-2527-2010 | Regression Analysis | Mathematics

Nat. Hazards Earth Syst. Sci.

, 10, 2527–2537, 2010
www.nat-hazards-earth-syst-sci.net/10/2527/2010/
doi:10.5194/nhess-10-2527-2010
© Author(s) 2010. CC Attribution 3.0 License.
Natural Hazards
and Earth
System Sciences
An attempt to model the relationship between MMI attenuation and
engineering ground-motion parameters using artificial neural
networks and genetic algorithms
G-A. Tselentis
1
and L. Vladutu
2
1
Seismological Laboratory, University of Patras, RIO 265 04, Patras, Greece
2
Mathematics Dept., Dublin City University, Dublin, Republic of Ireland
Received: 23 July 2010 – Accepted: 13 October 2010 – Published: 7 December 2010
Abstract. Complex application domains involve difficult
pattern classification problems. This paper introduces
a model of MMI attenuation and its dependence on
engineering ground motion parameters based on artificial
neural networks (ANNs) and genetic algorithms (GAs). The
ultimate goal of this investigation is to evaluate the target-
region applicability of ground-motion attenuation relations
developed for a host region based on training an ANN using
the seismic patterns of the host region. This ANN learning
is based on supervised learning using existing data from
past earthquakes. The combination of these two learning
procedures (that is, GA and ANN) allows us to introduce
a new method for pattern recognition in the context of
seismological applications. The performance of this new
GA-ANN regression method has been evaluated using a
Greek seismological database with satisfactory results.
1 Introduction
A common problem encountered in engineering seismology
involves the difficulty in assessing the damage potential of
an earthquake based on the distribution of macroseismic
intensity. This parameter is subject to interpretation due
to the wide variation in geological conditions, the response
of structures, uncertainty related to construction conditions
before the earthquake, the type of construction, and
population density. Obviously, a physically-based ground-
related MMI is required for engineering purposes (Tselentis
Correspondence to: G-A. Tselentis
(tselenti@upatras.gr)
and Danciu, 2008; Danciu and Tselentis, 2007). With the
advent of instrumental seismology, the relationship between
intensity and ground-motion parameters has become a topic
of increasing interest.
Instrumental seismology offers the possibility to transform
readily observed data (in this case, intensity) into widely-
used parameters useful for engineering purposes (in this
case, engineering ground-motion measures) and allows
seismologists to evaluate historical earthquakes for which
no instrumental data are available in order to assess seismic
hazard and damages, correlate different intensity scales, and
rapidly assess the severity of ground shaking.
Until recently, macroseismic intensity was related most
frequently to peak ground acceleration (PGA) because of
that parameter’s importance for seismic-resistant design.
This is due to the fact that the product of PGA and mass
represents the inertial force loading structures (Krinitzsky
and Chang, 1988). In recent years, research on earthquake
damage prediction has concluded that other ground-motion
characteristics such as duration, frequency, and energy
content all contribute to structural damage.
Focusing either on regional or worldwide data, many
empirical equations have been proposed to relate a seismic
event’s felt intensity with its peak ground velocity (PGV)
(Panza et al., 1997; Wald et al., 1999; Wu et al., 2003;
Kaka and Atkinson, 2004, 2007; Atkinson and Kaka, 2007),
duration of ground motion (Trifunac and Westermo, 1977),
response spectra (Kaka and Atkinson, 2007), Fourier spectra
(Sokolov, 1998, 2002), cumulative absolute velocity (CAV)
(Cabanas et al., 1997; Kostov, 2005), Arias intensity (I
a
)
(Margottini et al., 1992), Housner’s spectrum intensity, and
JMA instrumental intensity (Karim and Yamazaki, 2002).
Published by Copernicus Publications on behalf of the European Geosciences Union.
2528 G-A. Tselentis and L. Vladutu: MMI attenuation engineering parameters and neural networks
Recently, Tselentis and Danciu (2008) derived empirical
regression equations for modified Mercalli intensity (MMI)
and for various ground-motion parameters such as duration,
CAV, I
a
, characteristic intensity, Housner’s spectrum inten-
sity, and total elastic input energy index.
These relationships have been used to generate maps
of estimated shaking intensities within a few minutes of
the event based on recorded peak motions. These maps
provide a rapid visualisation of the extent of expected
damages following an earthquake and can be used for
emergency response, loss estimation, and the release of
public information through the media.
Another application of engineering ground-motion param-
eters involves the development of early warning systems.
These systems are low-cost solutions aimed at reducing
the seismic risk to vital facilities, such as nuclear power
plants, pipelines, and high-speed trains. The ground-motion
parameters and the damage potential threshold are essential
for these systems.
The objective of the present investigation is to uncover
the hidden, complex and often fuzzy relations between the
engineering ground-motion parameters and macroseismic
intensity and to express these relations in the form of
input/output dependencies. The emergence of neural
network technology (Haykin, 1999; Bishop, 1996) provides
valuable insights for addressing the complicated problem of
expression these relations. In this context, neural networks
can be viewed as advanced mathematical models used to
discover complex correlations between physical process
variables from a set of perturbed observations.
2 Engineering seismological parameters
Because structure and equipment damage is measured
according to its inelastic deformation, the potential for
earthquake damage depends on the time duration of motion,
the energy absorption capacity of the structure or equipment,
the number of strain cycles, and the energy content of the
earthquake. Therefore, for engineering purposes, param-
eters that incorporate in their definition these previously-
mentioned characteristics are more reliable predictors of an
earthquake’s damage potential than parameters related solely
to the amplitude of ground motion (such as peak ground
acceleration, or PGA), which are often poor indicators of
structural damage. The most commonly-used engineering
ground-motion parameters in addition to PGA and PGV are
Arias intensity (I
a
), acceleration response spectrum (S
a
), and
cumulative absolute velocity (CAV).
I
a
, as defined by Arias (1970), is the total energy per unit
weight stored by a set of undamped simple oscillators at the
end of ground motion. The Arias intensity for ground motion
in the x-direction (I
aX
), is calculated as follows:
I
aX
=
2
g
t

0
[a
X
(t )]
2
dt, (1)
where a
X
(t ) is the acceleration time history in the
x direction, and t is the total duration of ground motion.
Spectrum acceleration (S
a
) is the most common-response
spectral parameter and is related to spectrum velocity (S
v
)
and spectrum displacement (S
d
) according to the following
expression.
S
a
=

T
S
v
=ωS
v
=


T

S
d

2
S
d
, (2)
where T is the undamped natural period of a single-degree-
of-freedom (SDOF) oscillator.
Cumulative absolute velocity (CAV) is defined as the
integral of the absolute value of ground acceleration over the
following seismic time-history record.
CAV=
t

0
|a(t )|dt, (3)
where |a(t )| is the absolute value of acceleration, and t is
the total duration of ground motion.
3 Data set
The strong-motion records used for the present investigation
were provided by the European Strong Motion Database
(Ambraseys et al., 2004). More details on these data can
be found in Danciu and Tselentis (2007). The macroseismic
information was available in part from the digital database
of the European strong-motion data; it was also separately
estimated in part from the macroseismic data provided by the
Geodynamic Institute of the National Observatory of Athens
(Kalogeras et al., 2004). The general criterion of selecting
the appropriate MMI value was to allocate at each station
the nearest MMI values within an uncertainty of one unit to
every station. If more than one MMI value was observed near
the station location, at an equal distance from the station, the
average of the values was used (Tselentis and Danciu, 2008).
Maps of the reported MMI values together with the
strong-motion instrument locations were plotted to show
reasonable confidence that the allocated MMI values are
within one unit of the assigned value. This approach provides
rapid visualisation of the macroseismic distribution in the
area surrounding the recording stations and might be an
efficient approach to assigning MMI values to minimise
errors (Atkinson and Kaka, 2007).
At this preliminary stage of intensity ground-motion
investigation, we decided not to disaggregate the soil effect
because the soft soil subset comprises less than 30% of the
total data set.
Nat. Hazards Earth Syst. Sci., 10, 2527–2537, 2010 www.nat-hazards-earth-syst-sci.net/10/2527/2010/
G-A. Tselentis and L. Vladutu: MMI attenuation engineering parameters and neural networks 2529
23



Fig. 1

Fig. 2
Fig. 1. Epicentral distribution of earthquakes used in the present
analysis.
The final data set consists of 310 records from
151 earthquakes and is depicted in Fig. 1 and Table 1. Using
the recorded strong-motion data for these earthquakes, for
each horizontal component, we compute the previously
mentioned engineering seismological parameters. The
arithmetic average between the two horizontal components
of the independent variables was used for the present
investigation.
4 Artificial neural networks
An artificial neural network (ANN) is an information-
processing paradigm inspired by the way biological nervous
systems such as the brain process information. The most
basic element of the human brain involves a specific type
of cell that provides us with the ability to remember, think,
and apply previous experiences to our every action. These
cells are known as neurons. The power of the brain comes
from the number of neurons and the multiple connections (or
synapses) between them.
Figure 2 shows a simplified view of an ANN. It consists
of a network of simple processing elements (that is, artificial
neurons) that are organised in several layers, including an
input layer that shows the number of neurons linked to the
dimensionality of the input, one or several hidden layers, and
an output layer. The hidden layer represents network inputs.
When one presents the network with a form to be learned, the
neurons simultaneously start a state of activity that causes
a small modification of the synaptic forces between them.
This is followed by a quantitative reconfiguration of all of
the synapses: some of them become very strong, while others
23



Fig. 1

Fig. 2
23



Fig. 1

Fig. 2
Fig. 2. General topology of a feed-forward ANN with one hidden
layer.
24




Fig. 3



Fig. 4

Fig. 5
Fig. 3. Two classical nonlinear activation functions.
become weak. The learned form is not directly memorised at
a precise place; it corresponds to a particular energy state
of the network, which is a particular configuration of the
activity of each neuron, across a very large set of possible
configurations. This configuration is supported by the values
of the synaptic forces.
Let Y
s
j
represent the output of the j-th neuron at layer s;
W
s
i j
is the weight connecting the i-th neuron in layer s to the
j-th neuron at layer s-1. The neurons have their activation
function characterised by a nonlinear function, such as the
sigmoid function in Fig. 3. This function maps the output to
its input and can be expressed by the following equation.
Y
s
j
=f

b
+
R
¸
i=1
W
s
j i
· Y
s−1
i

(4)
where b is the bias.
www.nat-hazards-earth-syst-sci.net/10/2527/2010/ Nat. Hazards Earth Syst. Sci., 10, 2527–2537, 2010
2530 G-A. Tselentis and L. Vladutu: MMI attenuation engineering parameters and neural networks
Table 1. Database of strong motion records used.
No Earthquake Date Time Latitude Longitude Moment Focal Number
Name Magnitude Depth of Records
(dd-mm-yy) (

