See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/242083368

Radiative properties estimation with the Particle

Collision Algorithm based on a sensitivity analysis

Article  in  High Temperatures-High Pressures · January 2009

CITATIONS READS

15 955

3 authors:

Diego Campos Knupp Antônio Silva Neto

Rio de Janeiro State University Rio de Janeiro State University
83 PUBLICATIONS   453 CITATIONS    299 PUBLICATIONS   1,478 CITATIONS   

SEE PROFILE SEE PROFILE

Wagner F. Sacco
Universidade Federal do Oeste do Pará
50 PUBLICATIONS   622 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Heat and fluid flow analysis via integral transforms View project

Diagnóstico de fallos en sistemas industriales aplicando métodos de agrupamiento difuso View project

All content following this page was uploaded by Wagner F. Sacco on 10 April 2016.

The user has requested enhancement of the downloaded file.

High Temperatures-High Pressures, Vol. 38, pp. 137–151 ©2009 Old City Publishing, Inc.
Reprints available directly from the publisher Published by license under the OCP Science imprint,
Photocopying permitted by license only a member of the Old City Publishing Group

Radiative properties estimation with

the Particle Collision Algorithm based
on a sensitivity analysis†

Diego C. Knupp1, Antônio J. Silva Neto2

and Wagner F. Sacco3

Instituto Politécnico, IPRJ, Universidade do Estado do Rio de Janeiro, UERJ

P.O. Box 97282, 28630-050, Nova Friburgo, R.J., Brazil

Received: December 17, 2008. February 16, 2009.

Implicit formulations for parameter estimation inverse problems in which a

cost function is minimized have largely been employed in several applica-
tions related to heat and mass transfer. Even though gradient based methods
have been used in most cases, it has been observed an increasing interest in
the use of stochastic methods for the solution of inverse radiative transfer
problems. Ahybrid approach with the Particle CollisionAlgorithm (PCA) –
a recently developed stochastic method – and the Levenberg-Marquardt
(LM) – a deterministic method – has been successfully used by the authors
for the solution of the inverse problem of participating media radiative
properties estimation. In such approach it is required the solution of the
direct radiative transfer problem which is modeled by the linear version
of the Boltzmann equation. For that purpose it is used a discrete ordinates
method combined with the finite difference method. Even though good
results have been obtained, it has been observed that the PCA solution,
which is later used as the initial guess for the LM, shows lower quality
when experimental data with low sensitivity to the parameters we want to
estimate are used. Here we identify and use for the solution of the inverse
problem only the data with higher sensitivity. This approach yields better
results and prevents LM from not converging.
Keywords: Levenberg-Marquardt, Particle Collision Algorithm, radiative transfer,
sensitivity analysis, inverse problem, optimization, stochastic method.

∗ Corresponding author: E-mail: ajsneto@iprj.uerj.br

† Paper presented at the 19th International Congress of Mechanical Engineering November 5–9,
2007, Brasília, DF, Brazil

