Simulating Random. Multi-Phase Media. P.S. Koutsourelakis. Institute of Engineering Mechanics. University of Innsbruck. Baltimore 2005 â p.1/23 ...
A general framework for Simulating Random Multi-Phase Media P.S. Koutsourelakis Institute of Engineering Mechanics University of Innsbruck
Baltimore 2005 – p.1/23
Motivation
Soil Profile
Multiphase Fire-resistant
Concrete
Baltimore 2005 – p.2/23
Motivation
Soil Profile
Multiphase Fire-resistant
Concrete
Variability - Uncertainties in Various Length Scales •
Appropriate Probabilistic Model
•
Incorporate them in the analysis
Baltimore 2005 – p.2/23
Modeling Multi-Phase Media Random Field Approach
Baltimore 2005 – p.3/23
Modeling Multi-Phase Media Random Field Approach 1 Binary Fields: I (j) (x) = 0
if x ∈ phase j otherwise
Baltimore 2005 – p.3/23
Modeling Multi-Phase Media Random Field Approach 1 Binary Fields: I (j) (x) = 0 X
if x ∈ phase j otherwise
I (j) (x) = 1 ∀x
j
Baltimore 2005 – p.3/23
Modeling Multi-Phase Media Random Field Approach 1 Binary Fields: I (j) (x) = 0 X
if x ∈ phase j otherwise
I (j) (x) = 1 ∀x
j
Discrete-Valued Fields: I(x) = j
if x ∈ phase j
Baltimore 2005 – p.3/23
Modeling Multi-Phase Media Random Field Approach 1 Binary Fields: I (j) (x) = 0 X
if x ∈ phase j otherwise
I (j) (x) = 1 ∀x
j
• •
Discrete-Valued Fields: I(x) = j I (j) (x) and I(x) are random functions.
if x ∈ phase j
Their probabilistic characteristics are related to the morphological uncertainties. Baltimore 2005 – p.3/23
Incorporation of Uncertainties Monte Carlo Simulations
Baltimore 2005 – p.4/23
Incorporation of Uncertainties Monte Carlo Simulations
•
Generate samples of the medium (i.e. I(x))
•
Solve deterministically for each sample
•
Process statistics of the results
Baltimore 2005 – p.4/23
Incorporation of Uncertainties Monte Carlo Simulations
•
Generate samples of the medium (i.e. I(x))
•
Solve deterministically for each sample
•
Process statistics of the results
Baltimore 2005 – p.4/23
Problem Formulation Given a number of samples of the medium {Ik (x)}K k=1 we can extract statistics:
Baltimore 2005 – p.5/23
Problem Formulation Given a number of samples of the medium {Ik (x)}K k=1 we can extract statistics: 1st order - volume fraction of phase j : E[I (j) (x)] ≈
1 K
P
(j) k Ik (x).
Baltimore 2005 – p.5/23
Problem Formulation Given a number of samples of the medium {Ik (x)}K k=1 we can extract statistics: 1st order - volume fraction of phase j : E[I (j) (x)] ≈
1 K
2nd order -(cross)-correlations : E[I (j) (x1 )I (r) (x2 )] ≈
P 1 K
(j) k Ik (x).
P
(j) (r) k Ik (x1 )Ik (x2 ).
Baltimore 2005 – p.5/23
Problem Formulation Given a number of samples of the medium {Ik (x)}K k=1 we can extract statistics: 1st order - volume fraction of phase j : E[I (j) (x)] ≈
1 K
2nd order -(cross)-correlations : E[I (j) (x1 )I (r) (x2 )] ≈
P 1 K
(j) k Ik (x).
P
(j) (r) k Ik (x1 )Ik (x2 ).
Higher order - lineal path functions, statistics of Fourier or wavelet transforms of I k .
Baltimore 2005 – p.5/23
Problem Formulation Given a number of samples of the medium {Ik (x)}K k=1 we can extract statistics: 1st order - volume fraction of phase j : E[I (j) (x)] ≈
1 K
2nd order -(cross)-correlations : E[I (j) (x1 )I (r) (x2 )] ≈
P 1 K
(j) k Ik (x).
P
(j) (r) k Ik (x1 )Ik (x2 ).
Higher order - lineal path functions, statistics of Fourier or wavelet transforms of I k .