) (

) (km) B C D
1 Kefallinia Island 17/9/1972 14:07:12 38.245 20.263 5.6 1 0 0 2
2 Ionian 11/4/1973 15:52:12 38.78 20.55 5.8 7 1 0 0
3 Ionian 11/4/1973 16:11:36 38.76 20.65 4.9 15 1 0 0
4 Patras 29/1/1974 15:12:43 38.3 21.86 4.7 13 1 0 0
5 Amfissa 29/12/1977 16:52:59 38.55 22.35 5.1 10 1 0 0
6 Volvi 7/4/1978 22:23:28 40.7 23.106 5 6 1 0 0
7 Achaia 18/5/1978 00:18:49 38.3 21.79 4.5 26 1 0 0
8 Volvi 20/6/1978 20:03:22 40.729 23.254 6.2 6 3 1 0
9 Almiros (aftershock) 16/7/1980 00:06:58 39.21 22.76 5 12 0 1 0
10 Almiros (aftershock) 26/9/1980 04:19:21 39.27 22.75 4.8 5 0 1 0
11 Almiros (aftershock) 8/11/1980 09:15:59 39.3 22.83 5.2 5 0 1 0
12 Alkion 24/2/1981 20:53:39 38.099 22.842 6.6 10 2 0 0
13 Alkion 25/2/1981 02:35:53 38.135 23.05 6.3 8 1 0 0
14 Panagoula 25/5/1981 23:04:00 38.71 20.95 4.7 15 1 0 0
15 Levkas 27/5/1981 15:04:02 38.79 21.01 5.1 15 1 0 0
16 Preveza 3/10/1981 15:16:20 39.2 20.8 5.4 10 1 1 0
17 Paliambela 4/10/1981 08:33:32 38.91 21.02 4.7 10 1 0 0
18 Kefallinia Island 17/1/1983 12:41:31 37.96 20.26 6.9 5 1 1 0
19 Kefallinia (aftershock) 17/1/1983 15:53:57 38.13 20.49 5.2 11 0 1 0
20 Kefallinia (aftershock) 31/1/1983 15:27:02 38.12 20.49 5.4 4 0 1 0
21 Kyllini (foreshock) 20/2/1983 05:45:12 37.72 21.25 4.9 15 1 1 0
22 Etolia 16/3/1983 21:19:41 38.81 20.89 5.2 25 1 0 0
23 Kefallinia (aftershock) 23/3/1983 19:04:06 38.78 20.81 5.2 25 1 0 0
24 Kefallinia (aftershock) 23/3/1983 23:51:08 38.22 20.41 6.2 3 1 2 0
25 Off coast of Magion Oros peninsula 8/6/1983 15:43:53 40.08 24.81 6.6 22 0 1 2
26 Ierissos (foreshock) 14/6/1983 04:40:43 40.44 23.92 4.5 10 0 1 0
27 Ierissos 26/8/1983 12:52:09 40.45 23.92 5.1 12 0 1 2
28 Near southeast coast of Zakynthos Island 10/4/1984 10:15:12 37.64 20.85 5 6 0 1 0
29 Gulf of Corinth 17/8/1984 21:22:58 38.21 22.68 4.9 24 1 0 0
30 Arnissa 7/9/1984 18:57:12 40.66 21.89 5.2 5 0 1 1
31 Kranidia 25/10/1984 14:38:30 40.13 21.64 5.5 20 0 0 1
32 Kremidia (aftershock) 25/10/1984 09:49:15 36.93 21.76 5 11 0 0 2
33 Gulf of Amvrakikos 22/3/1985 20:38:39 38.99 21.11 4.5 6 0 0 1
34 Anchialos 30/4/1985 18:14:13 39.24 22.89 5.6 13 0 1 1
35 Gulf of Kiparissiakos 9/7/1985 10:20:51 37.24 21.25 5.4 10 0 0 1
36 Near coast of Preveza 31/8/1985 06:03:47 39 20.61 5.2 15 1 2 0
37 Drama 11/9/1985 23:30:43 41.26 23.98 5.2 18 0 2 2
38 Aghios Vasileios 18/2/1986 05:34:42 40.7 23.23 4.8 3 2 1 1
39 Skydra-Edessa 18/2/1986 14:34:04 40.79 22.07 5.3 10 0 1 1
40 Kalamata 13/9/1986 17:24:34 37.1 22.19 5.9 1 1 2 1
41 Kalamata (aftershock) 15/9/1986 11:41:28 37.03 22.13 4.9 12 0 3 1
42 Tsipiana 2/1/1987 05:35:36 37.86 21.77 4.5 20 0 1 0
43 Near northwest coast of Kefallinia Island 27/2/1987 23:34:52 38.46 20.33 5.7 5 1 2 0
44 Dodecanese 10/5/1987 09:27:02 36.29 28.46 5.3 6 0 1 1
45 Kounina 14/5/1987 06:29:11 38.17 22.06 4.6 9 0 1 0
46 Near northeast coast of Crete 2/9/1987 12:28:23 35.41 26.08 4.9 18 0 1 0
47 Kalamata (aftershock) 6/10/1987 14:50:12 37.24 21.48 5.3 30 1 1 1
48 Near southwest coast of Peloponnes 12/10/1987 22:51:14 36.65 21.68 5.2 18 0 0 1
49 Near northeast coast of Rodos Island 25/10/1987 13:02:00 36.3 28.36 5.1 10 0 0 1
50 Astakos 22/1/1988 06:18:55 38.64 21.02 5.1 10 1 0 0
51 Kefallinia Island 6/2/1988 10:35:25 38.32 20.43 4.8 10 0 1 0
52 Gulf of Corinth 4/3/1988 03:56:07 38.08 22.82 4.5 5 0 1 0
53 Ionian 24/4/1988 10:10:33 38.83 20.56 4.8 1 2 0 0
54 Gulf of Corinth 7/5/1988 20:34:52 38.1 22.86 4.9 10 1 1 0
55 Etolia 18/5/1988 05:17:42 38.35 20.47 5.3 26 0 2 1
56 Etolia 22/5/1988 03:44:15 38.35 20.54 5.4 15 2 2 0
57 Agrinio 3/8/1988 11:38:57 38.82 21.11 4.9 28 0 1 1
Nat. Hazards Earth Syst. Sci., 10, 2527–2537, 2010 www.nat-hazards-earth-syst-sci.net/10/2527/2010/
G-A. Tselentis and L. Vladutu: MMI attenuation engineering parameters and neural networks 2531
Table 1. Continued.
No Earthquake Date Time Latitude Longitude Moment Focal Number
Name Magnitude Depth of Records
(dd-mm-yy) (

) (

) (km) B C D
58 Kyllini (foreshock) 22/9/1988 12:05:39 37.93 21.08 5.3 12 0 2 0
59 Kyllini (foreshock) 30/9/1988 13:02:54 37.69 21.33 4.7 5 1 1 0
60 Kyllini 16/10/1988 12:34:05 37.9 20.96 5.9 4 1 5 0
61 Trilofon 20/10/1988 14:00:59 40.53 22.94 4.8 20 3 1 1
62 Kyllini (aftershock) 22/10/1988 14:58:18 37.88 21.02 4.5 20 1 1 0
63 Kyllini (aftershock) 31/10/1988 02:59:51 37.85 21.01 4.8 18 1 1 0
64 Kyllini (aftershock) 27/11/1988 16:38:45 37.88 20.99 4.5 8 1 1 0
65 Patras 22/12/1988 09:56:50 38.37 21.78 4.9 10 1 2 0
66 Patras 15/5/1989 22:40:04 38.28 21.79 4.8 1 0 1 0
67 Patras 31/8/1989 21:29:31 38.06 21.76 4.8 23 1 1 0
68 Near southeast coast of Sithonia peninsula 9/3/1990 05:35:50 39.93 23.97 4.6 10 0 1 0
69 Aigion 17/5/1990 08:44:06 38.39 22.22 5.1 26 0 1 0
70 Near east coast of Zakynthos Island 20/5/1990 05:57:24 37.76 20.85 4.5 11 0 1 0
71 Zakynthos Island 24/5/1990 18:51:49 37.73 20.97 4.5 1 0 1 0
72 Near east coast of Zakynthos Island 24/5/1990 19:59:06 37.8 20.91 4.8 1 0 1 0
73 Plati 8/8/1990 00:35:07 37.15 22.04 4.9 10 0 2 0
74 Kefallinia Island 24/8/1990 12:54:41 38.24 20.43 4.5 9 0 1 0
75 Near southeast coast of Sithonia peninsula 9/9/1990 19:00:39 39.9 24.02 5 1 0 1 0
76 Kefallinia Island 4/10/1990 03:19:16 38.21 20.43 4.5 6 0 1 0
77 Griva 21/12/1990 06:57:43 40.95 22.43 6.1 1 1 2 3
78 Near southeast coast of Crete 19/3/1991 12:09:23 34.673 26.358 5.5 5 0 1 0
79 Near southeast coast of Crete 19/3/1991 21:29:27 34.74 26.376 5.2 9 0 1 0
80 Kefallinia Island 26/6/1991 11:43:32 38.34 21.044 5.3 4 0 3 0
81 Near north coast of Kefallinia Island 2/1/1992 09:05:18 38.29 20.325 4.5 9 0 1 0
82 Kefallinia Island 23/1/1992 04:24:17 38.28 20.41 5.6 3 0 3 0
83 Near northwest coast of Kefallinia Island 25/1/1992 12:23:23 38.38 20.44 4.5 10 0 1 0
84 Mataranga 30/5/1992 18:55:40 38.04 21.45 5.2 12 1 5 0
85 Tithorea 18/11/1992 21:10:41 38.26 22.37 5.9 15 1 3 0
86 Pyrgos (foreshock) 14/2/1993 10:17:45 37.71 21.38 4.5 4 1 0 0
87 Pyrgos (foreshock) 25/3/1993 05:44:09 37.61 21.31 4.5 5 1 0 0
88 Pyrgos (foreshock) 26/3/1993 11:45:16 37.68 21.44 4.9 3 3 3 0
89 Pyrgos (aftershock) 26/3/1993 12:49:13 37.69 21.42 4.7 10 1 1 0
90 Pyrgos (aftershock) 26/3/1993 12:26:30 37.55 21.27 4.5 19 1 0 0
91 Pyrgos (aftershock) 30/3/1993 19:08:57 37.64 21.32 4.5 10 0 1 0
92 Gulf of Corinth 2/4/1993 02:22:59 38.16 22.62 5 5 1 0 0
93 Off coast of Levkas Island 6/4/1993 03:24:27 38.7 20.45 4.8 1 0 1 0
94 Gulf of Corinth 11/4/1993 05:18:37 38.34 21.91 5.3 10 1 2 0
95 Pyrgos (aftershock) 29/4/1993 07:54:29 37.76 21.46 4.8 0 1 0 0
96 Near coast of Filiatra 3/5/1993 06:55:06 37.07 21.46 5.2 1 0 1 3
97 Mouzakaiika 13/6/1993 23:26:40 39.25 20.57 5.3 5 2 3 0
98 Patras 14/7/1993 12:31:50 38.16 21.76 5.6 13 3 4 0
99 Patras (aftershock) 14/7/1993 12:39:13 38.18 21.64 4.6 10 1 2 0
100 Patras (aftershock) 14/7/1993 12:54:07 38.15 21.71 4.6 10 1 0 0
101 Pyrgos (aftershock) 7/10/1993 20:26:04 37.79 21.11 4.8 10 0 1 0
102 Near southwest coast of Levkas Island 9/10/1993 13:33:20 38.58 20.45 4.6 6 0 1 0
103 Off coast of Levkas Island 12/1/1994 07:32:57 38.702 20.359 4.6 7 0 1 0
104 Ionian 14/1/1994 06:07:48 37.61 20.88 4.9 10 0 1 0
105 Komilion 25/2/1994 02:30:50 38.73 20.58 5.4 5 2 3 0
106 Ionian 27/2/1994 22:34:52 38.69 20.46 4.8 10 2 1 0
107 Near southwest coast of Levkas Island 15/3/1994 22:41:04 38.602 20.459 4.5 0 0 1 0
108 Arta 14/4/1994 23:01:34 39.132 20.968 4.5 1 0 1 0
109 Levkas Island 18/7/1994 15:44:18 38.626 20.507 4.9 3 0 1 0
110 Paliouri 4/10/1994 19:46:21 39.976 23.643 5.1 10 0 1 0
111 Zakynthos Island 17/10/1994 09:02:17 37.756 20.912 4.6 14 0 1 0
112 Arnaia (foreshock) 5/3/1995 15:39:56 40.541 23.642 4.7 8 0 1 1
113 Arnaia (foreshock) 5/3/1995 21:36:54 40.58 23.65 4.6 8 0 1 2
114 Arnaia (foreshock) 4/4/1995 17:10:10 40.545 23.625 4.6 9 0 1 1
115 Arnaia 5/4/1995 0:34:11 40.59 23.6 5.3 14 0 2 2
www.nat-hazards-earth-syst-sci.net/10/2527/2010/ Nat. Hazards Earth Syst. Sci., 10, 2527–2537, 2010
2532 G-A. Tselentis and L. Vladutu: MMI attenuation engineering parameters and neural networks
Table 1. Continued.
No Earthquake Date Time Latitude Longitude Moment Focal Number
Name Magnitude Depth of Records
(dd-mm-yy) (