137

“HTHP” — “HTHP-1069” — 2009/7/31 — 14:05 — page 137 — #1

138 D. C. Knupp et al.

1 INTRODUCTION

Direct and inverse radiative transfer problems have several relevant appli-
cations such as optical tomography [1], computerized tomography [2],
hydrologic optics [3], and radiative properties estimation [4, 5], among many
others. A lot of effort has been devoted to the estimation of the optical thick-
ness, phase function of anisotropic scattering, and the absorption and scattering
coefficients [6–12].
When formulated implicitly [13, 14], inverse problems are usually written
as optimization problems, and the main focus becomes the minimization of
a cost function, for example the one given by the summation of the squared
residues between a calculated and a measured quantity.
In recent years we have used a number of deterministic, stochastic and
hybrid (combined) methods for the solution of inverse radiative transfer prob-
lems with particular emphasis on: (i) Levenberg-Marquardt method (LM);
(ii) Simulated Annealing (SA); (iii) Genetic Algorithms (GA); (iv) Artifi-
cial Neural Networks (ANN); (v) Ant Colony System (ACS); (vi) Particle
Swarm Optimization (PSO); (vii) Generalized Extremal Optimization (GEO);
(vii) Interior Points Method (IPM); and (ix) combinations of the previous
methods.
Sacco and co-workers proposed a novel stochastic optimization method
[15], the so called Particle Collision Algorithm (PCA), which is inspired by the
physics of the interaction of nuclear particles inside nuclear reactors. Knupp
et al. [16] applied PCA for the estimation of the optical thickness, single
scattering albedo and diffuse reflectivities at the inner surface of the boundaries
of one-dimensional participating media.
The great advantage of PCA in relation to other optimization algorithms
such as the genetic algorithm, simulated annealing or particle swarm opti-
mization is that, other than the number of iterations, it does not require any
additional parameters.
The PCA can be applied to continuous or discrete optimization problems
by just changing the perturbation function, while in genetic algorithms, for
example, it is necessary to apply special operators for discrete optimization
problems [17]. Results from the literature show that the PCA outperforms
other metaheuristics with less computational effort [18]. Last but not least, the
PCA is extremely easy to implement.
The purpose of hybridizing stochastic and deterministic methods is that,
while the former promotes a thorough exploration of the search space, the latter
exploits its most promising areas. We do believe that the future in optimization
lies in hybrid algorithms. In fact, there have been many recent efforts in this
research field [19–21].
Regarding the PCA, Sacco et al. [22] hybridized this algorithm with the
well-known Nelder-Mead Simplex [23], outperforming the canonical algo-
rithm, the genetic algorithm and particle swarm optimization in a challenging

“HTHP” — “HTHP-1069” — 2009/7/31 — 14:05 — page 138 — #2

 Radiative Properties Estimation with the PCA based on a Sensitivity Analysis 139

nuclear engineering optimization problem. A hybridization of PCA with the

Levenberg-Marquardt method [24, 25], LM, was also used, in which the for-
mer provides an initial guess for the latter. Such approach was denominated
PCA-LM (1). In order to speed up PCA the same authors have developed a
second hybridization of PCA with LM, denominated PCA-LM(2), in which
the LM is used in the exploitation step of the PCA [26].
Whatever the optimization method we use, we face a major problem when
estimating unknowns with poor sensitivity coefficients. In the present work we
improve the implementation of PCA for the solution of the inverse problem
of radiative properties estimation by performing a sensitivity analysis and
considering only the experimental data with higher sensitivity. We implement
also a five steps algorithm in which each one of the four unknowns is estimated
separately with PCA, and in the end of an iterative procedure LM is used for
the simultaneous estimation of all four unknowns.

2 MATHEMATICAL FORMULATION AND SOLUTION OF THE

DIRECT PROBLEM

Consider a one-dimensional, gray, homogeneous, isotropically scattering, par-

ticipating medium of optical thickness τ0 , with diffusely reflecting boundary
surfaces which are subjected to external radiation. The mathematical for-
mulation for such problem considering no emission inside the medium and
azymuthal symmetry is given by [27]

1
∂I (τ, µ) ω
µ + I (τ, µ) = I (τ, µ )dµ in 0 < τ < τ0 ,
∂τ 2 −1 (1a)
− 1 ≤ µ ≤ 1,

1
I (0, µ) = A1 + 2ρ1 I (0, −µ )µ dµ , µ>0 (1b)
0

1
I (τ0 , µ) = A2 + 2ρ2 I (τ0 , µ )µ dµ , µ<0 (1c)
0

where I (τ, µ) represents the dimensionless radiation intensity, τ is the optical

variable, µ is the cosine of the polar angle θ, i.e. the angle formed by the
radiation beam and the positive τ axis, ω is the single scattering albedo, ρ1 and
ρ2 are the diffuse reflectivities at boundaries τ = 0 and τ = τ0 , respectively,
and A1 and A2 represent the strength of the external sources.
The direct problem arises when the geometry, the radiative properties and
the boundary conditions are known. In that case, problem (1) may be solved
yielding the values of the radiation intensity I (τ, µ), for 0 ≤ τ ≤ τ0 and
−1 ≤ µ ≤ 1.