In general any function g(I) : E[g(I(x))] ≈
1 K
P
k
gm (Ik (x)).
Baltimore 2005 – p.5/23
Problem Formulation Given a number of samples of the medium {Ik (x)}K k=1 we can extract statistics: 1st order - volume fraction of phase j : E[I (j) (x)] ≈
1 K
2nd order -(cross)-correlations : E[I (j) (x1 )I (r) (x2 )] ≈
P 1 K
(j) k Ik (x).
P
(j) (r) k Ik (x1 )Ik (x2 ).
Higher order - lineal path functions, statistics of Fourier or wavelet transforms of I k .
In general any function g(I) : E[g(I(x))] ≈
1 K
P
k
gm (Ik (x)).
Goal: Digitally generate sample functions of the random field I(x) based on given probabilistic information (i.e. volume fractions, correlation functions etc).
Baltimore 2005 – p.5/23
Problem Formulation Let I denote the values of I(x) at a discrete domain i.e. an N × N lattice.
Baltimore 2005 – p.6/23
Problem Formulation Let I denote the values of I(x) at a discrete domain i.e. an N × N lattice. Consider M features-functions gm for which: E[gm (I)] = µm
Baltimore 2005 – p.6/23
Problem Formulation Let I denote the values of I(x) at a discrete domain i.e. an N × N lattice. Consider M features-functions gm for which: E[gm (I)] = µm
(known)
Baltimore 2005 – p.6/23
Problem Formulation Let I denote the values of I(x) at a discrete domain i.e. an N × N lattice. Consider M features-functions gm for which: E[gm (I)] = µm
(known)
Incomplete Description of I F=
f (I) :
Z
gm (I)f (I)dI = µm
Baltimore 2005 – p.6/23
Problem Formulation Let I denote the values of I(x) at a discrete domain i.e. an N × N lattice. Consider M features-functions gm for which: E[gm (I)] = µm
(known)
Incomplete Description of I F=
f (I) :
Z
gm (I)f (I)dI = µm
•
Real pdf f ∗ (unknown) belongs in F.
•
Are all pdfs in F equivalent?
Baltimore 2005 – p.6/23
Existing Methods Most commonly use maps from Gaussian fields that are able to reproduce at best 2nd order information.
Baltimore 2005 – p.7/23
Existing Methods Most commonly use maps from Gaussian fields that are able to reproduce at best 2nd order information. Deficiencies •
2nd order characteristics give an incomplete description of the uncertainties in the medium.
Baltimore 2005 – p.7/23
Existing Methods Most commonly use maps from Gaussian fields that are able to reproduce at best 2nd order information. Deficiencies •
2nd order characteristics give an incomplete description of the uncertainties in the medium.
(Yeong & Torquato 1998)
Baltimore 2005 – p.7/23
Existing Methods Most commonly use maps from Gaussian fields that are able to reproduce at best 2nd order information. Deficiencies •
2nd order characteristics give an incomplete description of the uncertainties in the medium.
•
All pdfs f ∈ F are considered equivalent for simulation purposes.
Baltimore 2005 – p.7/23
Revised Requirements - Goals •
The simulation method should be able to incorporate as much information as possible.
Baltimore 2005 – p.8/23
Revised Requirements - Goals •
The simulation method should be able to incorporate as much information as possible.
•
A rationale should be developed to select the most appropriate f ∈ F.
Baltimore 2005 – p.8/23
Proposed Framework - Theory Maximum Entropy Principle R p(I) = arg maxf ∈F − f (I) log f (I)dI
subject to f (I) ∈ F, i.e: Z f (I)dI = 1 and Ef [gm (I)] = µm
∀m
Baltimore 2005 – p.9/23
Proposed Framework - Theory Maximum Entropy Principle R p(I) = arg maxf ∈F − f (I) log f (I)dI
subject to f (I) ∈ F, i.e: Z f (I)dI = 1 and Ef [gm (I)] = µm •
∀m
Information of any order can be introduced.
Baltimore 2005 – p.9/23
Proposed Framework - Theory Maximum Entropy Principle R p(I) = arg maxf ∈F − f (I) log f (I)dI
subject to f (I) ∈ F, i.e: Z f (I)dI = 1 and Ef [gm (I)] = µm
∀m
•
Information of any order can be introduced.