) (

) (km) B C D
116 Kozani 13/5/1995 8:47:15 40.183 21.66 6.5 14 0 3 5
117 Kozani (aftershock) 13/5/1995 11:43:31 40.1 21.6 5.2 10 0 0 1
118 Kozani (aftershock) 13/5/1995 18:06:01 40.28 21.52 4.6 29 0 0 1
119 Kozani (aftershock) 14/5/1995 14:46:57 40.13 21.66 4.5 0 0 1 0
120 Kozani (aftershock) 15/5/1995 04:13:57 40.08 21.65 5.2 9 0 3 1
121 Kozani (aftershock) 16/5/1995 23:00:42 40.02 21.56 4.7 0 0 1 0
122 Kozani (aftershock) 16/5/1995 23:57:28 40.09 21.62 4.9 0 0 2 0
123 Kozani (aftershock) 16/5/1995 04:37:28 40.01 21.58 4.8 9 0 1 0
124 Kozani (aftershock) 17/5/1995 04:14:25 40.046 21.58 5.3 10 0 3 1
125 Kozani (aftershock) 17/5/1995 09:45:07 40.01 21.56 5 0 0 2 0
126 Kozani (aftershock) 18/5/1995 06:22:55 40.03 21.56 4.6 0 0 1 0
127 Kozani (aftershock) 19/5/1995 06:48:49 40.09 21.6 5.2 7 0 4 0
128 Kozani (aftershock) 19/5/1995 07:36:19 40.06 21.61 4.8 0 0 1 0
129 Kozani (aftershock) 6/6/1995 04:36:00 40.14 21.61 4.8 0 0 4 1
130 Aigion 15/6/1995 00:15:51 38.362 22.2 6.5 10 1 8 1
131 Kozani (aftershock) 17/7/1995 23:18:15 40.21 21.55 5.2 22 0 1 1
132 Kozani (aftershock) 18/7/1995 07:42:54 40.101 21.575 4.7 10 0 1 0
133 Aigion (aftershock) 13/8/1995 05:17:29 38.101 22.81 4.5 8 0 1 0
134 Kozani (aftershock) 6/11/1995 18:51:48 39.92 21.62 4.8 13 0 5 1
135 East of Kithira Island 29/6/1996 01:09:03 36.351 23.179 4.5 6 0 0 1
136 Pyrgos 8/11/1996 11:43:45 37.684 21.425 4.7 0 0 1 0
137 Zakynthos Island 16/2/1997 11:03:19 37.676 20.723 4.9 8 0 1 0
138 Strofades (foreshock) 26/4/1997 22:18:34 37.181 21.385 5 7 0 0 1
139 Strofades (foreshock) 29/4/1997 23:52:17 37.416 20.713 4.5 2 0 1 0
140 South of Vathi 11/5/1997 10:27:56 38.412 23.588 4.6 30 0 3 0
141 Itea 11/5/1997 21:10:28 38.44 22.28 5.6 24 0 3 1
142 South of Rhodos 17/7/1997 13:21:01 36.412 28.192 4.5 5 0 1 1
143 South of Rhodos 18/7/1997 1:45:23 36.38 28.188 4.6 14 0 1 1
144 Varis 22/8/1997 03:17:47 40.148 21.572 4.5 23 0 0 1
145 Northwest of Makrakomi 21/10/1997 17:57:47 38.971 22.073 4.7 14 0 0 1
146 Strofades 18/11/1997 13:07:41 37.482 20.692 6.6 10 2 6 2
147 Strofades (aftershock) 18/11/1997 13:13:46 37.229 21.057 6 10 1 3 2
148 Strofades (aftershock) 18/11/1997 15:23:35 37.334 21.191 5.3 30 1 1 1
149 Strofades (aftershock) 18/11/1997 13:44:05 37.309 21.047 4.8 10 0 1 0
150 Strofades (aftershock) 19/11/1997 00:33:07 37.458 20.764 4.8 10 0 1 0
151 Ano Liosia 9/7/1999 11:56:51 38.08 23.58 6 17 2 7 0
Total no. of records 75 197 63
This relation, by establishing the input of the first layer
of the network, allows us to gradually calculate the value of
the global output of the network, thus ensuring its forward
propagation. When one compares this output with the
desired output, one can calculate the error function, generally
given by
e =
1
2

Y −
¯
Y

2
(5)
where Y is the desired output, and

¯
Y

is the obtained output.
The direction in which the weights are updated is given
by the negative of the gradient of (e) with respect to every
element of the weight. This process consists in minimising
(e) by a gradient descent. Thus, we try to modify the synaptic
weights to reduce (e). This is carried out using the following
relation.
w
s
ij
=−µ

e
s
j
Y
s−1
i

n
+

w
s
ji

n−1
(6)
where µ is the learning rate parameter, which usually takes
values between 0 and 0.5. The quantity e
s
j
is the locale of
the error of the j-th neuron in the s layer. Weights and bias
terms are first initialised at random values. In general, there
are no strict rules for determining the network configuration
for optimum training and prediction.
One representation of the ANN used in the present
investigation is shown in Fig. 4, which includes the most
commonly-used engineering seismological ground-motion
parameters as inputs.
Nat. Hazards Earth Syst. Sci., 10, 2527–2537, 2010 www.nat-hazards-earth-syst-sci.net/10/2527/2010/
G-A. Tselentis and L. Vladutu: MMI attenuation engineering parameters and neural networks 2533
24