“HTHP” — “HTHP-1069” — 2009/7/31 — 14:05 — page 139 — #3

140 D. C. Knupp et al.

In order to solve the direct problem we have used Chandrasekhar’s discrete

ordinates method [28], in which the polar angle domain is discretized, and the
integral term on the right hand side of Eq. (1a) is replaced by a gaussian quadra-
ture. We then used a finite-difference approximation for the terms on the left
hand side of Eq. (1a), and by performing forward and backward sweeps, from
τ = 0 to τ = τ0 and from τ = τ0 to τ = 0, respectively, I (τ, µ) is determined
for all spatial and angular nodes of the discretized computational domain.

3 MATHEMATICAL FORMULATION OF THE

INVERSE PROBLEM

In the inverse radiative transfer problem considered in the present work, from
the measured experimental data on the intensity of the exit radiation we want
to obtain estimates for the optical thickness, single scattering albedo, and the
boundary diffuse reflectivities of one-dimensional homogeneous participating
media. That is, we are interested in the following radiative properties, which
are considered unknowns
 = {τ0 , ω, ρ1 , ρ2 }T
Z (2)
As mentioned above, measured data on the intensity of the exit radiation
at the boundaries τ = 0 and τ = τ0 , i.e., Yi , i = 1, 2, . . . , Nd , are considered
available, where Nd represents the total number of experimental data. Because
the number of measured data, Nd , is usually much larger than the number of
estimated parameters, the inverse problem is formulated as a finite dimensional
optimization problem in which the following cost function is minimized (also
referred to as the objective function)

Nd
 =
Q(Z)  − Yi ]2
[Icalci (Z) (3)
i=1
where Icalci represents the calculated value of the radiation intensity (using
estimates for the unknown radiative properties Z)  at the same boundary, and
at the same polar angle, for which the experimental value Yi is obtained.
For the solution of the inverse problem we have used a variant of a stochastic
method recently developed, the Particle Collision Algorithm [15], in which
the exploitation step of the algorithm, instead of consisting in a random search
in the vicinity of a promising solution, is replaced by the deterministic gradient
based Levenberg-Marquardt method [24, 25].

4 THE PARTICLE COLLISION ALGORITHM

The PCA resembles in its structure that of simulated annealing [29]: first an ini-
tial configuration is chosen; then there is a modification of the old configuration

“HTHP” — “HTHP-1069” — 2009/7/31 — 14:05 — page 140 — #4

 Radiative Properties Estimation with the PCA based on a Sensitivity Analysis 141

FIGURE 1
Pseudo code for PCA.

into a new one. The qualities of the two configurations are compared. A deci-
sion then is made on whether the new configuration is “acceptable”. If it is,
it serves as the old configuration for the next step. If it is not acceptable, the
algorithm proceeds with a new change of the old configuration. PCA can also
be considered a Metropolis [30] algorithm, as a trial solution can be accepted
with a certain probability. This acceptance may avoid the convergence to local
optima.
The pseudo code description of the PCA is shown in Fig. 1. The algorithm’s
default is for maximization problems. For minimization, just multiply the
objective function by −1 and invert the ratio in pscattering.
The “stochastic perturbations” in the beginning of the loop are random vari-
ations in each variable’s values within their ranges. We based this mechanism
on that used in the Simulated Annealing algorithm. For details, see function
“Perturbation” in Fig. 2.
If the quality or fitness of the new configuration is better than the fit-
ness of the old configuration, then the “particle” is “absorbed”, there is an
exploitation of the boundaries searching for an even better solution. Function

“HTHP” — “HTHP-1069” — 2009/7/31 — 14:05 — page 141 — #5

142 D. C. Knupp et al.

FIGURE 2
Function “Perturbation”.

FIGURE 3
Function “Small_Perturbation”.