•
The fusion of the available information is done in an optimal manner.
Baltimore 2005 – p.9/23
Proposed Framework - Theory Maximum Entropy Principle R p(I) = arg maxf ∈F − f (I) log f (I)dI
subject to f (I) ∈ F, i.e: Z f (I)dI = 1 and Ef [gm (I)] = µm
∀m
•
Information of any order can be introduced.
•
The fusion of the available information is done in an optimal manner.
(Jaynes 1979): “ Given incomplete information, the distribution of maximum entropy is not only the one that can be realized in the greatest number of ways; for large sample size the overwhelming majority of distributions compatible with our information have entropy very close to the maximum.” Baltimore 2005 – p.9/23
Proposed Framework - Theory Maximum Entropy Distribution P 1 p(I) = Z(λ) exp {− m λm gm (I)}
Baltimore 2005 – p.10/23
Proposed Framework - Theory Maximum Entropy Distribution P 1 exp {− m λm gm (I)} p(I) = Z(λ) R P partition function : Z(λ) = exp {− m λm gm (I)} dI
Baltimore 2005 – p.10/23
Proposed Framework - Theory Maximum Entropy Distribution P 1 p(I) = Z(λ) exp {− m λm gm (I)} The vector λ = (λ1 , λ2 , . . . , λm ) is determined by solving the equations: R Ep [gm (I)] = gm (I)p(I)dI = µm ∀m
Baltimore 2005 – p.10/23
Proposed Framework - Theory Maximum Entropy Distribution P 1 p(I) = Z(λ) exp {− m λm gm (I)} The vector λ = (λ1 , λ2 , . . . , λm ) is determined by solving the equations: R Ep [gm (I)] = gm (I)p(I)dI = µm ∀m Equivalence with Maximum Log-Likelihood Estimation P P L(λ) = k log p(Ik ) = − m λm µm − log Z(λ)
Baltimore 2005 – p.10/23
Proposed Framework - Theory Maximum Entropy Distribution P 1 p(I) = Z(λ) exp {− m λm gm (I)} The vector λ = (λ1 , λ2 , . . . , λm ) is determined by solving the equations: R Ep [gm (I)] = gm (I)p(I)dI = µm ∀m Equivalence with Maximum Log-Likelihood Estimation P P L(λ) = k log p(Ik ) = − m λm µm − log Z(λ)
If gm are linearly independent then L(λ) is concave and the solution vector λ is unique.
Baltimore 2005 – p.10/23
Proposed Framework - Algorithms •
Find λ that maximizes L(λ) = − find p(I).
P
m
λm µm − log Z(λ) in order to
Baltimore 2005 – p.11/23
Proposed Framework - Algorithms P
•
Find λ that maximizes L(λ) = − find p(I).
•
Draw samples from the maximum entropy distribution P 1 exp {− m λm gm (I)}. p(I) = Z(λ)
m
λm µm − log Z(λ) in order to
Baltimore 2005 – p.11/23
Proposed Framework - Algorithms P
•
Find λ that maximizes L(λ) = − find p(I).
•
Draw samples from the maximum entropy distribution P 1 exp {− m λm gm (I)}. p(I) = Z(λ)
m
λm µm − log Z(λ) in order to
These tasks can be carried out simultaneously using Markov Chain Monte Carlo
Baltimore 2005 – p.11/23
Proposed Framework - Maximization of L(λ) L(λ) = −
X
λm µm −log Z(λ)
m
Baltimore 2005 – p.12/23
Proposed Framework - Maximization of L(λ) L(λ) = −
X m
λm µm −log Z(λ)
Z(λ) =
R
P
exp − m λm gm (I) dI not known analytically
Baltimore 2005 – p.12/23
Proposed Framework - Maximization of L(λ) L(λ) = −
X
λm µm −log Z(λ)
Z(λ) =
m
R
P
exp − m λm gm (I) dI not known analytically
Importance Sampling •
For λ0 = 0 (uniform distribution p0 ), Z(0) =
R
1dI is known.