Fig. 3



Fig. 4

Fig. 5
Fig. 4. Topology of the feed-forward k-NN-type ANN used in the present study.
5 Data processing
We consider automatic MMI assessment based on ground-
motion parameters as part of a larger category of problems
encountered in pattern recognition (Poggio and Girosi, 1990.
In the present investigation, we consider the following four
phases: (1) feature extraction, (2)classification, (3) pre-
processing and optimisation and (4) regression.
5.1 Feature extraction
During this phase, we combined the enhanced selection
capacities offered by GA with the performance of an ANN
as a classifier. At first, we used all nine commonly-used
input parameters that characterise earthquake ground motion
at a site corresponding to an MMI value. These measures
include M, log(R), PGV, log(PGV), S
a
, PGA, log(PGA), I
a
and CAV (Fig. 4). During this procedure, we use 9-bit string
(S) quantification with binary field-values as follows.
S
=
¸
M,logR,S
a
,PGV,logPGV,PGA,logPGA,I
a
,CAV
¸
(7)
For example, the three-parameter string [logR, I
a
, CAV] is
represented as the string 010000011. A genetic algorithm
was used to generate populations of strings out of the
512 possible combinations from 000000000 to 111111111.
The total number of available strings (that is, the equivalent
of chromosomes) at a certain time (i.e., after the Max-No-
Generations), which is known as the genome, was evaluated
by an ANN implemented as a k-nearest neighbour (k-NN).
This was achieved by comparing the corresponding MMI
(that is, the outputs) with the selected inputs out of the nine
possible inputs represented by the strings generated by the
genetic algorithm. In our implementation, we have allowed
a population size of 20, where the population size indicates
the maximum number of chromosomes (or strings) allowed
in a generation.
5.2 Classification
The second phase, which deals with classification, is
implemented using a k-NN-type neural network. The
k-nearest neighbours (k-NN) algorithm is a method for
classifying objects based on the closest training examples in
the feature space. The k-NN algorithm is a type of instance-
based learning, or lazy learning, where the function is only
approximated locally, and all computation is deferred until
classification.
In the machine-learning community, Instance-Based
Learning (IBL) (Aha et al., 1991), also known as memory-
based learning, is a family of learning algorithms that,
instead of performing explicit generalisations, compares new
instances with instances that have been observed during
training. It is called instance-based because it constructs
hypotheses directly from the training instances themselves.
A direct consequence of this approach is that the complexity
of the problem grows with the amount of data available for
training and testing.
Of the data set of 310 values, some data points were left
for testing, while most were considered for training. We
have used a GA-ANN with IBL approach in order to avoid
data changes produced by normalisation techniques. The
Euclidean metric is used to assess distances between the
training and testing epochs.
In our approach, we used one of the simplest examples
of IBL, namely, k-NN classifier and its Java implementation
based on the Weka-toolbox (Witten and Frank, 2005).
www.nat-hazards-earth-syst-sci.net/10/2527/2010/ Nat. Hazards Earth Syst. Sci., 10, 2527–2537, 2010
2534 G-A. Tselentis and L. Vladutu: MMI attenuation engineering parameters and neural networks
The k-NN classifier is amongst the simplest machine-
learning algorithms. In our case, it selects the optimal
combination of the nine inputs (see Eq. 7). An object is
an instance out of the 310 data values given in the features
space; it is a five-parameter string consisting of [M, log(R),
PGA, I
a
, CAV]. It is classified by a majority vote of its
neighbours, with the object being assigned to the class most
common among its k-nearest neighbours.
5.3 Pre-processing and optimisation
In the previous two phases, we considered all inputs as they
were, but for processing purposes, we converted all inputs
into integers by multiplying them by powers of 10 (H¨ ardle et
al., 1995; Mierswa et al., 2006).
Candidate solutions to the optimisation problem act like
individuals in a population, and a fitness function determines
the environment within which these solutions “live” (e.g.,
a cost function). Genetic algorithms are a particular class
of evolutionary algorithms (EA) (Fonseca and Fleming,
1995) that use techniques inspired by evolutionary biology
such as inheritance, mutation, selection, and crossover (or
recombination).
In other words, a GA quantifies information (namely, the
parameters of the k-NN classifier) in the form of strings (that
is, the chromosomes), and through the EA, only the fittest
chromosomes survive over the generations of the evolution.
Therefore, an important parameter for the proposed GA-
ANN method is Max-No-Gener, which allows the GA
algorithm to evolve to achieve the optimal solution. In our
case, there are several parameters that have to be modified (or
fine-tuned) to achieve optimal behaviour according to ANN.
These include the number of hidden layers and the number
of neurons in each hidden layer.
After finding the optimal Max-No-Gener parameter, we
determined k and the selection scheme (or the size of the
tournament). k was found by a trial-and-error procedure to
have a value of 2. One advantage of the selection mechanism
of a GA (Tobias and Lothar, 1995) is the independence of the
representation of the individuals. Only the fitness values of
individuals are taken into account.
A fitness function is a particular type of objective function
that prescribes the optimality of a solution (or chromosome)
in a genetic algorithm so that a particular chromosome may
be ranked among all other chromosomes from the genome.
Chromosomes which are found to be “more optimal” are
allowed to breed. That is, further binary combinations will
be created by the GA on the “skeleton” of these “more
optimal” chromosomes. Data sets can then be mixed using
any of several available techniques, thus producing a new
generation of chromosomes. This mix in our case can be
represented by taking the first four digits from one “optimal”
string and the last five digits from another “more optimal”
one, thereby creating a new chromosome (9-digit string)
Fig. 5. One-point crossover.
that can possibly perform better under k-NN classification.
This simplifies the analysis of the selection methods and
allows a comparison that can be used for all kinds of genetic
algorithms.
One of the frequently-used selection schemes is tourna-
ment selection. In this scheme, we run a tournament among
a few individuals chosen at random from the population
(i.e., from the genome) and select the one with the best
fitness for crossover as the winner, adjusted by variances in
tournament size. In genetic algorithm theory, the crossover
is the genetic operator used to modify the programming
of a chromosome or a group of chromosomes from one
generation to the next. It is analogous to natural reproduction
and biological crossover upon which the simplified GA
theory for computational intelligence was built.
First, a single crossover point on the strings of both parent
organisms is selected while avoiding extreme points. All
data beyond that point in either organism string are swapped
between the two organism strings. The resulting organisms
are the children (or offspring), as shown in Fig. 5.
If the tournament size is larger (chromosomes for which
the objective function, i.e., the error has a higher value),
weak individuals have a smaller chance of being selected for
breeding, crossover, and perpetuation in the next generations.
Performance was quantified by RMS criteria and the squared
error. A flow chart describing all of the above operations is
presented in Fig. 6.
To validate the performance of the above-mentioned GA-
ANN selection schemes, we selected a validation scheme
based on regression performance. The results for the
selection of the first optimal parameter from those described
above (that is, Max-No-Gener) are presented in Table 2.
From this table, we note the following cases.
Supposing that Max-No-Gener =75 and that we obtained
S
1
as the fittest string, where S
1
is given by the GA+k-NN
algorithm and corresponds to the combination [M, log(R),
log(PGV), I
a
, CAV]. This is represented as 110010011.
The fittest chromosome is considered the one for which the
objective function has an extreme value. In our case, the
objective function is the squared error of the k-NN type of
ANN, and so the objective function must have a minimum
value.
Nat. Hazards Earth Syst. Sci., 10, 2527–2537, 2010 www.nat-hazards-earth-syst-sci.net/10/2527/2010/
G-A. Tselentis and L. Vladutu: MMI attenuation engineering parameters and neural networks 2535
Table 2. The parameters retained by the GA-ANN selection algorithm and corresponding regression performance (columns 2 and 3).
Max-No-Gener. Sq error RMS error root relative The retained parameters
squared error
75 0.329±0.170 0.553±0.152 0.608±0.178 S
1
=[M,logR,logPGV,I
a
,CAV]
80 0.311±0.154 0.456±0.169 0.593±0.195 S
2
=[M,logR,PGA,I
a
,CAV]
90 0.339±0.231 0.548±0.196 0.618±0.189 S
3
=[logR,I
a
,CAV]
94 0.324±0.194 0.571±0.176 0.608±0.194 S
4
=[M,logR,S
a
,logPGV,PGA,I
a
,CAV]
Fig. 6. Flowchart showing the sequences of feature-selection and
classification.
If we consider the case with Max-No-Gener =80 with S
2
as the fittest string, this corresponds to the combination [M,
log(R), PGA, I
a
, CAV] and is represented as 110001011.
In this situation, S
2
is retained as having the best fitness
because the output of the k-NN classifier for S
1
is given by
the squared error 0.329±0.170 and for S
2
is given by of the
squared error of 0.311±0.154. In other words, both the error
and its standard deviation are lower than in the case of S
1
.
Setting Max-No-Gener higher than optimally at 90, S
3
then corresponds to the combination [logR, I
a
, CAV]
and is represented as 010000011; this solution is rejected
due to a worse regression performance (that is, a higher
RMS error). The solution with Max-No-Gener =94 and
string S
4
corresponding to the combination [M, logR, S
a
,
PGV, logPGV, PGA, logPGA, I
a
, CAV] and represented by
111111111 is also rejected because it accepted almost all of
the input parameters and is a trivial solution.
Finally, an underdetermined solution, which results if
values of Max-No-Gener are in the range of 25 to 70,
is rejected because GAs require a minimum number of
generations to obtain optimal fitness among all available
individuals.
Judging from the above, an optimal value of Max-No-
Gener =80 with a minimum RMS was selected. In this case,
the optimal selected input parameters are M, logR, PGA,
I
a
, and CAV. These are the parameters used to express MMI
throughout the regression process.
5.4 Regression
In a case in which the existing data are not sufficient
for analysis because we must use part of the data for
validation and test sets, it is common to use a cross-
validation or rotation estimation method (Kohavi, 1995).
This is a technique for assessing how the results of a
statistical analysis will generalise to an independent data
set. It is mainly used in applications in which the goal
is prediction, particularly when one wants to estimate how
accurately a predictive model will perform in practice.
Cross-validation involves partitioning a portion of the data
into complementary subsets, performing analysis on one
subset called the training set, and validating the analysis on
the other subset, which is called the validation or testing
set. For relatively large datasets, a higher cross-validation
is usually used (25 in our case).
Next, the selected optimal input parameters [M, logR,
PGA, I
a
, and CAV] are considered as X multivariate
inputs, and the response variable Y is understood to
indicate MMI. To calculate the coefficients that link Y
to the five-dimensional X variable, we used the linear
www.nat-hazards-earth-syst-sci.net/10/2527/2010/ Nat. Hazards Earth Syst. Sci., 10, 2527–2537, 2010
2536 G-A. Tselentis and L. Vladutu: MMI attenuation engineering parameters and neural networks
Table 3. The parameters for linear regression obtained by the
program XploRe for the input data selected by the proposed GA-
ANN method.
PARA- Beta SE StandB t-test p-value
METERS
b[0, ] 8.8236 4.0481 0.0000 1.439 0.1513
b[1, ] 0.4173 0.9553 0.2733 0.437 0.6625
b[2, ] –7.9601 12.2908 –0.4084 –0.648 0.5177
b[3, ] 0.3801 0.4533 0.1499 0.839 0.4024
b[4, ] 1.1046 0.5022 0.7718 2.200 0.0286
b[5, ] –0.5508 0.5965 –0.2046 –0.923 0.3566
Table 4. The statistical parameters obtained for the linear regression
using the ANOVA test.
ANOVA SS df MSS F-test p-value
Regression 155.893 5 31.179 67.556 0.0000
Residuals 140.30 4 304 0.462
Total Variation 296.197 309 0.959
Multiple R =0.72548
R
2
=0.52632
Adjusted R
2
= 0.51853
Standard Error =0.67936
regression approach described by the following equation
(H¨ ardle et al., 1995).
Y =b
0
+
5
¸
i=1
b
i
∗X
i
(8)
where b
0
is the intercept, and b
1
,...,b
5
are the coefficients
for the ground parameters.
The obtained b values are presented in Table 3, and the
results of the ANOVA statistical test are shown in Table 4.
Accordingly, the relation that describes MMI as a function
of [M, logR, PGA, I
a
, CAV] is as follows.
MMI =8.824+0.417M−7.960logR+0.380PGA
+1.105I
a
−0.551CAV (9)
Figure 7 shows the time series corresponding to the
original data and the results of the regression analysis.
For the GA-ANN selection scheme, we used a Java-based
implementation built around the Weka (the IBk lazy learner)
data-mining system (Witten and Frank, 2005).
26










Fig. 7. Output of the regression algorithm (in red) and the original
MMI data (in blue) for the 310 considered data points.
6 Conclusions
In this research, we presented a new approach based on
ANN and GA to model the relationship between MMI
attenuation and engineering ground motion parameters. The
performance of this new regression approach has been tested
using a Greek strong motion database with satisfactory
results.
We note that not all of the features selected in the GA-
ANN approach have the same influence on MMI attenuation.
An approach based on evolutionary algorithms can be useful
in weighting the importance of those features. Additionally, a
new type of neural network (namely, an evolutionary neural
network) can be used to replace the classical k-NN used in
the current paper. If we implement an expert system, we
can derive a real-time signal-processing system, which can
be used for near real-time damage assessment and shake map
construction.
Acknowledgements. We are grateful to Laurentin Danciu for
his most usefull comments. The help of M. E. Condakis and
2 anonymous reviewers is greately appreciated.
Edited by: M. E. Contadakis
Reviewed by: two anonymous referees
References
Ambraseys, N., Smit, P., Douglas, J., Margaris, B., Sigbjornsson,
R., Olafsson, S., Suhadolc, P., and Costa, G.: Internet-site for
European strong-motion data, Boll. Geofis. Teor. Appl. 45(3),
113–129, 2004.
Aha, D. W.: Instance-based Learning, Machine Learning, Kluwer
Academic Publishers, 6, 36–77, 1991.
Nat. Hazards Earth Syst. Sci., 10, 2527–2537, 2010 www.nat-hazards-earth-syst-sci.net/10/2527/2010/
G-A. Tselentis and L. Vladutu: MMI attenuation engineering parameters and neural networks 2537
Arias, A.: A measure of earthquake intensity, Cambridge, MA,
1970.
Atkinson, G. M. and Kaka, S. I.: Relationship between felt intensity
and instrumental ground motion in the central United States and
California, B. Seismol. Soc. Am., 97(2), 497–510, 2007.
Bishop, C. M.: Neural Networks for Pattern Recognition, Oxford
University Press, UK, 1996.
Cabanas, L., Benito, B., and Herraiz, M.: An approach to the
measurement of the potential structural damage of earthquake
ground motions, Earthq. Eng. Struct. D., 26, 79–92, 1997.
Danciu, L. and Tselentis, G-A.: Engineering ground motion
parameters attenuation relationships for Greece, B. Seismol. Soc.
Am., 97(1B), 162–183, 2007.
Fonseca, C. M. and Fleming, P. J.: Multi-objective genetic
algorithms made easy: Selection, sharing and mating restriction,
in: Proceedings of the first international conference on generic
algorithms in engineering systems: Innovations and applications,
Sheffield, UK, 45–52, 1995.
Fujinawa, Y., Matsumoto, T., and Takahashi, K.: Method for
estimating origin time, hypocentral distance and scale based on
electric field observation, and apparatus for prediction, Patent
No. US 6,885,945 B2, 2005.
H¨ ardle, W.: Applied Nonparametric Regression, Humboldt-
Universit¨ at zu Berlin Wirtschaftswissenschaftliche Fakult¨ at
Institut f¨ ur Statistik und
¨
Okonometrie, 1994.
H¨ ardle, W., Klinke, S., and Turlach, B.: XploRe – an Interactive
Statistical Computing, Environment, Springer, Heidelberg, 1995.
Haykin, S.: Neural Networks, 2nd edn., MacMillan College
Publishing Company, 1999.
Kaka, S. I. and Atkinson, G. M.: Relationships between felt
intensity and instrumental ground motion in the central United
States and California, B. Seismol. Soc. Am., 97(2), 497–510,
2004.
Kalogeras, I. S., Marketos, G., and Theodoridis, Y.: A tool for
collecting, uerying, and mining macroseismic data, B. Geol. Soc.
Greece, XXXVI, 2004.
Karim, K. R. and R. Yamazaki, R.: Correlation of JMAinstrumental
seismic intensity with strong motion parameters, Earthq. Eng.
Struct, D., 31(5), 1191–1212, 2002.
Kohavi, R.: A study of cross-validation and bootstrap for
accuracy estimation and model selection, Proceedings of
the Fourteenth International Joint Conference on Artificial
Intelligence, (Morgan Kaufmann, San Mateo), 2(12), 1137–
1143, 1995.
Koliopoulos, P. K., Margaris, B. N., and Klims, N. S.: Duration and
energy characteristics of Greek strong motion records, J. Earthq.
Eng., 391–417, 1998.
Kostov, M. K.: Site specific estimation of cumulative absolute
velocity, 18th International Conference on Structural Mechanics
in ReactorTechnology, Beijing, China, 2005.
Krinitzsky, E. L. and Chang, F. K.: Intensity-related earthquake
ground motion, Bull. Int. Assoc. Eng. Geol. 4, 425–435, 1988.
Margottini, C., Molin, D., and Serva, L.: Intensity versus ground
motion: a new approach using Italian data, Eng. Geol., 33, 45–
58, 1992.
Mierswa, I., Wurst, M., Klinkenberg, R., Scholz, M., and
Euler, T.: YALE: Rapid prototyping for Complex Data Mining
Tasks, in Proceedings of the 12th ACM SIGKDD International
Conference on Data Mining and Knowledge Discovery, (KDD-
06), 2006.
Panza, G. F., Cazzaro, R., and Vaccari, F.: Correlation between
macroseismic intensities and seismic ground motion parameters,
Ann. Geofis., 40(5), 1371–1382, 1997.
Poggio, T. and Girosi, F.: Regularization algorithms for learning
that are equivalent to multilayer perceptrons, Science, 247, 978–
982, 1990.
Sokolov, Y. V.: Spectral parameters of the ground motions in
Caucasian seismogenic zones, B. Seismol. Soc. Am., 88(6),
1438–1444, 1998.
Skolov, Y. V.: Seismic Intensity and Fourier Acceleration Spectra:
Revised Relationship, Earthq. Spectra, 18(1), 161–187, 2002.
Tobias, B. and Lothar, T.: A comparison of selection schemes used
in evolutionary algorithms, Evol. Comput., 4(4), 361–394, 1995.
Trifunac, M. D. and Westermo, B.: A note on the correlation
of frequency-dependent duration of strong earthquake ground
motion with the modified Mercalli intensity and the geologic
conditions at the recording stations, B. Seismol. Soc. Am., 67(3),
917–927, 1977.
Tselentis, G.-A. and Danciu, L.: Empirical relationships between
MMI and engineering ground motion parameters in Greece, B.
Seimol. Soc. Am., 98(4), 1863–1875, 2008.
Witten, I. H. and Frank, E.: Data Mining – Practical Machine
Learning Tools and Techniques, Elsevier, 2005.
Wald, D. J., Quitoriano, V., Heaton, T. H., and Kanamori, H.:
Relationships between peak ground acceleration, peak ground
velocity, and modified mercalli intensity in California, Earthq.
Spectra, 15(3), 557–564, 1999.
Wu, Y. M., Teng, T. I., Shin, T. C., and Hsiao, N. C.: Relationship
between peak ground acceleration, peak ground velocity and
intensity in Taiwan, B. Seismol. Soc. Am., 93, 386–396, 2003.
www.nat-hazards-earth-syst-sci.net/10/2527/2010/ Nat. Hazards Earth Syst. Sci., 10, 2527–2537, 2010

