Introduction of Genetic

Algorithms for geophysical

applications

Gaspar Monsalve
 Genetic Algorithms (GA)
9 Strategy to fit a model to a data set using
forward modeling.
9 It uses the principles of genetics / evolution,
such as natural selection, in order to find the
most appropriate model for the set of data.
9 The selection is based on some measurement of
data misfit.
9 It facilitates fast convergence and prevents the
user from purely random tests.
Advantages of Genetic Algorithms

9 They efficiently search the model space, so they

are more likely (than local optimization
techniques) to converge toward a global minima.
9 There is no need of linearization of the problem.
9 There is no need to compute partial derivatives.
9 More probable models are sampled more
frequently than less probable ones.
 Basic operations in the genetic
algorithms
9 Given a data set, the evolution starts with a
population of randomly selected individuals
(models). The initial population should cover the
entire range of possible solutions.
9 Subsequent generations are constructed by the
action of three reproduction operators:
- Selection (based on misfit)
- Cross-over (swapping information)
- Mutation (Changes in individual
genes)
 Selection
9 A filter to select the model that best fit the
data is commonly used.
9 After defining a probability of reproduction
(which can be calculated as a function of
misfit), the strongest (fittest) individuals are
selected.
9 A good balance should be found in the choice
of the “amount of pressure” of the selection
algorithm. “High pressure”: fast convergence
and risk of a local minimum. “Low pressure”:
Very slow convergence.
 Reproduction
9 Pairs of “parent” solutions are selected to generate
“child” solutions with many of the characteristics of
their “parents”.
9 Since only the best individuals are selected for
breeding, the average fitness of the second
generation will increase.
9 Reproduction includes cross over (high probability)
and mutation (low probability).
9 There is usually one offspring per “parent”, so that
the population size remains constant.
9 Sometimes individuals can be ranked according to
their fitness; the number of offspring will be
proportional to the rank position.
 Cross-over
9 Exchange (swapping) of genes between “parent” models
in order to produce “child” models (next generation).
9 Controls the degree of mixing and sharing of information.
9 It is performed only by those that pass the fitness test.
9 Cross-over favors stronger models and ensures that
diversity is kept.
9 “Parents” stronger than their offsprings can be included
in the next generation instead of the “child”.
 Models have string
Cross-over methods representations (binary,
decimal, etc..)

Single point cross-over Two point cross-over

Cut and splice cross-over Uniform cross-over

Individual bits in the string are
compared between the two
parents. The bits are
swapped with a fixed
probability
 Mutation

9 Making random changes in genes.

9 It is made in order to maintain some
degree of randomness in the population
(helps to avoid local minima).
9 The probability of mutation should be kept
low in order to prevent excess of
randomness.
 Sequence of steps

1. Input: A generation of individuals.

2. Calculation of misfit and selection.
3. Reproduction (cross-over + mutation) =
New population
4. Go to 1. with a new generation
 When to stop? (termination criteria)

9 Misfit (a minimum criteria)

9 Number of generations (maximum
reached)
9 Allocated budget reached (time, money,
etc.)
9 Solution reached a plateau
9 Combinations of previous criteria
Disadvantages of genetic algorithms
9 They show a very fast initial convergence, followed
by progressive slower improvements (sometimes is
good to combine it with a local optimization method).
9 In presence of lots of noise, convergence is difficult
and the local optimization technique might be
useless.
9 Models with many parameters are computationally
expensive.
9 Sometimes not particularly good models are better
than the rest of the population and cause premature
convergence to local minima.
9 The fitness of all the models may be similar, so
convergence is slow
 Examples: Hypocenter location
(Billings, Kennett and Sambridge, GJI, 1994)
9 Population of hypocenters is generated randomly.
9 Subsequent generations are constructed by the
action of the reproduction operators with 90%
probability of cross-over and 4% probability of
mutation. Local
minimum
9 Each earthquake is
located 1000
times.They expect the Frequency
locations to be tightly
clustered about a
single region in the
parameter space, and
with similar misfits.
Misfit
 Examples: Hypocenter location
(Billings, Kennett and Sambridge, GJI, 1994)

9 Parameter correlations in the hypocenter

location problem:
- Strong correlation between depth and origin time.
- Weak correlation between epicenter and depth.
9 The genetic algorithm search is done in latitude-
longitude.
9 After finding an epicenter, they fix lat.-lon., and
for each depth they calculate the optimum origin
time.
 Examples: Velocity structure using
waveform fitting
(Bhattacharyya et al.,BSSA, 1999, Sambridge and
Drijkoningen, 1992, GJI)

9 They use 1-D velocity models and a

known seismic source, so the number of
parameters in each model can be
reasonably small.
9 Fitness (the cost function) is a function
of the cross-correlation coefficient.
9 Tests with synthetic data are useful in
order to check for convergence.
 Examples: Velocity structure using
waveform fitting
(Bhattacharyya et al.,BSSA, 1999)

Test with synthetic data Results with real data

 Examples: Seismic Refraction data
(Boschetti et al., 1996, Geophysics, Jin and Madariaga,
1993, G.R.L.)

9 They want to solve for velocity structure in a grid

of nodes.
9 They produce synthetic data: waveforms and/or
arrival times and/or migrated data.
9 Population: P-wave velocity values in a grid of
nodes. Each member of the population can be
represented in binary or decimal form.
9 High dimensionality might be a problem:
Subdivisions of the model space might be
necessary (pseudospace methods).
Examples: Seismic Refraction data
(Boschetti et al., 1996, Geophysics)

Synthetic data for slowness

Examples: Seismic Refraction data
(Boschetti et al., 1996, Geophysics)

Results of the regular Genetic Algorithm

Examples: Seismic Refraction data
(Boschetti et al., 1996, Geophysics)

Genetic algorithm + pseudospace method

 Examples: Seismic Refraction data
(Boschetti et al., 1996, Geophysics)

100 generations of GA + 150 generations of GA +

local search local search
 Conclusion

„ The main advantage of the GA is the

ability to converge to a global minimum.
„ It is useful for geophysical applications
when combined with local optimization
methods.