Baltimore 2005 – p.12/23
Proposed Framework - Maximization of L(λ) L(λ) = −
X
λm µm −log Z(λ)
Z(λ) =
m
R
P
exp − m λm gm (I) dI not known analytically
Importance Sampling •
For λ0 = 0 (uniform distribution p0 ), Z(0) =
•
For λ1 6= λ0 : 1
Z(λ )
= =
Z
e−
P
Z(λ0 )
1 m λm gm (I)
Z
p0 (I)
R
1dI is known.
0
p0 (I)dI = Z(λ )
Z
e
−
P
1 0 m (λm −λm )gm (I) p0 (I)dI
w(I)p0 (I)dI
Baltimore 2005 – p.12/23
Proposed Framework - Maximization of L(λ) L(λ) = −
X
λm µm −log Z(λ)
Z(λ) =
m
R
P
exp − m λm gm (I) dI not known analytically
Importance Sampling •
For λ0 = 0 (uniform distribution p0 ), Z(0) =
•
For λ1 6= λ0 : 1
Z(λ )
= =
Z
e−
0
P
Z(λ )
1 m λm gm (I)
p0 (I)
Z
R
1dI is known.
0
p0 (I)dI = Z(λ )
Z
e
−
P
1 0 m (λm −λm )gm (I) p0 (I)dI
1 X w(I )p0 (I )dI ≈ Z(λ ) w(I (j) ) N j 0
Baltimore 2005 – p.12/23
Proposed Framework - Maximization of L(λ) L(λ) = −
X
λm µm −log Z(λ)
Z(λ) =
m
R
P
exp − m λm gm (I) dI not known analytically
Importance Sampling •
For λ0 = 0 (uniform distribution p0 ), Z(0) =
•
For λ1 6= λ0 : 1
Z(λ )
= =
Z
e−
0
P
Z(λ )
1 m λm gm (I)
p0 (I)
Z
R
1dI is known.
0
p0 (I)dI = Z(λ )
Z
e−
P
1 0 m (λm −λm )gm (I)
p0 (I)dI
1 X w(I)p0 (I )dI ≈ Z(λ ) w(I (j) ) N j 0
Baltimore 2005 – p.12/23
Proposed Framework - Maximization of L(λ) L(λ) = −
X
λm µm −log Z(λ)
Z(λ) =
m
R
P
exp − m λm gm (I) dI not known analytically
Importance Sampling •
For λ0 = 0 (uniform distribution p0 ), Z(0) =
•
For λ1 6= λ0 : 1
Z(λ )
= =
Z
e−
0
P
Z(λ )
1 m λm gm (I)
p0 (I)
Z
R
1dI is known.
0
p0 (I)dI = Z(λ )
Z
e−
P
1 0 m (λm −λm )gm (I)
p0 (I)dI
1 X w(I)p0 (I )dI ≈ Z(λ ) w(I (j) ) N j 0
- Noisy function evaluations
Baltimore 2005 – p.12/23
Proposed Framework - Maximization of L(λ) L(λ) = −
X
λm µm −log Z(λ)
Z(λ) =
m
R
P
exp − m λm gm (I) dI not known analytically
Importance Sampling •
For λ0 = 0 (uniform distribution p0 ), Z(0) =
•
For λ1 6= λ0 : 1
Z(λ )
= =
Z
e−
0
P
Z(λ )
1 m λm gm (I)
p0 (I)
Z
R
1dI is known.
0
p0 (I)dI = Z(λ )
Z
e−
P
1 0 m (λm −λm )gm (I)
p0 (I)dI
1 X w(I)p0 (I )dI ≈ Z(λ ) w(I (j) ) N j 0
- Noisy function evaluations - The closer λ1 is to λ0 , the smaller the variance of w(I) and the less the noise. Baltimore 2005 – p.12/23
Proposed Framework - Maximization of L(λ) Conjugate Gradients 1 Iteration i = 0: Set λ0 = 0, u0 = −∇L(λ0 ) and v0 = u0 . 2 Let a∗ = arg maxa L(λi + αvi ). Set λi+1 = λi + a∗ vi . 3 Set ui+1 = −∇L(λi+1 ) and vi+1 = ui+1 + γvi . 4 Set i = i + 1 and goto step 2.