it was also separately estimated in part from the macroseismic data provided by the Geodynamic Institute of the National Observatory of Athens (Kalogeras et al. 2010 The strong-motion records used for the present investigation were provided by the European Strong Motion Database (Ambraseys et al. These relationships have been used to generate maps of estimated shaking intensities within a few minutes of the event based on recorded peak motions.. The macroseismic information was available in part from the digital database of the European strong-motion data. 2004). Cumulative absolute velocity (CAV) is defined as the integral of the absolute value of ground acceleration over the following seismic time-history record. Therefore. These maps provide a rapid visualisation of the extent of expected damages following an earthquake and can be used for emergency response. The objective of the present investigation is to uncover the hidden. complex and often fuzzy relations between the engineering ground-motion parameters and macroseismic intensity and to express these relations in the form of input/output dependencies. and the release of public information through the media. characteristic intensity. Ia . These systems are low-cost solutions aimed at reducing the seismic risk to vital facilities. Tselentis and Danciu (2008) derived empirical regression equations for modified Mercalli intensity (MMI) and for various ground-motion parameters such as duration. The ground-motion parameters and the damage potential threshold are essential for these systems. and high-speed trains. neural networks can be viewed as advanced mathematical models used to discover complex correlations between physical process variables from a set of perturbed observations. (3) where |a(t)| is the absolute value of acceleration. The general criterion of selecting the appropriate MMI value was to allocate at each station the nearest MMI values within an uncertainty of one unit to every station. Tselentis and L. the average of the values was used (Tselentis and Danciu. the number of strain cycles. pipelines. Hazards Earth Syst. 2 IaX = g 0 [aX (t)]2 dt. 1999. we decided not to disaggregate the soil effect because the soft soil subset comprises less than 30% of the total data set. parameters that incorporate in their definition these previouslymentioned characteristics are more reliable predictors of an earthquake’s damage potential than parameters related solely to the amplitude of ground motion (such as peak ground acceleration. the potential for earthquake damage depends on the time duration of motion. T T (2) where T is the undamped natural period of a single-degreeof-freedom (SDOF) oscillator. is calculated as follows: t Recently. as defined by Arias (1970). at an equal distance from the station. Maps of the reported MMI values together with the strong-motion instrument locations were plotted to show reasonable confidence that the allocated MMI values are within one unit of the assigned value. 2008). and t is the total duration of ground motion. In this context. t CAV = 0 |a(t)|dt. The most commonly-used engineering ground-motion parameters in addition to PGA and PGV are Arias intensity (Ia ). If more than one MMI value was observed near the station location.. CAV. Housner’s spectrum intensity. 2007). Ia . Another application of engineering ground-motion parameters involves the development of early warning systems. for engineering purposes. which are often poor indicators of structural damage. and t is the total duration of ground motion. Sa = 2π 2π Sv = ωSv = Sd = ω2 Sd . 3 Data set 2 Engineering seismological parameters Because structure and equipment damage is measured according to its inelastic deformation. At this preliminary stage of intensity ground-motion investigation. Sci. loss estimation. 2527–2537. (1) where aX (t) is the acceleration time history in the x direction. such as nuclear power plants. and cumulative absolute velocity (CAV). 1996) provides valuable insights for addressing the complicated problem of expression these relations. 2004). is the total energy per unit weight stored by a set of undamped simple oscillators at the end of ground motion. Spectrum acceleration (Sa ) is the most common-response spectral parameter and is related to spectrum velocity (Sv ) and spectrum displacement (Sd ) according to the following expression.net/10/2527/2010/ . Vladutu: MMI attenuation engineering parameters and neural networks in the x-direction (IaX ). and total elastic input energy index. www. acceleration response spectrum (Sa ). and the energy content of the earthquake. More details on these data can be found in Danciu and Tselentis (2007). This approach provides rapid visualisation of the macroseismic distribution in the area surrounding the recording stations and might be an efficient approach to assigning MMI values to minimise errors (Atkinson and Kaka.. the energy absorption capacity of the structure or equipment. The Arias intensity for ground motion Nat. The emergence of neural network technology (Haykin. or PGA).nat-hazards-earth-syst-sci. 10. Bishop.2528 G-A.

the neurons simultaneously start a state of activity that causes a small modification of the synaptic forces between them. for each horizontal component. Nat. including an input layer that shows the number of neurons linked to the dimensionality of the input. Using the recorded strong-motion data for these earthquakes. When one presents the network with a form to be learned. and apply previous experiences to our every action. 1 and Table 1. This function maps the output to its input and can be expressed by the following equation. General topology of a feed-forward ANN with one hidden layer. we compute the previously mentioned engineering seismological parameters. 1 2529 Fig. Sci.. This is followed by a quantitative reconfiguration of all of the synapses: some of them become very strong. Fig. These cells are known as neurons. think. The learned form is not directly memorised at a precise place. 1 G-A. Tselentis and L. Figure 2 shows a simplified view of an ANN. 2010 . The arithmetic average between the two horizontal components of the independent variables was used for the present Fig. Epicentral distribution of earthquakes used in the present analysis. Two classical nonlinear activation functions. 2 synapses) between them. 4 Artificial neural networks Fig. and an output layer. 4 i=1 s Wj i · Yis−1 (4) where b is the bias. 1 Fig. such as the sigmoid function in Fig. The power of the brain comes from the number of neurons and the multiple connections (or Fig.net/10/2527/2010/ Fig. become weak. 2. It consists of a network of simple processing elements (that is. 3. 10. This configuration is supported by the values of the synaptic forces. artificial neurons) that are organised in several layers. 2527–2537. Hazards Earth Syst. across a very large set of possible configurations. The hidden layer represents network inputs.Fig. 3. 1. 2 The final data set consists of 310 records from 151 earthquakes and is depicted in Fig. R Yjs = f b+ Fig. The most basic element of the human brain involves a specific type of cell that provides us with the ability to remember. Let Yjs represent the output of the j -th neuron at layer s.nat-hazards-earth-syst-sci. while others www. s is the weight connecting the i-th neuron in layer s to the Wi j 23 j -th neuron at layer s-1. which is a particular configuration of the activity of each neuron. 3 Fig. one or several hidden layers. it corresponds to a particular energy state of the network. 2 investigation. 23 An artificial neural network (ANN) is an informationprocessing paradigm inspired by the way biological nervous systems such as the brain process information. The neurons have their activation function characterised by a nonlinear function. Vladutu: MMI attenuation engineering parameters and neural networks Fig.

