Upload
theodore-sims
View
221
Download
0
Embed Size (px)
Citation preview
Címlap
Ernő Keszei Eötvös Loránd UniversityBudapest, HUNGARY
http://keszei.chem.elte.hu/
Efficient model-free deconvolution of measured
femtosecond kinetic datausing a genetic algorithm
OutlineGenetic algorithms: a ”historical” intro
A few words about femtochemical data and convolutionA brief summary of deconvolution methods
Genetic algorithms: how they work in general
Implementation of a genetic algorithm for deconvolution
Examples of the performance: on a simulated data set on an experimental data setConclusions and perspectives
no
idézet2
And God said, Behold, I have given you every herb bearing seed,which is upon the face of all the earth, and every tree,in which is the fruit of a tree yielding seed; to you it shall be for meat.
So God created man in his own image, in the image of God created he him; male and female created he them. And God blessed them, and God said unto them,Be fruitful, and multiply, and replenish the earth, and subdue it:and have dominion over the fish of the sea, and over the fowl of the air, and over every living thing that moveth upon the earth.
(Genezis 1.27-1.29, authorized King James version)
idézet2
So God created man in his own image,in the image of God created he him;Be fruitful, and multiply,
and replenish the earth,
(Genezis 1.27-1.29, authorized King James version)
genalg
C. Darwin: On the Origin of Species, John Murray, London, 1859
J. H. Holland. Adaptation in Natural and Artificial Systems, The University of Michigan Press, Michigan, 1975
.
.
.
.
.
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2008
.
.
.
???? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Femtochemistry
• Aim: time-resolved data on elementary
reactions
• Time-resolution needed : 10–11 -10–14 seconds
10–15 seconds = 1 femtosecond
• problem: electronically accessible time
resolution
not less than 10–9 s (nanosecond)
• Ahmed Zewail (1987)
first time-resolved results
on an elementary reaction (Nobel-prize
1999)
femtochemist
ry
10-10000 fs
CPM lézererősítő
Nd:YAG lézer
Ar - ionlézer
detektor
D2O
minta
Kísérleti berendezés
CPM laseramplifier
pumping laser
drivinglaser
detector
D2O
pump
probe
reference
delay line
Femtosecond pump-probe measurement
sample
0.3 μm = 1 fs
Lézerfotolízis
A– B – C A + BC
ground state
excited state
higher excited state
Pote
nti
al energ
y
A – BC distance
Femtosecond pump-probe measurement
határozatlansági reláció
tetfF ti d)()( 2
d)()( 2 tieFtf
ttftN
t d)(1 222
d)(1 222
FN
dd22
FttfN
0lim 2
tftt
2
1 t
Let f (t) and F () be each others Fourier-transforms in time and frequency domain:
Let us define their ”widths” as their second moments:
N being the 2-norm:
If f is differentiable and , then
Consequences of the uncertainty relation
Visible range: Δ t ~ 100 fs Δ ω ~ 5 nm
Matematikai leírás
S ( )
Im( – t’ )
't
Ig(t) f (t’– t) dt dt ’
0<' if ,0)'(f tttt
f),(corr)( mg IIS n
Detected signal can be written as a convolution:
Maths of the detected femtosecond signal
pump (Ig) probe (Im)
time
(n is the number of exciting photons)
:),(corr mg II n
instrument response function
Torzítás a kinetikában
Distortion of the signal due to convolution
time
kinetic signal
instrument response function
kinetic signal
time
Distortion of the signal due to convolution
measured signal
kinetic signal
time
Distortion of the signal due to convolution
instrument response function
=
Needed: reconstruction of the undistorted object from the image
object spread = image
It can be found as the solution of the integral equation i = o s
)'(to)(ti
)'( tts dt '
or more explicitlyo bject
s pread
i mage
Reformulation using image processing terms
Problem: there exists an infinite number of solutions
Dekonvolúciós eljárások
iterative parameter estimation of the convolved model
• a known model function is needed
• computationally extensive (convolution at each iteration)
• estimated parameters are correlated with IRF parameters
• simple algorithms
• short computation time
• examples: Van Cittert iteration
inverse filtering
• complicated algorithms
• long computation time
• easily adapted as ”ad hoc ” methods to a given problem
Linear methods Nonlinear methods
Most widely used: reconvolution
Model-free deconvolution methods
Methods of deconvolution
|F()|am
plitúdó
frekvencia
f(t)
amplitúdó
csatorna
|F()|am
plitúdó
frekvencia
f(t)
amplitúdó
csatorna
Fourier-transform of a continuous function:
tdtfeF ti )()(
Discrete Fourier-transform:
1
0
2)()(N
NmnienfmF
Fourier-transzformáció
Fourier-transformation
time, t frequency, ω
am
plit
ude
am
plit
ude
dOeto ti
)(2
1)(
The undistorted object o can be computed (in principle) by a simple inverse Fourier-transformation:
Convolution in frequency space:
I () = S () · O ()
Deconvolution in frequency space:
O () = S ()I ()
Inverz szűrés
Inverse filtering
”filtering”
”inverse filtering”
Deconvolution by inverse filtering
0 25 50 75 1000.00.20.40.60.8
500 750 1000 1250 1500
0.0
1.0
2.0
3.0
4.0
5.0
ampl
itude
channelAmplitude spectrumof the filtered deconvolved signal
In addition to inverse filtering,a smoothing filter is also usedto damp high frequenciesin order to filter out noise
deconvolved
0 25 50 75 1000.00.20.40.60.8
500 750 1000 1250 1500
0.0
1.0
2.0
3.0
4.0
5.0
ampl
itude
undistorted signal
Deconvolution by inverse filtering
channelAmplitude spectrumof the filtered deconvolved signal
deconvolved
In addition to inverse filtering,a smoothing filter is also usedto damp high frequenciesin order to filter out noise
Iterációs módszerek
Iteration methods
o (i +1) = o
(i) (x) + [i(x) – s(x) o (i) (x)]
is a suitable function to ensure convergence
If is a constant: linear iterative deconvolution
If is a function of x : nonlinear iterative deconvolution
is called the relaxation function
500 750 1000 1250 1500
0
1
2
3
4
5
am
plit
údó
csatorna
Bayes: 4. lépésDeconvolution by (Bayesian) iteration
step4.
deconvolved
image
am
plit
ud
e
channel
500 750 1000 1250 1500
0
1
2
3
4
5
am
plit
údó
csatorna
Bayes: 16. lépésDeconvolution by (Bayesian) iteration
step16.
deconvolved
image
am
plit
ud
e
channel
500 750 1000 1250 1500
0
1
2
3
4
5
am
plit
údó
csatorna
Bayes: 128. lépésDeconvolution by (Bayesian) iteration
step128.
deconvolved
image
am
plit
ud
e
channel
500 750 1000 1250 1500
0
1
2
3
4
5
am
plit
údó
csatorna
Bayes: 512. lépésDeconvolution by (Bayesian) iteration
step512.
deconvolved
image
am
plit
ud
e
channel
500 750 1000 1250 1500
0
1
2
3
4
5
am
plit
údó
csatorna
Bayes: 1883. lépésDeconvolution by (Bayesian) iteration
step1883.
deconvolved
undistorted signal
am
plit
ud
e
channel
genetikus algoritmusok
Genetic algorithms (”eugenics”)
create an initial population
measure the fitness of each individual
select individuals to reproduce (parents)
let parents mate (crossover)
perform mutation on each offspring
select individuals of the new generation
repeat production of new generations (evolution) until you find an individual with the expected features
result: individual(s) with optimal features
production
of a
new
generation
Creation of the initial population („genesis”)
The initial population should be made via inversion of the above distortion effects
convolution makeswiden the signal temporally,diminish its amplitude,shallow its rise and descent,smooth out steplike jumps
Creation of the initial population („genesis”)
From the experiment, the image i (and the spread s ) is known
Creation of the initial population („genesis”)
To reconstruct the object o :
compress the image temporally,
From the experiment, the image i (and the spread s ) is known
Creation of the initial population („genesis”)
increase its amplitude,
To reconstruct the object o :
compress the image temporally,
From the experiment, the image i (and the spread s ) is known
Creation of the initial population („genesis”)
increase the steepness of its rise and decay,increase its amplitude,
To reconstruct the object o :
compress the image temporally,
From the experiment, the image i (and the spread s ) is known
Creation of the initial population („genesis”)
restitute the stepwise jump by ”cutting” the first few data
To reconstruct the object o :
From the experiment, the image i (and the spread s ) is known
increase the steepness of its rise and decay,increase its amplitude,compress the image temporally,
Creation of the initial population („genesis”)
random factors are used in all the operations for the
compression ratio,amplitude increase,steepness increase of the rise and decaylocation of the initial cut
The resulting initial population is made of different ”individuals”:
Reproduction of the population (”evolution”)
1. computation of the suitability (fitness) of individuals to be a proper object function:
large fitness = small difference between reconvolved individual and image (measured by the sum of squared differences)
2. selection of 2 parents with a probability proportional to their fitness
3. crossover of selected parents results in a would-be offspring (simple average or fitness-weighted average of parents)
4. mutation of the would-be offspring, to get an individual of the new generation
5. after sufficient new individuals, select the new generation (”elitism”: if the most fit parent(s) are also selected)
To get another new generation, repetition of 1-5. is performed, until a satisfactory deconvolved will be found.
