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

Global optimization algorithm:Search

for the Lost Valley (SLV)


Manuel Abarca

Urubamba, Perú
Email: manuel.z.abarca@outlook.es

A new global search algorithm is proposed here. This technique to find a


minimum (or maximum) of an objective function begins with a population of
models randomly generated . The global minimum of this function must be at
some point of the variables space. Search for that minimum uses of an arithmetic
mean (centroid) between neighbourhoods models. There is also a similar tactic to
bring outside of a local minimum. The new algorithm is tested with a seismological
inversion problem, modelling the Earth through receiver function. First with
synthetic data and finally with real data of receiver functions.

Keywords: computation of global search; optimization; geophysical inversion; seismology;


receiver function
Received ; revised

1. INTRODUCTION near the global minimum. In the first case we have


a tactic to bring outside the local minimum, will be
The global search algorithms are computational tools described in the next section. In the second case we
used in many fields of science and industry; the goal is have a mechanism to track in the vicinity of that point
to solve a problem through the minimization (or maxi- as if it were located in a valley, until to reach the bottom
mization) of an objective function (or error function). of the valley (in the hope that is the global minimum).
The proposal of an algorithm to solve an also known as The name of our algorithm explains this search for an
optimization problem means we know two things: one, unknown valley and the tracking by their surface until
the way to sampling the variables space and second, an to find the global minimum, Search for the Lost Valley
search mechanism. The focus of our research is on the (SLV).
development of a new search mechanism.
There are some kind of search mechanisms based in 2. METHOD
the gradient analysis (steepest descent [1];conjugate
gradient [2]). Another search mechanisms are purely First than all, some definitions here. We call ”model”
random or controlled random (Monte Carlo [5] , CRS to the group of variables able to describe a physical
[7]). The process by which molten metal crystallizes situation. For example the variables x1 , x2 , ..., xp can
and reaches equilibrium in the process of annealing represent seismic velocities and thicknesses of layers
can, for instance, gives place to a method of optimiza- in a sedimentary basin. One model explaining the
tion called Simulated Annealing [4]. A new criteria geophysical relationship between strata in the basin
in optimization were introduced by evolutionary algo- could be m ⃗ 1 (x11 , x12 , ..., x1p ); another model would be
rithms (Genetic Algorithm [3] ) which uses operations m⃗ 2 (x1 , x2 , ..., x2p ); where p is the number of variables in
2 2

as cross-over and mutation to find the best fitted indi- the model.
vidual. We call ”forward problem” to the physical-
mathematical equations which relates a ”model”
Our algorithm implements a search strategy similar with an ”measurable” response in some point of the
to the Simplex [6] but combined with a tracking in the space; in a general form is a f (m). ⃗ This function of
neighbourhood. The main difference respect to Simplex the model has unique solution, so, for each model m ⃗i
is that this uses the point of centroid over all population there is one and only one set of observed responses in
in variables space. While our logical reasoning is to find the space f (m ⃗ i ).
the centroid around a (transitory) minimum point plus The ”real” data or ”observed” data d⃗o are the mea-
two points in their neighbourhood. The idea below this sures of some physical magnitude in points of the space
search mechanism is that one point of minimum in a (generally on surface earth).
surface representing an error function could mean one We call ”inverse problem” to the mathematical pro-
of this things: it is a local minimum or it is a point cedure to find the model which best fit the observed
2 M. Abarca

data. This is made minimizing an objective (or error) If there is a lost valley in any region of the variables
function, space it will be possible to find with this mechanism.
2
We present the same algorithm in a flux diagram in
τ (m) ⃗ − d⃗o
⃗ = f (m) (1)
2 the next figure,
for instance we are using here the least square norm.