25 20.2 5.6 6.55 20.net/10/2527/2010/ .8 5.8 5.49 21.6 5.85 22.02 20.1 5.79 39.11 22.3 40.2 6.65 21. 2010 www.7 40.9 5.78 38.46 36.27 39.9 5.9 5.2 5.08 40.6 4.9 5.01 20.26 20.2530 G-A.2 5.1 5 4.9 4.48 21.4 5.26 40.68 21.93 38.05 20.81 23.099 38.92 23.4 4.7 5.21 40.21 39.8 4.135 38.8 4.54 21.47 20.89 21.99 39.89 21.4 4.7 6.02 20.25 20.76 21.1 4.46 22.81 38.1 37.106 21.91 37.5 5.2 6.6 5.79 37.3 5.64 38.3 38.35 38.nat-hazards-earth-syst-sci.1 38.24 37.98 23. Database of strong motion records used.75 22.3 5.2 6.33 28.6 4.9 4.3 4.5 5 4.82 Longitude (◦ ) 20.7 38.06 26.65 36.71 38.1 5 4.263 20. Tselentis and L.2 5.254 22. No Earthquake Name Date (dd-mm-yy) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 Kefallinia Island Ionian Ionian Patras Amfissa Volvi Achaia Volvi Almiros (aftershock) Almiros (aftershock) Almiros (aftershock) Alkion Alkion Panagoula Levkas Preveza Paliambela Kefallinia Island Kefallinia (aftershock) Kefallinia (aftershock) Kyllini (foreshock) Etolia Kefallinia (aftershock) Kefallinia (aftershock) Off coast of Magion Oros peninsula Ierissos (foreshock) Ierissos Near southeast coast of Zakynthos Island Gulf of Corinth Arnissa Kranidia Kremidia (aftershock) Gulf of Amvrakikos Anchialos Gulf of Kiparissiakos Near coast of Preveza Drama Aghios Vasileios Skydra-Edessa Kalamata Kalamata (aftershock) Tsipiana Near northwest coast of Kefallinia Island Dodecanese Kounina Near northeast coast of Crete Kalamata (aftershock) Near southwest coast of Peloponnes Near northeast coast of Rodos Island Astakos Kefallinia Island Gulf of Corinth Ionian Gulf of Corinth Etolia Etolia Agrinio 17/9/1972 11/4/1973 11/4/1973 29/1/1974 29/12/1977 7/4/1978 18/5/1978 20/6/1978 16/7/1980 26/9/1980 8/11/1980 24/2/1981 25/2/1981 25/5/1981 27/5/1981 3/10/1981 4/10/1981 17/1/1983 17/1/1983 31/1/1983 20/2/1983 16/3/1983 23/3/1983 23/3/1983 8/6/1983 14/6/1983 26/8/1983 10/4/1984 17/8/1984 7/9/1984 25/10/1984 25/10/1984 22/3/1985 30/4/1985 9/7/1985 31/8/1985 11/9/1985 18/2/1986 18/2/1986 13/9/1986 15/9/1986 2/1/1987 27/2/1987 10/5/1987 14/5/1987 2/9/1987 6/10/1987 12/10/1987 25/10/1987 22/1/1988 6/2/1988 4/3/1988 24/4/1988 7/5/1988 18/5/1988 22/5/1988 3/8/1988 14:07:12 15:52:12 16:11:36 15:12:43 16:52:59 22:23:28 00:18:49 20:03:22 00:06:58 04:19:21 09:15:59 20:53:39 02:35:53 23:04:00 15:04:02 15:16:20 08:33:32 12:41:31 15:53:57 15:27:02 05:45:12 21:19:41 19:04:06 23:51:08 15:43:53 04:40:43 12:52:09 10:15:12 21:22:58 18:57:12 14:38:30 09:49:15 20:38:39 18:14:13 10:20:51 06:03:47 23:30:43 05:34:42 14:34:04 17:24:34 11:41:28 05:35:36 23:34:52 09:27:02 06:29:11 12:28:23 14:50:12 22:51:14 13:02:00 06:18:55 10:35:25 03:56:07 10:10:33 20:34:52 05:17:42 03:44:15 11:38:57 Time Latitude (◦ ) 38. 2527–2537.82 20.9 5.79 23. Sci.7 5.89 20.35 38. Hazards Earth Syst.23 22.36 21.72 38.2 38.19 22.96 38.86 22.76 22.8 21..45 37.92 20.32 38.49 20.3 38.2 5 4.76 38.86 20.56 22.03 37.4 4.245 38.55 40.29 38.08 38. Vladutu: MMI attenuation engineering parameters and neural networks Table 1.07 22.24 36.2 4.43 22.5 4.729 39.22 40.77 20.13 21.3 5.83 22.5 6.9 Moment Magnitude Focal Depth (km) 1 7 15 13 10 6 26 6 12 5 5 10 8 15 15 10 10 5 11 4 15 25 25 3 22 10 12 6 24 5 20 11 6 13 10 15 18 3 10 1 12 20 5 6 9 18 30 18 10 10 10 5 1 10 26 15 28 Number of Records B 0 1 1 1 1 1 1 3 0 0 0 2 1 1 1 1 1 1 0 0 1 1 1 1 0 0 0 0 1 0 0 0 0 0 0 1 0 2 0 1 0 0 1 0 0 0 1 0 0 1 0 0 2 1 0 2 0 C 0 0 0 0 0 0 0 1 1 1 1 0 0 0 0 1 0 1 1 1 1 0 0 2 1 1 1 1 0 1 0 0 0 1 0 2 2 1 1 2 3 1 2 1 1 1 1 0 0 0 1 1 0 1 2 2 1 D 2 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 2 0 0 1 1 2 1 1 1 0 2 1 1 1 1 0 0 1 0 0 1 1 1 0 0 0 0 0 1 0 1 Nat.7 5.5 5.64 38.3 38.64 21.17 35.95 21.35 23.41 24.13 38.3 4.41 37.24 39 41.83 38.86 38.2 5.81 20.1 5.842 23.8 4.13 36.61 23.44 40.68 28.12 37. 10.5 5.9 4.11 5.66 40.08 21.78 38.

62 20.3 5.06 39.325 20.5 5 4.91 21.39 37.602 39.27 21.44 21.976 37.69 37.1 4.02 20.38 38.358 26.69 38. Tselentis and L.58 20.28 38.61 38.88 37.6 4.8 4.58 40.5 4.8 37.55 37.43 26.04 38.71 21.7 5..73 37.41 20.5 4.643 20.5 4.31 21.15 37.8 4.64 21.88 20.11 20.8 5.61 37.08 21.nat-hazards-earth-syst-sci.79 21.44 21.43 22. Vladutu: MMI attenuation engineering parameters and neural networks Table 1.26 37. 2010 .673 34.6 4.25 38.5 5.93 37.642 23. 10.7 4.59 Longitude (◦ ) 21. Hazards Earth Syst.24 39.8 4.6 4.3 4.99 21.2 5.9 5.6 5.57 21.net/10/2527/2010/ Nat.6 4.5 4.64 38.93 38.37 38.45 22.5 5.76 21.16 38.1 5.34 37.76 37.88 38.6 4.33 20.507 23.29 38.912 23.73 38.3 4.76 23.42 21.8 4.626 39.4 4.702 37.968 20.16 38. Continued.376 21. 2527–2537.46 21.18 38.9 4.7 38.9 4.625 23.8 4.359 20.97 20.6 4.94 21.5 5.5 4.5 6.22 20.78 21.68 37.044 20.132 38.21 40.74 38.02 21.5 5 4. Sci.85 37.46 20.79 38.37 21.6 5.91 22.6 5.9 4.6 4.34 38.07 39.5 4.04 20.G-A.53 37.43 24.45 20.6 4.97 22.545 40.69 37.541 40.5 4.9 4.85 20.9 4.58 38.2 5.8 5.5 4.5 4.46 20.71 37.9 38.2 5.756 40.3 Moment Magnitude Focal Depth (km) 12 5 4 20 20 18 8 10 1 23 10 26 11 1 1 10 9 1 6 1 5 9 4 9 3 10 12 15 4 5 3 10 19 10 5 1 10 0 1 5 13 10 10 10 6 7 10 5 10 0 1 3 10 14 8 8 9 14 2531 Number of Records B 0 1 1 3 1 1 1 1 0 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 1 1 1 1 3 1 1 0 1 0 1 1 0 2 3 1 1 0 0 0 0 2 2 0 0 0 0 0 0 0 0 0 C 2 1 5 1 1 1 1 2 1 1 1 1 1 1 1 2 1 1 1 2 1 1 3 1 3 1 5 3 0 0 3 1 0 1 0 1 2 0 1 3 4 2 0 1 1 1 1 3 1 1 1 1 1 1 1 1 1 2 D 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 3 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 2 1 2 www.9 5.28 38.15 38. No Earthquake Name Date (dd-mm-yy) 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 Kyllini (foreshock) Kyllini (foreshock) Kyllini Trilofon Kyllini (aftershock) Kyllini (aftershock) Kyllini (aftershock) Patras Patras Patras Near southeast coast of Sithonia peninsula Aigion Near east coast of Zakynthos Island Zakynthos Island Near east coast of Zakynthos Island Plati Kefallinia Island Near southeast coast of Sithonia peninsula Kefallinia Island Griva Near southeast coast of Crete Near southeast coast of Crete Kefallinia Island Near north coast of Kefallinia Island Kefallinia Island Near northwest coast of Kefallinia Island Mataranga Tithorea Pyrgos (foreshock) Pyrgos (foreshock) Pyrgos (foreshock) Pyrgos (aftershock) Pyrgos (aftershock) Pyrgos (aftershock) Gulf of Corinth Off coast of Levkas Island Gulf of Corinth Pyrgos (aftershock) Near coast of Filiatra Mouzakaiika Patras Patras (aftershock) Patras (aftershock) Pyrgos (aftershock) Near southwest coast of Levkas Island Off coast of Levkas Island Ionian Komilion Ionian Near southwest coast of Levkas Island Arta Levkas Island Paliouri Zakynthos Island Arnaia (foreshock) Arnaia (foreshock) Arnaia (foreshock) Arnaia 22/9/1988 30/9/1988 16/10/1988 20/10/1988 22/10/1988 31/10/1988 27/11/1988 22/12/1988 15/5/1989 31/8/1989 9/3/1990 17/5/1990 20/5/1990 24/5/1990 24/5/1990 8/8/1990 24/8/1990 9/9/1990 4/10/1990 21/12/1990 19/3/1991 19/3/1991 26/6/1991 2/1/1992 23/1/1992 25/1/1992 30/5/1992 18/11/1992 14/2/1993 25/3/1993 26/3/1993 26/3/1993 26/3/1993 30/3/1993 2/4/1993 6/4/1993 11/4/1993 29/4/1993 3/5/1993 13/6/1993 14/7/1993 14/7/1993 14/7/1993 7/10/1993 9/10/1993 12/1/1994 14/1/1994 25/2/1994 27/2/1994 15/3/1994 14/4/1994 18/7/1994 4/10/1994 17/10/1994 5/3/1995 5/3/1995 4/4/1995 5/4/1995 12:05:39 13:02:54 12:34:05 14:00:59 14:58:18 02:59:51 16:38:45 09:56:50 22:40:04 21:29:31 05:35:50 08:44:06 05:57:24 18:51:49 19:59:06 00:35:07 12:54:41 19:00:39 03:19:16 06:57:43 12:09:23 21:29:27 11:43:32 09:05:18 04:24:17 12:23:23 18:55:40 21:10:41 10:17:45 05:44:09 11:45:16 12:49:13 12:26:30 19:08:57 02:22:59 03:24:27 05:18:37 07:54:29 06:55:06 23:26:40 12:31:50 12:39:13 12:54:07 20:26:04 13:33:20 07:32:57 06:07:48 02:30:50 22:34:52 22:41:04 23:01:34 15:44:18 19:46:21 09:02:17 15:39:56 21:36:54 17:10:10 0:34:11 Time Latitude (◦ ) 37.1 4.65 23.459 20.9 40.76 37.01 20.8 4.7 4.3 4.8 4.45 21.32 22.96 22.38 21.95 34.