Baltimore 2005 – p.13/23
Proposed Framework - Maximization of L(λ) Conjugate Gradients 1 Iteration i = 0: Set λ0 = 0, u0 = −∇L(λ0 ) and v0 = u0 . 2 Let a∗ = arg maxa L(λi + αvi ). Set λi+1 = λi + a∗ vi . 3 Set ui+1 = −∇L(λi+1 ) and vi+1 = ui+1 + γvi . 4 Set i = i + 1 and goto step 2. Advantages •
Robust for noisy functions
Baltimore 2005 – p.13/23
Proposed Framework - Maximization of L(λ) Conjugate Gradients 1 Iteration i = 0: Set λ0 = 0, u0 = −∇L(λ0 ) and v0 = u0 . 2 Let a∗ = arg maxa L(λi + αvi ). Set λi+1 = λi + a∗ vi . 3 Set ui+1 = −∇L(λi+1 ) and vi+1 = ui+1 + γvi . 4 Set i = i + 1 and goto step 2. Advantages •
Robust for noisy functions
•
Allows for small enough steps that result in reduced noise
Baltimore 2005 – p.13/23
Proposed Framework - Maximization of L(λ) Conjugate Gradients 1 Iteration i = 0: Set λ0 = 0, u0 = −∇L(λ0 ) and v0 = u0 . 2 Let a∗ = arg maxa L(λi + αvi ). Set λi+1 = λi + a∗ vi . 3 Set ui+1 = −∇L(λi+1 ) and vi+1 = ui+1 + γvi . 4 Set i = i + 1 and goto step 2. Advantages •
Robust for noisy functions
•
Allows for small enough steps that result in reduced noise
•
Makes use of only the derivatives of L instead of the curvature matrix
∂ 2 L(λ) ∂λi λj
which
requires more storage and its estimation contains higher error.
Baltimore 2005 – p.13/23
Proposed Framework - Calculation of L(λ) and ∇L(λ) i
L(λ ) = −
P
i m λm µ m
i
− log Z(λ )
∂L(λi ) ∂λm
= −µm + Ep(i) [gm (I)]
Baltimore 2005 – p.14/23
Proposed Framework - Calculation of L(λ) and ∇L(λ) i
L(λ ) = − Z(λi ) =
R
P
i m λm µ m
i
− log Z(λ )
∂L(λi ) ∂λm
= −µm + Ep(i) [gm (I)]
P exp − m λim gm (I) dI
Baltimore 2005 – p.14/23
Proposed Framework - Calculation of L(λ) and ∇L(λ) i
L(λ ) = − Z(λi )
=
R
P
exp −
i m λm µ m
P
i
− log Z(λ )
i m λm gm (I)
dI
∂L(λi ) ∂λm
p(i) (I)
=
= −µm + E
1 Z(λi )
exp −
P
p(i)
[gm (I)]
i m λm gm (I)
Baltimore 2005 – p.14/23
Proposed Framework - Calculation of L(λ) and ∇L(λ) i
L(λ ) = − Z(λi )
=
R
P
exp −
i m λm µ m
P
i
− log Z(λ )
i m λm gm (I)
dI
∂L(λi ) ∂λm
p(i) (I)
=
= −µm + E
1 Z(λi )
exp −
P
p(i)
[gm (I)]
i m λm gm (I)
Importance Sampling Estimators ˆ Z(λ
i+1
N ˆ i) X Z(λ )= w(I (j) ) N j=1
Epi [gm (I)] ≈
µ ˆi+1 m
N ˆ i+1 ) X Z(λ (j) (j) = w(I )g (I ) m N Z(λi ) j=1
where: •
w(I) = e
−
P
i+1 i m (λm −λm )gm (I)
Baltimore 2005 – p.14/23
Proposed Framework - Calculation of L(λ) and ∇L(λ) i
L(λ ) = − Z(λi )
=
R
P
exp −
i m λm µ m
P
i
− log Z(λ )
i m λm gm (I)
dI
∂L(λi ) ∂λm
p(i) (I)
=
= −µm + E
1 Z(λi )
exp −
P
p(i)
[gm (I)]
i m λm gm (I)
Importance Sampling Estimators ˆ Z(λ
i+1
N ˆ i) X Z(λ )= w(I (j) ) N j=1
Epi [gm (I)] ≈
µ ˆi+1 m
N ˆ i+1 ) X Z(λ (j) (j) = w(I )g (I ) m N Z(λi ) j=1
where: −
P
i+1 i m (λm −λm )gm (I)
•
w(I) = e
•
The samples I (j) are drawn from p(i) (I).