The inverse procedure is generally dealing with ill


posed problems. This is because the noise (natural,
instrumental, anthropogenic) included in the real
data. Then to stabilize the solution and to get
uniqueness in the inverse problem we have to apply
some regularization criteria. We choose the Tijonov
criteria ([9]), which implies to add a-priori information
to the objective function,
2
2
⃗ = f (m)
τ (m) ⃗ − d⃗o + λ2 ∥W · m∥
⃗ 2 (2)
2

d⃗o : real data;


W : regularization function;
λ : equalization factor .

Another measure necessary to our search strategy is


the ”distance” between models,
q
s̄ = (x11 − x1 )2 + (x12 − x2 )2 + ... + (x1p − xp )2 (3)

Now, we want to minimize the objective function τ


applying the SLV algorithm:
1. Create a random population of models,
m⃗1, m⃗ 2 , ..., m⃗n , with n number of elements; insert
a threshold value e.
2. Evaluate (2) for each model.
3. Sort the population of models by the magnitude
of error function. The lowest error is the best model,
until this step.
4. Measure the distance between best model and all
elements of population, (3).
5. Sort models by distance to the best model.
6. Choose 3 models with shortest distance; this group
includes the best (distance zero).
7. Take the centroid from 3 elements of group.
8. Evaluate (2) for the centroid.
8.1 Compare objective function of centroid with thresh-
old value; if τ is less to e the process ends. Another FIGURE 1. SLV algorithm for global search in
optimization problems. Begin (inicio), end (fin).
stop criteria could be the number of iterations, defined
by n.
9. If the objective function of centroid is better than
2.1. An improvement to search mechanism
τ of best model, then centroid takes the place of best
model; go to step 4. When we evaluate the centroid over 3 models, in
10. Else, if the error of centroid is not better than best mathematical terms we are taking the arithmetic mean,
model, then takes the next 2 models by distance and
m
⃗ c = (m
⃗1 + m
⃗2 + m
⃗ 3 )/3 (4)
the best one.
11. Go to step 7. so, we can refine our tracking by the surface of the
valley putting different weights to each one of this 3
The tactic to bring outside a local minimum is given models.
in step 10. This search mechanism is not totally ran-
dom because we are using the best model as pivot. ⃗ c = (k1 · m
m ⃗ 1 + k2 · m
⃗ 2 + k3 · m
⃗ 3 )/6 (5)
3

[hhh]
where,
k1 = 1; k2 = 2; k3 = 3 , for the first weighted mean; TABLE 1. Seismic model of a sedimentary basin.
k1 = 3; k2 = 1; k3 = 2 , for the second weighted mean; Depht m Vp m/s
k1 = 2; k2 = 3; k3 = 1 , for the third weighted mean. 1500 1800
Then, inside the step 8 of SLV we choose the centroid 4000 2500
with the lowest error function. 90000 5000

3. INVERSE PROBLEM IN GEOPHYSICS


value because it is representing a semi-inifinite medium.
We can take a geophysical example to show how to
apply the SLV algorithm in solution of inverse problems.
Theoretical response of a seismic P-wave crossing this
In a sedimentary basin the layers would be characterized
pack o layers is given by a RF (Fig. 2)
by parameters as seismic velocity, density of rocks
and thicknesses of strata. Knowing some physical-
mathematical formulae we are able to calculate the
travel time of a seismic wave, from the bottom of
sediments to top (air-earth interface). This is the
forward problem in geophysics,

T (m(x
⃗ j , xj+1 , ..., xp−1 , xp )) (6)
where,
xj : seismic velocity in the first layer ;
xj+1 : depth of first layer;
xp−1 : seismic velocity of semi-infinite medium;
xp : depth trending to infinite (90000 m in table 1
means semi-infinite);
T : travel time.
FIGURE 2. Synthetic Receiver Function (RF), obtained
But, in a real seismological investigation we got time from the model of table 1.
series (seismograms) registered by instruments (called
seismometers) where we can read arrival times of seis- The test consists in applying our SLV algorithm to
mic waves. In Receiver Function method (RF) we have inverse problem; in that way we have to recovery the
not just the travel time of one kind of seismic wave, but same model (or approximately the same) given in table
also the times of converted waves (P to S) and rever- 1, from inversion of ”observed” RF (Fig. 2). Our RF
berated waves; further we have the complete waveform gets a more realistic shape including 5% of noise.
of RF. In any case the relevant parameter is time,
t0 observed in a seismogram or in a pseudo-seismogram. Minimization of( 7) will produce a best fitted RF
associated to a seismic model which must be similar
We don’t describe the RF method in their seis- to the proposed in the premises of our test,
mological intricacies for one reason: this is not a
seismological study. This a research with focus in the
area of optimization, or could be also in geophysical
inversion techniques, so our proposal is to make known
a new method in global search (or maybe in non linear 1.0
1000