188 21. This is carried out using the following relation.713 23.183 40.7 6.01 40.5 4. Thus.5 4.08 40..412 36. One representation of the ANN used in the present investigation is shown in Fig.44 36.191 21.01 40.58 Moment Magnitude Focal Depth (km) B Number of Records C 3 0 0 1 3 1 2 1 3 2 1 4 1 4 8 1 1 1 5 0 1 1 0 1 3 3 1 1 0 0 6 3 1 1 1 7 197 D 5 1 1 0 1 0 0 0 1 0 0 0 0 1 1 1 0 0 1 1 0 0 1 0 0 1 1 1 1 1 2 2 1 0 0 0 63 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 Total no. 2010 s s wij = −µ ej Yis−1 + n s wj i n−1 (6) where µ is the learning rate parameter. 2527–2537.7 4.412 38.723 21. which includes the most commonly-used engineering seismological ground-motion parameters as inputs. allows us to gradually calculate the value of the global output of the network.65 21.62 21.575 22.13 40.8 6 14 10 29 0 9 0 0 9 10 0 0 7 0 0 10 22 10 8 13 6 0 8 7 2 30 24 5 14 23 14 10 10 30 10 10 17 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2 1 1 0 0 2 75 This relation.9 5 4.21 40.588 22.7 4. Hazards Earth Syst.03 40.5 5. of records Kozani Kozani (aftershock) Kozani (aftershock) Kozani (aftershock) Kozani (aftershock) Kozani (aftershock) Kozani (aftershock) Kozani (aftershock) Kozani (aftershock) Kozani (aftershock) Kozani (aftershock) Kozani (aftershock) Kozani (aftershock) Kozani (aftershock) Aigion Kozani (aftershock) Kozani (aftershock) Aigion (aftershock) Kozani (aftershock) East of Kithira Island Pyrgos Zakynthos Island Strofades (foreshock) Strofades (foreshock) South of Vathi Itea South of Rhodos South of Rhodos Varis Northwest of Makrakomi Strofades Strofades (aftershock) Strofades (aftershock) Strofades (aftershock) Strofades (aftershock) Ano Liosia 13/5/1995 13/5/1995 13/5/1995 14/5/1995 15/5/1995 16/5/1995 16/5/1995 16/5/1995 17/5/1995 17/5/1995 18/5/1995 19/5/1995 19/5/1995 6/6/1995 15/6/1995 17/7/1995 18/7/1995 13/8/1995 6/11/1995 29/6/1996 8/11/1996 16/2/1997 26/4/1997 29/4/1997 11/5/1997 11/5/1997 17/7/1997 18/7/1997 22/8/1997 21/10/1997 18/11/1997 18/11/1997 18/11/1997 18/11/1997 19/11/1997 9/7/1999 8:47:15 11:43:31 18:06:01 14:46:57 04:13:57 23:00:42 23:57:28 04:37:28 04:14:25 09:45:07 06:22:55 06:48:49 07:36:19 04:36:00 00:15:51 23:18:15 07:42:54 05:17:29 18:51:48 01:09:03 11:43:45 11:03:19 22:18:34 23:52:17 10:27:56 21:10:28 13:21:01 1:45:23 03:17:47 17:57:47 13:07:41 13:13:46 15:23:35 13:44:05 00:33:07 11:56:51 40.192 28.5 4.nat-hazards-earth-syst-sci.62 23.52 21.6 5.8 4.572 22.2 4.58 21.057 21.28 40.81 21.6 21.482 37.764 23.56 21.6 6 5. 10.5 4. Sci.14 38.28 28.09 40.676 37.net/10/2527/2010/ .5 5.229 37. we try to modify the synaptic weights to reduce (e).2532 Table 1.047 20.458 38.38 40.56 21.101 39.2 4.046 40.351 37.55 21.5 5.181 37.334 37. No Earthquake Name G-A.684 37. The quantity ej is the locale of the error of the j -th neuron in the s layer.1 40. thus ensuring its forward propagation.61 22. 4.8 6. Weights and bias terms are first initialised at random values.3 5 4.92 36. by establishing the input of the first layer of the network.179 21.08 6.6 4. there are no strict rules for determining the network configuration for optimum training and prediction. one can calculate the error function.6 4.6 5.101 38.58 21. When one compares this output with the desired output. This process consists in minimising (e) by a gradient descent.309 37. www.385 20. In general.8 4.2 4.6 4. which usually takes s values between 0 and 0.56 21.971 37. Continued.3 4.416 38.06 40.425 20. Nat.6 21.09 40.5.7 4.66 21.8 4. The direction in which the weights are updated is given by the negative of the gradient of (e) with respect to every element of the weight.2 21. Tselentis and L.9 4.692 21.61 21.148 38.8 5.2 4. generally given by 1 ¯ 2 e = Y −Y (5) 2 ¯ where Y is the desired output. and Y is the obtained output.5 4.362 40.073 20.02 40.66 21. Vladutu: MMI attenuation engineering parameters and neural networks Date (dd-mm-yy) Time Latitude (◦ ) Longitude (◦ ) 21.

log(PGA). This was achieved by comparing the corresponding MMI (that is. was evaluated by an ANN implemented as a k-nearest neighbour (k-NN). 5. A genetic algorithm was used to generate populations of strings out of the 512 possible combinations from 000000000 to 111111111. is a family of learning algorithms that. In our approach. and all computation is deferred until classification.e. we used all nine commonly-used input parameters that characterise earthquake ground motion at a site corresponding to an MMI value. Instance-Based Learning (IBL) (Aha et al.Ia . Sci. Sa . Topology of the feed-forward k-NN-type ANN used in the present study. Ia and CAV (Fig.logR. Nat. (3) preprocessing and optimisation and (4) regression. where the function is only approximated locally. compares new instances with instances that have been observed during training. which deals with classification.. In the machine-learning community.Fig. A direct consequence of this approach is that the complexity of the problem grows with the amount of data available for training and testing.PGA. The total number of available strings (that is.CAV (7) For example. (2)classification. we consider the following four phases: (1) feature extraction. CAV] is represented as the string 010000011. the equivalent of chromosomes) at a certain time (i.Sa . It is called instance-based because it constructs hypotheses directly from the training instances themselves. log(PGV). instead of performing explicit generalisations. The k-nearest neighbours (k-NN) algorithm is a method for classifying objects based on the closest training examples in the feature space. after the Max-NoGenerations). 4). while most were considered for training. We have used a GA-ANN with IBL approach in order to avoid data changes produced by normalisation techniques. During this procedure.1 Fig. At first. also known as memorybased learning. 2527–2537. 2005). 4 5 Data processing We consider automatic MMI assessment based on groundmotion parameters as part of a larger category of problems encountered in pattern recognition (Poggio and Girosi. the outputs) with the selected inputs out of the nine possible inputs represented by the strings generated by the www. 4. S = M.net/10/2527/2010/ The second phase.logPGA. Tselentis and L. namely. The k-NN algorithm is a type of instance24 based learning. the three-parameter string [logR. we use 9-bit string (S) quantification with binary field-values as follows.nat-hazards-earth-syst-sci. we have allowed a population size of 20. In our implementation. 10. Of the data set of 310 values.. Vladutu: MMI attenuation engineering parameters and neural networks 2533 Fig. 2010 . Hazards Earth Syst. Fig. or lazy learning. In the present investigation. 3 G-A. is implemented using a k-NN-type neural network. PGA. Feature extraction 5 genetic algorithm. The Euclidean metric is used to assess distances between the training and testing epochs. These measures include M. which is known as the genome..2 Classification During this phase. 1990. log(R). we used one of the simplest examples of IBL.PGV. 5. 1991). k-NN classifier and its Java implementation based on the Weka-toolbox (Witten and Frank. PGV.logPGV. we combined the enhanced selection capacities offered by GA with the performance of an ANN as a classifier. where the population size indicates the maximum number of chromosomes (or strings) allowed in a generation. some data points were left for testing. Ia .

5. First...2534 G-A.3 Pre-processing and optimisation that can possibly perform better under k-NN classification. Ia . It is analogous to natural reproduction and biological crossover upon which the simplified GA theory for computational intelligence was built. 1995. One-point crossover. the chromosomes). These include the number of hidden layers and the number of neurons in each hidden layer. but for processing purposes. log(PGV). weak individuals have a smaller chance of being selected for breeding. Ia . In our case. a cost function). 2010 . an important parameter for the proposed GAANN method is Max-No-Gener. the error has a higher value). Genetic algorithms are a particular class of evolutionary algorithms (EA) (Fonseca and Fleming. we considered all inputs as they were. and perpetuation in the next generations. and through the EA.g. 10. adjusted by variances in tournament size. there are several parameters that have to be modified (or fine-tuned) to achieve optimal behaviour according to ANN. further binary combinations will be created by the GA on the “skeleton” of these “more optimal” chromosomes. a GA quantifies information (namely. Fig. where S1 is given by the GA+k-NN algorithm and corresponds to the combination [M. In other words. we note the following cases. and a fitness function determines the environment within which these solutions “live” (e. the crossover is the genetic operator used to modify the programming of a chromosome or a group of chromosomes from one generation to the next. An object is an instance out of the 310 data values given in the features space. 2006). In genetic algorithm theory. mutation. a single crossover point on the strings of both parent organisms is selected while avoiding extreme points. It is classified by a majority vote of its neighbours. we determined k and the selection scheme (or the size of the tournament). www. In this scheme. Vladutu: MMI attenuation engineering parameters and neural networks The k-NN classifier is amongst the simplest machinelearning algorithms. This mix in our case can be represented by taking the first four digits from one “optimal” string and the last five digits from another “more optimal” one. the parameters of the k-NN classifier) in the form of strings (that is. That is. A flow chart describing all of the above operations is presented in Fig. If the tournament size is larger (chromosomes for which the objective function. All data beyond that point in either organism string are swapped between the two organism strings. 1995) that use techniques inspired by evolutionary biology such as inheritance. as shown in Fig. Sci. After finding the optimal Max-No-Gener parameter. One advantage of the selection mechanism of a GA (Tobias and Lothar. 1995) is the independence of the representation of the individuals. From this table. we run a tournament among a few individuals chosen at random from the population (i. the objective function is the squared error of the k-NN type of ANN. and crossover (or recombination). 6. The resulting organisms are the children (or offspring). CAV].e. with the object being assigned to the class most common among its k-nearest neighbours.. selection.. crossover. i. In our case.nat-hazards-earth-syst-sci. from the genome) and select the one with the best fitness for crossover as the winner. Supposing that Max-No-Gener = 75 and that we obtained S1 as the fittest string. To validate the performance of the above-mentioned GAANN selection schemes. The fittest chromosome is considered the one for which the objective function has an extreme value. CAV].net/10/2527/2010/ In the previous two phases. Hazards Earth Syst. PGA. it is a five-parameter string consisting of [M. which allows the GA algorithm to evolve to achieve the optimal solution. 2527–2537. Data sets can then be mixed using any of several available techniques.. This simplifies the analysis of the selection methods and allows a comparison that can be used for all kinds of genetic algorithms. log(R). Therefore. Tselentis and L. In our case. Chromosomes which are found to be “more optimal” are allowed to breed. k was found by a trial-and-error procedure to have a value of 2. One of the frequently-used selection schemes is tournament selection. it selects the optimal combination of the nine inputs (see Eq. Mierswa et al. we converted all inputs into integers by multiplying them by powers of 10 (H¨ rdle et a al. 5. Max-No-Gener) are presented in Table 2. Performance was quantified by RMS criteria and the squared error. Candidate solutions to the optimisation problem act like individuals in a population. thus producing a new generation of chromosomes. Only the fitness values of individuals are taken into account. and so the objective function must have a minimum value. thereby creating a new chromosome (9-digit string) Nat. we selected a validation scheme based on regression performance. A fitness function is a particular type of objective function that prescribes the optimality of a solution (or chromosome) in a genetic algorithm so that a particular chromosome may be ranked among all other chromosomes from the genome.. only the fittest chromosomes survive over the generations of the evolution. log(R). 7). 5. This is represented as 110010011.e. The results for the selection of the first optimal parameter from those described above (that is.