Baltimore 2005 – p.14/23
Simulation of samples I (j) from p(i) (I) Gibbs Sampler (or Metropolis-Hastings sampler).
Baltimore 2005 – p.15/23
Simulation of samples I (j) from p(i) (I) Gibbs Sampler (or Metropolis-Hastings sampler). 1. Given I (j) , select randomly a point x ∈ D under the uniform distribution.
Baltimore 2005 – p.15/23
Simulation of samples I (j) from p(i) (I) Gibbs Sampler (or Metropolis-Hastings sampler). 1. Given I (j) , select randomly a point x ∈ D under the uniform distribution. 2. Draw Ix from the conditional distribution p(i) (Ix /I−x ) where Ix is the value of I (j) at x and I−x denotes the rest of the values of I (j) (at every point y ∈ D\{x}).
Baltimore 2005 – p.15/23
Simulation of samples I (j) from p(i) (I) Gibbs Sampler (or Metropolis-Hastings sampler). 1. Given I (j) , select randomly a point x ∈ D under the uniform distribution. 2. Draw Ix from the conditional distribution p(i) (Ix /I−x ) where Ix is the value of I (j) at x and I−x denotes the rest of the values of I (j) (at every point y ∈ D\{x}). (j+1)
3. Set Ix(j+1) = Ix and I−x
(j)
= I−x .
Baltimore 2005 – p.15/23
Simulation of samples I (j) from p(i) (I) Gibbs Sampler (or Metropolis-Hastings sampler). 1. Given I (j) , select randomly a point x ∈ D under the uniform distribution. 2. Draw Ix from the conditional distribution p(i) (Ix /I−x ) where Ix is the value of I (j) at x and I−x denotes the rest of the values of I (j) (at every point y ∈ D\{x}). (j+1)
3. Set Ix(j+1) = Ix and I−x
(j)
= I−x .
4. Set j = j + 1. If j ≤ N goto Step 1.
Baltimore 2005 – p.15/23
Simulation of samples I (j) from p(i) (I) Gibbs Sampler (or Metropolis-Hastings sampler). 1. Given I (j) , select randomly a point x ∈ D under the uniform distribution. 2. Draw Ix from the conditional distribution p(i) (Ix /I−x ) where Ix is the value of I (j) at x and I−x denotes the rest of the values of I (j) (at every point y ∈ D\{x}). (j+1)
3. Set Ix(j+1) = Ix and I−x
(j)
= I−x .
4. Set j = j + 1. If j ≤ N goto Step 1. Samples {I (j) }N j=1 are correlated but asymptotically distributed according to p(i) (independently of the initial sample I (0) ).
Baltimore 2005 – p.15/23
Simulation of samples I (j) from p(i) (I) Gibbs Sampler (or Metropolis-Hastings sampler). 1. Given I (j) , select randomly a point x ∈ D under the uniform distribution. 2. Draw Ix from the conditional distribution p(i) (Ix /I−x ) where Ix is the value of I (j) at x and I−x denotes the rest of the values of I (j) (at every point y ∈ D\{x}). (j+1)
3. Set Ix(j+1) = Ix and I−x
(j)
= I−x .