optimization). 2000

2.0 3000
The final subject of an inversion is to get a
realistic model of sedimentary basin from RF pseudo-
3.0
seismogram. This implies to minimize an objective 5000

function,
4.0

2 2
⃗ = ∥T (m)
τ (m) ⃗ − t0 ∥ 2 + λ ∥W ·
2
m∥
⃗ 2 (7)
5.0

4. TESTING SLV WITH A SYNTHETIC RF


It is a theoretical model, with variables xj represent- FIGURE 3. ”Real” waveform of RF and best fitted RF
ing seismic velocity of P-wave and depth of layers (at the left); seismic model of a sedimentary basin result of
into a sedimentary basin , Last depth has a very large inversion (right side of figure).
4 M. Abarca

Considering the difficulties of RF inversion, results of


the test are satisfactory.
inicio

5. EVALUATION OF GOODNESS OF SLV


ALGORITHM
final

SLV algorithm begins generating a random population


of models; the ideal number of models in this population
is in the range 50-60. Test of sensitivity shows that
an initial population too great no correlates with an
increasing in fitness of final result.

FIGURE 5. Best model from initial population in green


square (inicio); final solution for inversion in blue square;
path of search in orange lines; the nodes of path are points
of centroid, this explains why there are not nodes in the
borders of surface of objective function.

6. TESTING SLV WITH A REAL GEOPHYS-


ICAL PROBLEM
The waveform of a RF is the subject of present
evaluation of SLV in their capacity to solve inverse
problems. This RF (Fig.6)was obtained in a study
developed over sedimentary basin of Parana river ([11]).

0,4
FR ajustada
FR observada

FIGURE 4. Surface of objective function from 512 models 0,3

randomly created. 0,2


FR amplitud

0,1

Creation of random models open the possibility to -0,1

put an interval for each variable. This is a kind of soft


regularization.
-0,2
0 1 2 3 4 5
Tiempo - seg.

The first best model (from initial random population)


is signed with a green square. In figure (Fig. 5) the path FIGURE 6. Best fitted RF in black line; observed RF in
of search by new local minima is indicated with orange green dotted line.
straight lines. When SLV found a general minima marks
this with a clear blue square. Note that the area around
the blue square is not revealed as a valley in the initial After running our program which realize the SLV
population (Fig. 4). algorithm we got inversion of RF, obtaining following
0
5
-1000
a surficial volcanic rocks cover; this basalts layer has
-2000
higher seismic velocity than sediments below. Both
-3000 algorithms were capable to define the high velocity
Profundidad - m

-4000 upper layer in the range 4250-4500 m/s for P-wave.


-5000
GA uses more layers to fit the RF, this is the reason
by which we can see a thin low velocity layer at the
-6000
top of basin; but SLV (using fewer layers) can fit the
-7000 waveform with an average-velocity layer in surface.
-8000

-9000
7. CONCLUSIONS
2000 2500 3000 3500 4000 4500 5000
Velocidad de P - m/s
SLV algorithm meets two important proposals of all
global optimization strategy,. First one. is capable of
an uniform and complete sampling of variables space;
FIGURE 7. Seismic model of sedimentary basin in Parana this quality is given by a good random numbers gen-
(Brazil), obtained from inversion of RF. erator. And the second one, has a search mechanism
Dois RF inversion - capb stn.
able to track the variables space until to find any local
0
minimum, and probably the general minima. Looking
-1
Modelo sismico in the neighbourhood of a transitory best model with 4
centroids increase probability to get a new best model.
-2 It is a notable feature of algorithm the pivoting on each
new best model, re-beginning the search for a new min-
Profundidade (km)