Ia . 5. we used the linear Nat. Sa . which is called the validation or testing set. PGA. Max-No-Gener. both the error and its standard deviation are lower than in the case of S1 .Ia .CAV] 2535 Setting Max-No-Gener higher than optimally at 90.CAV] S2 = [M.608 ± 0.608 ± 0. S3 then corresponds to the combination [logR.329 ± 0. PGA.195 0. CAV] and represented by 111111111 is also rejected because it accepted almost all of the input parameters and is a trivial solution. logPGV.Sa .CAV] S3 = [logR.PGA. logR. particularly when one wants to estimate how accurately a predictive model will perform in practice. This is a technique for assessing how the results of a statistical analysis will generalise to an independent data set.194 RMS error 0.231 0. For relatively large datasets. Ia .net/10/2527/2010/ . Flowchart showing the sequences of feature-selection and classification.G-A.196 0. In other words. and CAV] are considered as X multivariate inputs.Ia . Vladutu: MMI attenuation engineering parameters and neural networks Table 2. 2527–2537. 10. log(R). Tselentis and L.PGA. this solution is rejected due to a worse regression performance (that is. logR. logPGA.logPGV. the selected optimal input parameters [M. Cross-validation involves partitioning a portion of the data into complementary subsets.152 0.CAV] S4 = [M. this corresponds to the combination [M. CAV] and is represented as 010000011. which results if values of Max-No-Gener are in the range of 25 to 70.4 Regression Fig.logPGV. PGV. It is mainly used in applications in which the goal is prediction.logR.170 and for S2 is given by of the squared error of 0. Judging from the above.logR.178 0.553 ± 0. 75 80 90 94 Sq error 0.Ia .169 0. 1995).170 0.154. These are the parameters used to express MMI throughout the regression process. PGA.339 ± 0. the optimal selected input parameters are M. and the response variable Y is understood to indicate MMI. 2010 www.311 ± 0.logR. Sci. performing analysis on one subset called the training set. Ia . 6. and validating the analysis on the other subset.Ia . In this case. an optimal value of Max-NoGener = 80 with a minimum RMS was selected. PGA. Ia .194 The retained parameters S1 = [M. it is common to use a crossvalidation or rotation estimation method (Kohavi.456 ± 0. If we consider the case with Max-No-Gener = 80 with S2 as the fittest string.. CAV] and is represented as 110001011. In this situation. The parameters retained by the GA-ANN selection algorithm and corresponding regression performance (columns 2 and 3).618 ± 0. logR. an underdetermined solution. The solution with Max-No-Gener = 94 and string S4 corresponding to the combination [M. S2 is retained as having the best fitness because the output of the k-NN classifier for S1 is given by the squared error 0. Next. Finally. is rejected because GAs require a minimum number of generations to obtain optimal fitness among all available individuals.571 ± 0. In a case in which the existing data are not sufficient for analysis because we must use part of the data for validation and test sets.311±0. To calculate the coefficients that link Y to the five-dimensional X variable.329 ± 0.189 0. a higher cross-validation is usually used (25 in our case). a higher RMS error). Hazards Earth Syst.593 ± 0. Ia .324 ± 0.176 root relative squared error 0.548 ± 0.154 0. and CAV.nat-hazards-earth-syst-sci.

Teor..72548 R2 = 0. 10.462 0. An approach based on evolutionary algorithms can be useful in weighting the importance of those features. Acknowledgements. E.2536 G-A. 2004. ] b[5.1513 0.648 0. and the results of the ANOVA statistical test are shown in Table 4.4024 0. S. ] b[2. We note that not all of the features selected in the GAANN approach have the same influence on MMI attenuation. Tselentis and L..6625 0. The parameters for linear regression obtained by the program XploRe for the input data selected by the proposed GAANN method.. the relation that describes MMI as a function of [M.556 p-value 0..0481 0. The obtained b values are presented in Table 3.417M − 7. Boll. Olafsson.960logR + 0. Nat..9553 12. we can derive a real-time signal-processing system. a new type of neural network (namely. 1995). Vladutu: MMI attenuation engineering parameters and neural networks Table 3.0000 Fig.0000 0. Suhadolc. we presented a new approach based on ANN and GA to model the relationship between MMI attenuation and engineering ground motion parameters. 2010 www.net/10/2527/2010/ . Hazards Earth Syst.. The statistical parameters obtained for the linear regression using the ANOVA test. The performance of this new regression approach has been tested using a Greek strong motion database with satisfactory results. 7.5022 0.3566 Table 4. We are grateful to Laurentin Danciu for his most usefull comments.2908 0.. ] b[1. Edited by: M.4533 0.197 df 5 4 304 309 MSS 31. E.2733 –0.0286 0. Smit.51853 Standard Error = 0. PGA.179 0..52632 Adjusted R 2 = 0. MMI = 8. an evolutionary neural network) can be used to replace the classical k-NN used in the current paper. Output of the regression algorithm (in red) and the original MMI data (in blue) for the 310 considered data points. Kluwer Academic Publishers.. PARAMETERS b[0. 6. Aha. Appl. Sigbjornsson. Machine Learning.. 2005). P. For the GA-ANN selection scheme.67936 6 Conclusions regression approach described by the following equation (H¨ rdle et al.200 –0. B. which can be used for near real-time damage assessment and shake map construction.105Ia − 0.nat-hazards-earth-syst-sci.824 + 0..551CAV In this research.2046 t-test 1. 1991. ] b[4. CAV] is as follows.. Douglas.8236 0. 36–77.7718 –0.: Internet-site for European strong-motion data.1046 –0. and Costa. W.5508 SE 4. Multiple R = 0. 45(3). Sci.5965 StandB 0. 113–129.5177 0. ] b[3.30 296.839 2. a 5 Y = b0 + i=1 bi ∗ Xi (8) where b0 is the intercept. logR. ] Beta 8. Condakis and 2 anonymous reviewers is greately appreciated. Additionally. P.. N. Geofis.4173 –7.1499 0.893 140.437 –0.959 F-test 67. we used a Java-based implementation built around the Weka (the IBk lazy learner) data-mining system (Witten and Frank. Figure 7 shows the time series corresponding to the original data and the results of the regression analysis. Contadakis Reviewed by: two anonymous referees 26 (9) References Ambraseys. R. Accordingly. J. 2527–2537.b5 are the coefficients for the ground parameters. The help of M.4084 0.439 0. Ia .9601 0. ANOVA Regression Residuals Total Variation SS 155. Margaris.380PGA +1.3801 1.: Instance-based Learning.923 p-value 0. and b1 . G. If we implement an expert system. D.

2537 Kostov. R. and Klims.: Duration and energy characteristics of Greek strong motion records.. H¨ rdle. Ann. Geol. 2005. Am. 4. E. 40(5). 978– 982. 1996. 162–183. M. R.: A study of cross-validation and bootstrap for accuracy estimation and model selection. Seismol. Seismol.: Relationships between felt intensity and instrumental ground motion in the central United States and California. I. W. Assoc..: A tool for collecting. Heaton. W.: Data Mining – Practical Machine Learning Tools and Techniques. hypocentral distance and scale based on electric field observation. Environment. Geofis. Springer. 247. 88(6). 1371–1382. Y. 15(3).. Am. K.: YALE: Rapid prototyping for Complex Data Mining Tasks.: Correlation of JMA instrumental seismic intensity with strong motion parameters. I. 79–92. Eng... Soc. V. Marketos... I. L.: Relationship between peak ground acceleration. S.. F. T. in: Proceedings of the first international conference on generic algorithms in engineering systems: Innovations and applications. Y. D. J. (Morgan Kaufmann. 1994. 1998. B. N. L.: A comparison of selection schemes used in evolutionary algorithms. C. 4(4).. Sci. M. XXXVI. www. Cambridge. 1438–1444. and Danciu. peak ground velocity. B. Krinitzsky. Scholz. Quitoriano. Soc.: Empirical relationships between MMI and engineering ground motion parameters in Greece. China. and Kaka. P. G... D. Spectra. A. S. Seismol. Vladutu: MMI attenuation engineering parameters and neural networks Arias. 2003. 45–52. Eng. Poggio. M. Panza..: A measure of earthquake intensity..: Relationship between felt intensity and instrumental ground motion in the central United States and California. 45– 58. and Vaccari. 26. Patent No. 1191–1212. T. Karim. Tselentis and L.. San Mateo). V. H. Earthq. 2005. 2007. 2004. Koliopoulos.. Soc. 18th International Conference on Structural Mechanics in ReactorTechnology. 1970. Comput. B.: An approach to the measurement of the potential structural damage of earthquake ground motions. Geol.. 1137– 1143. 1995.. 2004.. K. Molin.. B. MA. Soc. T. and Atkinson. 2(12). and apparatus for prediction. Hazards Earth Syst. 2002. B. H. Kaka. 98(4). 2010 . Mierswa. D.. Haykin. Fonseca.. Benito.: Neural Networks. UK. 2527–2537. G. T. Margottini. and Frank.: Engineering ground motion parameters attenuation relationships for Greece. 1998. B. 33.: Regularization algorithms for learning that are equivalent to multilayer perceptrons. F. 1988. 1997. and Kanamori. 1999. and Euler. Evol.885. and Chang. Beijing. sharing and mating restriction. M. and R. Struct. 1863–1875.: Intensity versus ground motion: a new approach using Italian data. Klinke. J.: Multi-objective genetic algorithms made easy: Selection. Earthq. 1990. Yamazaki. Klinkenberg. R. Matsumoto. G. G-A.. Heidelberg. C. T. 361–394. S. L. Y. MacMillan College Publishing Company. Oxford University Press. S. M. I. and Westermo. Am. B. Kalogeras. Soc. K. C.. R. I.. Seismol. N. Atkinson. T. Tselentis. Humboldta Universit¨ t zu Berlin Wirtschaftswissenschaftliche Fakult¨ t a a ¨ Institut f¨ r Statistik und Okonometrie. Am. D. Geol. Wu.. 93. Science. Cazzaro. K. Y. Am. M. and Takahashi. S.: Method for estimating origin time. 1977. B. Proceedings of the Fourteenth International Joint Conference on Artificial Intelligence. T.: Seismic Intensity and Fourier Acceleration Spectra: Revised Relationship.. and Turlach. Shin. 391–417. 31(5). and Herraiz. peak ground velocity and intensity in Taiwan. M. US 6. C. 1997. Am. 2nd edn. 161–187. L. Fujinawa.. M. G. and Tselentis. Elsevier. and Fleming. Seismol. Int. Teng.945 B2. uerying. Struct. 2006.. Kohavi. Margaris. H. u H¨ rdle. Sokolov. N. 1992.. 557–564.. 1995. E. R. 497–510. M. V. Wurst. Soc.: A note on the correlation of frequency-dependent duration of strong earthquake ground motion with the modified Mercalli intensity and the geologic conditions at the recording stations.: Relationships between peak ground acceleration. 97(1B). I. Seismol. 2005. 2008..: Site specific estimation of cumulative absolute velocity.: Intensity-related earthquake ground motion. L. P. Tobias. F.. Bull. S. B. G. 1995. F. and Serva. and Hsiao. Y. 2007. Earthq. UK. and Lothar.: Applied Nonparametric Regression. Earthq. Eng.. Spectra. 425–435. D. 1995. Cabanas. B. and mining macroseismic data. and Girosi.. and modified mercalli intensity in California. 18(1).. C. B.. Danciu.: Correlation between macroseismic intensities and seismic ground motion parameters.: Neural Networks for Pattern Recognition. J.: XploRe – an Interactive a Statistical Computing. Soc.nat-hazards-earth-syst-sci.: Spectral parameters of the ground motions in Caucasian seismogenic zones. Greece. Eng. 10.. 917–927. Sheffield.. 2002..net/10/2527/2010/ Nat. Am. 386–396. Seimol. Wald. Trifunac. K. and Theodoridis. Bishop. Soc.-A. M. Witten. 97(2). B. 67(3). (KDD06).G-A. in Proceedings of the 12th ACM SIGKDD International Conference on Data Mining and Knowledge Discovery. Skolov. Earthq. 1999. 97(2). Eng. 497–510. B.

Sign up to vote on this title
UsefulNot useful