“Exploitation( )” performs this local search, generating a small stochastic

perturbation of the solution inside a loop. In PCA’s current version, it is
a one-hundred-iteration loop. The “small stochastic perturbation” is sim-
ilar to the previous stochastic perturbation, but in a smaller range (see
“Small_Perturbation” in Fig. 3).
Otherwise, if the quality of the new configuration is worse than the old con-
figuration’s, the “particle” is “scattered”. By scattering we mean that the new
configuration receives random values between the upper and lower bounds
of each design variable. The scattering probability (pscattering) is inversely
proportional to its quality. A low-fitness particle will have a greater scattering
probability.
The solution of the inverse radiative transfer problem under analysis con-
sists on obtaining estimates for the unknowns shown in Eq. (2) by minimizing
the cost function given by Eq. (3) using the recently developed stochastic Par-
ticle Collision Algorithm (PCA) [15], whose pseudo-code is shown in Fig. 1.
Stochastic methods are known to be time consuming, mainly for the solu-
tion of inverse problems in which the solution of the direct problem, such
as the one modeled by Eqs. (1a–c), is required. Therefore, we have opted
to use hybrid approaches through the coupling of PCA with the deterministic

“HTHP” — “HTHP-1069” — 2009/7/31 — 14:05 — page 142 — #6

 Radiative Properties Estimation with the PCA based on a Sensitivity Analysis 143

FIGURE 4
Flow diagram of the developed methodology for the estimation of radiative properties based on
a sensitivity analysis. Dτ0 , Dω , Dρ1 and Dρ2 represent the sets of experimental data with higher
sensitivity for each unknown.

gradient based Levenberg-Marquardt method. In a previous work [16] we used

a “weaker” version of PCA (100 iterations in the outer loop, and 10 iterations
in the exploitation loop) in order to generate an initial guess for the LM. Such
approach was denominated PCA-LM (1). In another work [26] we replaced
the exploitation step with a random search, as shown in Fig. 1, by the LM
method. Such approach was denominated PCA-LM (2).
In the present work we use a variation of PCA-LM (1) for the estimation of
the unknowns given by Eq. (2), in which, as shown in Fig. 4, first an iterative
procedure is used being each unknown estimated separately with PCA, but in
order to improve the results we previously perform a sensitivity analysis and
use, for the inverse problem solution, only the experimental data with higher
sensitivity. The sensitivity analysis is presented in the next section.

5 SENSITIVITY ANALYSIS FOR THE INVERSE RADIATIVE

TRANSFER PROBLEM

The sensitivity analysis plays a major role in several aspects related to the
formulation and solution of inverse problems [31]. Here we use the modified
or scaled sensitivity coefficients
 

∂Icalci Z
XZj = Zj , i = 1, 2, . . . , Nd and j = 1, 2, . . . , N (4)
∂Zj

“HTHP” — “HTHP-1069” — 2009/7/31 — 14:05 — page 143 — #7

144 D. C. Knupp et al.

where Zj represents a specific unknown, with j = 1, 2, . . . , N, and N is the

total number of unknowns. In the present work N = 4.
In order to obtain good estimates, i.e. within reasonable confidence bounds,
it is required that the sensitivity coefficients be high. Furthermore, when two
or more unkowns are estimated simultaneously their sensitivity coefficients
must be uncorrelated. Graphically it means that they should not present the
same shape, otherwise two or more parameters may affect in the same way
the observed quantity, Icalci , which is in fact measured experimentally.
In the present work, it has been developed a method based on a sensitivity
analysis in a way that the parameters are estimated separately and only the data
with higher sensitivity are used in the estimation of each variable, as it can be
observed in the flow diagram shown in Fig. 4. Dτ0 ,Dω , Dρ1 and Dρ2 represent
the sets of data used in the estimation of τ0 , ω, ρ1 and ρ2 , respectively. These
data are selected by the user after a sensitivity analysis of the problem under
consideration.
As the number of experimental data may be different for the estimation of
each property, we define the cost function used in the PCA as
Ndτ0

 =
Qτ0 (Z)  − Y i ]2
[Icalci (Z) (5a)
i=1


Ndω
 =
Qω (Z)  − Y i ]2
[Icalci (Z) (5b)
i=1
Ndρ1

 =
Qρ1 (Z)  − Y i ]2
[Icalci (Z) (5c)
i=1
Ndρ2

 =
Qρ2 (Z)  − Y i ]2
[Icalci (Z) (5d)
i=1

so that when the properties τ0 , ω, ρ1 and ρ2 are estimated the respective cost
functions Qτ0 , Qω , Qρ1 and Qρ2 are used. Ndτ0 , Ndω , Ndρ1 and Ndρ2 are the
number of experimental data related to Dτ0 , Dω , Dρ1 and Dρ2 , respectively.