-3
ima, or jumping outside this local minimum but track-
-4 ing again all the space.
The new global search algorithm SLV passes satisfacto-
-5
rily two tests. The first test with synthetic RF, and the
-6 second test inverting a real RF.
-7
2,5 3 3,5 4 4,5 5 5,5 6
Vp (km/s) DATA AVAILABILITY
The data underlying this article will be shared on
reasonable request to the corresponding author.
FIGURE 8. Result of RF inversion with GA in Parana
basin ([10]), using same seismic station and the same
waveform of RF used in this section. ACKNOWLEDGEMENTS
I have to thank to mi sister Martha by her financial
support during research epoch.
results (Fig. 7).
In the doubt smoothes, says some geophysicist who’s
REFERENCES
name I don’t remember. So, we apply smoothness
in the regularization function of our τ . This type [1] R. Fletcher, M. J. D. Powell (1963) A Rapidly
of regularization try to oblige the model to minimize Convergent Descent Method for Minimization. The
differences between velocities of layers; but abrupt Computer Journal, 6-2, 163–168.
changes in velocity can appear if the data indicates [2] R. Fletcher, C. M. Reeves (1964) Function minimiza-
necessity of that jumps in velocity. This is a good proof tion by conjugate gradients. The Computer Journal,
for our algorithm and the model in following figure is 7-2, 149–154.
signalling that strong differences in velocities are real. [3] J.H. Holland (1975) Adaptation in Natural and
Artificial Systems. University of Michigan Press, Ann
In this case we have not a previously well known model
Arbor, Michigan, re-issued by MIT Press.
to compare with the result. However we have made an
[4] Scott Kirkpatrick, C. D. Gelatt, Jr., and M. P. Vecchi
study some years ago with the same RF . In that old
(1983) Optimization by simulated annealing. Science,
study we used a Genetic Algorithm (GA) in inversion of 220(4598), 671–680.
RF, obtaining the model exhibited in Fig.8. Comparing [5] Nicholas Metropolis, S. Ulam (1949) The Monte
the resulting seismic models for Parana basin, obtained Carlo Method. Journal of the American Statistical
with two different algorithms, SLV Fig. 7 and GA Association, 44, No. 247, 335–341.
Fig. 8, we can see many similarities. One of the [6] Nelder, John A.; R. Mead (1965) A simplex method for
main goals of that old study was to determine depth minimization. The ComputerJournal , 7, 308–313.
of sedimentary basin (until basement). In both models [7] W. L. PRICE (1983) Global Optimization by Con-
the basement depth is located near 3500 m. A second trolled Random Search. Journal of optimization theory
feature distinctive of Parana basin is the existence of and applications, 40, 333–348.
6 M. Abarca

[8] Tarantola, A. (2005) Inverse problem theory and


methods for model parameter estimation. Society for
Industrial and Applied Mathematics, Philadelphia.
[9] Tijonov,Andrei Nikolayevich (1963) Solution of in-
correctly formulated problems and the regularization
method. Soviet Mathematics , 4, 1035–1038.
[10] Zevallos Abarca, Ivan (2004) Modelamento da Bacia
do Parana - reservatorio Capivara - atraves da inversao
conjunta de Funcao do Receptor e de sondagem
Magnetotelurica. Universidade de Sao Paulo, Sao
Paulo.
[11] Iván Zevallos, Marcelo Assumpc̃ão, A. L. Padilha.
(2009) Inversion of teleseismic receiver function and
magnetotelluric sounding to determine basement depth
in the Paraná basin, SE Brazil. Journal of Applied
Geophysics , 68, 231–242.

You might also like