4. Set j = j + 1. If j ≤ N goto Step 1. Samples {I (j) }N j=1 are correlated but asymptotically distributed according to p(i) (independently of the initial sample I (0) ). The respective estimators converge asymptotically as N → ∞. Baltimore 2005 – p.15/23
Applications in 1D - Hard Rods
Baltimore 2005 – p.16/23
Applications in 1D - Hard Rods
Baltimore 2005 – p.16/23
Applications in 1D - Hard Rods Simulation based on Autocorrelation: gm (I) = Ii Ii+m
Baltimore 2005 – p.16/23
Applications in 1D - Hard Rods Simulation based on Autocorrelation & Lineal-Path Function: gm (I) = Ii Ii+m
gm (I) = Ii Ii+1 . . . Ii+m
Baltimore 2005 – p.16/23
Applications in 2D - Hard Disks Simulation based on the Lineal-Path Function:
Baltimore 2005 – p.17/23
Applications in 2D - Hard Disks Simulation based on the Lineal-Path Function:
128 × 128 pixels Baltimore 2005 – p.17/23
Applications in 2D - Hard Disks Simulation based on the Lineal-Path Function:
128 × 128 pixels Baltimore 2005 – p.17/23
Applications in 2D - Hard Disks Simulation based on the Lineal-Path Function:
128 × 128 pixels Baltimore 2005 – p.17/23
Applications in 2D - Hard Disks Simulation based on the Lineal-Path Function:
128 × 128 pixels Baltimore 2005 – p.17/23
Applications in 2D - Hard Disks Simulation based on the Lineal-Path Function:
128 × 128 pixels Baltimore 2005 – p.17/23
Applications in 2D - Hard Disks Simulation based on the Lineal-Path Function:
128 × 128 pixels Baltimore 2005 – p.17/23
Applications in 2D - Hard Disks Simulation based on the Lineal-Path Function:
128 × 128 pixels Baltimore 2005 – p.17/23
Applications in 2D - Three-phase medium Simulation based on the Lineal-Path Function:
Baltimore 2005 – p.18/23
Applications in 2D - Three-phase medium Simulation based on the Lineal-Path Function:
128 × 128 pixels Baltimore 2005 – p.18/23
Applications in 2D - Functionally Graded Material Simulation based on the Volume Fraction: gm (I) = Im,j
Baltimore 2005 – p.19/23
Applications in 2D - Functionally Graded Material Simulation based on the Volume Fraction: gm (I) = Im,j
128 × 128 pixels
Baltimore 2005 – p.19/23
Feature Selection Maximum Entropy Distribution P 1 p(I) = Z(λ) exp {− m λm gm (I)}
Baltimore 2005 – p.20/23
Feature Selection Maximum Entropy Distribution P 1 p(I) = Z(λ) exp {− m λm gm (I)} Ep [gm (I)] = µm
Baltimore 2005 – p.20/23
Feature Selection Maximum Entropy Distribution P 1 p(I) = Z(λ) exp − m λm gm (I)
Baltimore 2005 – p.20/23
Feature Selection Maximum Entropy Distribution P 1 p(I) = Z(λ) exp − m λm gm (I)
We are given a number of samples of the medium {Ik }K k=1 drawn from the actual (unknown) distribution f ∗ (I).
Baltimore 2005 – p.20/23
Feature Selection Maximum Entropy Distribution P 1 p(I) = Z(λ) exp − m λm gm (I)
We are given a number of samples of the medium {Ik }K k=1 drawn from the actual (unknown) distribution f ∗ (I). Question Which features-functions gm (I) should be selected so that the maximum entropy distribution is as close as possible to the actual.
Baltimore 2005 – p.20/23
Feature Selection Kullback-Leibler Divergence (distance) ∗
D(f , p) = −
R
f ∗ (I) log
p(I) f ∗ (I) dI
Baltimore 2005 – p.21/23
Feature Selection Kullback-Leibler Divergence (distance) ∗
D(f , p) = −
R
f ∗ (I) log
p(I) f ∗ (I) dI
actual (unknown) pdf
Baltimore 2005 – p.21/23
Feature Selection Kullback-Leibler Divergence (distance) ∗
D(f , p) = −
R
f ∗ (I) log
p(I)
f ∗ (I) dI
max. entropy pdf
Baltimore 2005 – p.21/23
Feature Selection Kullback-Leibler Divergence (distance) ∗
D(f , p) = −
R
f ∗ (I) log
p(I) f ∗ (I) dI
strictly positive (except f ∗ (I) = p(I) ∀I)
Baltimore 2005 – p.21/23
Feature Selection Kullback-Leibler Divergence (distance) ∗
D(f , p) = −
R
f ∗ (I) log
p(I) f ∗ (I) dI
strictly positive (except f ∗ (I) = p(I) ∀I) D(f ∗ , p) = Entropy(p) − Entropy(f ∗ )
Baltimore 2005 – p.21/23
Feature Selection Kullback-Leibler Divergence (distance) ∗
D(f , p) = −
R
f ∗ (I) log
p(I) f ∗ (I) dI
strictly positive (except f ∗ (I) = p(I) ∀I) D(f ∗ , p) = Entropy(p) − Entropy(f ∗ ) Given a number of features {gm }m of which we need to select M , then the optimal choice is the M − sized batch that minimizes (w.r.t. g m ) the entropy of the maximum entropy distribution p.