Stopping: MSE error, Durbin-Watson statistics, No. of generations
Balancing creation and evolution
a carefully generated initial population is usually quite close to a suitable deconvolved – a fairly good estimate of the object
To get the right initial population, well-chosen parameters(compression, amplitude increase, steepness enhancement, initial cut) are needed – but random parameter variation is also necessary !
during reproduction of the population, randomness is also important (selection of parents, mutation), but mutation is a key element determining the quality of solution !
- too large mutations lead to noisy deconvolved data set - too small mutations result in a wavy deconvolved data set
a „smooth” correction in a larger interval avoids both noisy and wavy behavior
(actual implementation: correction by adding a random Gaussian)
Applied genetic algorithm in technical terms
Data structure: a chromosome is the deconvolved data set (coded genes are floating point numbers - ∞ alleles)
Individuals: single-chromosome haploid gene-sequence; no phenotype
Fitness: a scaled inverse of the sum of squared differences between the image and the reconvolved individual
Parent selection: fitness-proportional probability, roulette-wheel (natural selection, not breeding)
Crossover: arithmetic; non-weighted average or fitness-weighted average of 2 parents
Mutation: changes neighbouring genes in a given interval by adding a smooth random function
Selection of the new generation: one-parent elitism offsprings make the new generation, except for the fittest parent
eredmény ek1
Deconvolution of synthetic data
eredmény ek1
Deconvolution of synthetic data
eredmény ek1
Deconvolution of synthetic data
50 100 150 200
0,0
0,5
1,0object
dots: residuals
winneram
plit
ud
e
channel
reconvolvedimage
eredmények2
Deconvolution of synthetic data
50 100 150 200
1E-4
1E-3
0,01
0,1
1 object
winner
am
plit
ud
e
channel
reconvolved
image
eredmények2
Deconvolution of synthetic data
10 20 30 40 50 60 70 80 90 1001
10
object
winner
spec
tral
am
plitu
de
channel
reconvolved
image
eredmény3
20 40 60 80
0,0
0,5
1,0
1,5
2,0
2,5
dots: residuals
winnerampl
itude
channel
reconvolved
image
Deconvolution of experimental data
fluorescence of adenosine monophosphatein waterupconversion detectionexcited at 267 nmobserved at 310 nmBányász & Gustavsson
eredmény4
Deconvolution of experimental data
20 40 60 80
1E-4
1E-3
0,01
0,1
1winner
ampl
itude
channel
reconvolved
image
eredmény4
Deconvolution of experimental data
10 20 30
1
10
winner
spec
tral
am
plitu
de
channel
reconvolved
image
Conclusions
Genetic algorithms are suitable deconvolution methods
They can be well adapted to deconvolve femtochemical data (or transient responses in general)
Deconvolved data sets do not contain neither enhanced noise nor extra low-frequency oscillations
The entire frequency range of the undistorted signal can be reconstructed
The method performs excellently on experimental data
There are good perspectives to develop a largely automated version with an easy-to-use Graphical User Interface
Moral: 1. it is worth reading even the oldest literature 2. both creation and evolution have their place in science
Acknowledgement
Ákos Bányász & Thomas Gustavsson CNRS Saclay (experimental data)
Péter Pataki, grad. student in mathematicsEötvös Loránd University Budapest(parts of the Matlab code)
€ € € €............
Hungarian National Research Fund (OTKA)
Balaton / TéT bilateral exchange program (France-Hungary)
R & D Ulrafast Lasers Kft. (Róbert Szipőcs)
vége
eredmény3
Smoothing effect – synthetic data
100 150 200
0,0
0,1
0,2
object
dots: residuals
winnera
mp
litu
de
channel
reconvolved
image
eredmény4
Smoothing effect – synthetic data
10 20 30 40 50 60 70 80 90 1001E-4
1E-3
0,01
0,1
1
10
object
winner
spec
tral
am
plitu
de
channel
reconvolved
image
eredmény3
Effect of mutations
MSE: 0.06 DW: 0.07
2 generations
MSE: 0.001 DW: 1.93
2000 generations