Baltimore 2005 – p.21/23
Feature Selection - Algorithmic Considerations Performing an optimization over all M − sized batches of features is practically infeasible.
Baltimore 2005 – p.22/23
Feature Selection - Algorithmic Considerations Performing an optimization over all M − sized batches of features is practically infeasible. Greedy-Stepwise Algorithm: If p
(M )
=
1 − Z(λ) e
PM m
λm gm (I)
then:
Baltimore 2005 – p.22/23
Feature Selection - Algorithmic Considerations Performing an optimization over all M − sized batches of features is practically infeasible. Greedy-Stepwise Algorithm: If p
(M )
=
1 − Z(λ) e
PM m
λm gm (I)
then:
gM +1 = arg max δ(gm ) = Entropy(p(M ) ) − Entropy(p(M +1) ) gm
Baltimore 2005 – p.22/23
Feature Selection - Algorithmic Considerations Performing an optimization over all M − sized batches of features is practically infeasible. Greedy-Stepwise Algorithm: If p
(M )
=
1 − Z(λ) e
PM m
λm gm (I)
then:
gM +1 = arg max δ(gm ) = Entropy(p(M ) ) − Entropy(p(M +1) ) gm
For features of the same scale: δ(gm ) ≈k Ep(M ) [gm ] − Ef ∗ [gm ] k
Baltimore 2005 – p.22/23
Conclusions •
Discrete-valued random fields can be used to model random heterogeneous materials.
Baltimore 2005 – p.23/23
Conclusions • Discrete-valued random fields can be used to model random heterogeneous
materials. •
In practice, these are not uniquely defined by the available probabilistic information.
Baltimore 2005 – p.23/23
Conclusions • Discrete-valued random fields can be used to model random heterogeneous
materials. • In practice, these are not uniquely defined by the available probabilistic information. •
Simulation methods should be able to incorporate as much probabilistic information as possible and not be restricted to 2nd order statistics which provide an inaccurate description.
Baltimore 2005 – p.23/23
Conclusions • Discrete-valued random fields can be used to model random heterogeneous
materials. • In practice, these are not uniquely defined by the available probabilistic information. • Simulation methods should be able to incorporate as much probabilistic information
as possible and not be restricted to 2nd order statistics which provide an inaccurate description. •
In the proposed framework, the fusion of the available information is done in an optimal manner by employing the maximum entropy principle.
Baltimore 2005 – p.23/23
Conclusions • Discrete-valued random fields can be used to model random heterogeneous
materials. • In practice, these are not uniquely defined by the available probabilistic information. • Simulation methods should be able to incorporate as much probabilistic information
as possible and not be restricted to 2nd order statistics which provide an inaccurate description. • In the proposed framework, the fusion of the available information is done in an
optimal manner by employing the maximum entropy principle. •
An MCMC-based procedure is used to incorporate the target probabilistic characteristics and draw samples from the maximum entropy distribution.
Baltimore 2005 – p.23/23
Conclusions • Discrete-valued random fields can be used to model random heterogeneous
materials. • In practice, these are not uniquely defined by the available probabilistic information. • Simulation methods should be able to incorporate as much probabilistic information
as possible and not be restricted to 2nd order statistics which provide an inaccurate description. • In the proposed framework, the fusion of the available information is done in an
optimal manner by employing the maximum entropy principle. • An MCMC-based procedure is used to incorporate the target probabilistic
characteristics and draw samples from the maximum entropy distribution. •
The Kullback-Leibler divergence can be used to assess the descriptive power of the probabilistic features selected. Baltimore 2005 – p.23/23
Conclusions • Discrete-valued random fields can be used to model random heterogeneous
materials. • In practice, these are not uniquely defined by the available probabilistic information. • Simulation methods should be able to incorporate as much probabilistic information
as possible and not be restricted to 2nd order statistics which provide an inaccurate description. • In the proposed framework, the fusion of the available information is done in an
optimal manner by employing the maximum entropy principle. • An MCMC-based procedure is used to incorporate the target probabilistic
characteristics and draw samples from the maximum entropy distribution. • The Kullback-Leibler divergence can be used to assess the descriptive power of the
probabilistic features selected. Baltimore 2005 – p.23/23