6 RESULTS AND DISCUSSION

As real experimental data were not available, we generated synthetic experi-

mental data using
 exact ) + σe · r
Yi = Icalci (Z (6)
where r is a random number, from a uniform distribution, in the range [−1, 1],
 exact is a vector with the exact values of the unknowns that we want to
Z
estimate, and σe simulates the standard deviation of measurement errors.

“HTHP” — “HTHP-1069” — 2009/7/31 — 14:05 — page 144 — #8

 Radiative Properties Estimation with the PCA based on a Sensitivity Analysis 145

We are interested in the estimation of the four unknown radiative properties

given in Eq. (2). In order to evaluate the performance of the approach presented
here we chose a very difficult test case with

 exact = {τ0 , ω, ρ1 , ρ2 }T = {4.00, 0.30, 0.10, 0.90}

Z (7)

The incident radiation intensity was taken as A1 = 1.0 and A2 = 0.0 in

Eqs. (1b) and (1c), respectively. The ranges considered in the PCA for the
search of unknowns was [0,5.0] for τ0 , and [0,1.0] for ω, ρ1 and ρ2 . These
are the real physical bounds for the unknowns, except for τ0 which may have
a higher value than the upper bound considered. Nonetheless, it must be
stressed that τ0 = 4.0 is already a very high value if one wants to consider
the information on the transmitted radiation for the inverse problem solution.
In Fig. 5 are shown the sensitivity coefficients for the test case considered. It
can be observed that with the exception of ω, all variables present sensitivity
coefficients very close to zero for all data, what demonstrates how difficult is
the estimation of the parameters in this case.
We have considered a angular domain discretization with 20 nodes, as
shown in Fig. 6.
As experimental data we consider the emerging radiation intensities Yi with
i = 1, 2, . . . , M 2 being acquired at τ = τ0 , at the polar angles related to µi
with i = 1, 2, . . . , M 2 , i.e. µ > 0, and the emerging radiation intensities Yi
being acquired at τ = 0, at the polar angles related to µi with i = M 2 + 1, 2 +
M

2, . . . , M, i.e. µ < 0.
Due to the sensitivity analysis performed we then consider the follow-
ing sets of experimental data for the estimation of each unknown with
PCA:Dτ0 = {Y6 , . . . , Y10 }; Dω = {Y11 , . . . , Y20 }; Dρ1 = {Y16 , . . . , Y20 }
and Dρ2 = {Y6 , . . . , Y10 }.
In Tables 1–4 presented next are shown the initial guess for the vector
of unknowns obtained with the first run of PCA, Z  P CA-I NI , the estimates
obtained with PCA and the sets of data with higher sensitivity as shown in the
flow diagram of Fig. 6, Z  P CA−SA , being in this case each unknown estimated
separately, and the estimates obtained with the Levenberg-Marquardt method
(LM), Z  LM , using Z
 P CA−SA as the initial guess. In LM all unknowns were
estimated simultaneously and all experimental data available was considered,
i.e. Yi with i = 1, 2, . . . , M. In Tables 2–4 are also shown the average, z̄, and
the standard deviation, σ , of the runs.
In Table 1 are presented the results obtained with noiseless data, in a single
run. From the second and third columns it can be observed the improvement in
the estimates with respect to the ones obtained with PCA and all experimental
data as shown in the first column of the table.
In Tables 2–4 are shown the results obtained in five runs using noisy
experimental data with up to 5% , 8% and 11% error, respectively.
In Tables 2–4 it can be observed that in the second step of the pro-
cess, i.e. when the estimations are performed separately, using only the

“HTHP” — “HTHP-1069” — 2009/7/31 — 14:05 — page 145 — #9

146 D. C. Knupp et al.

FIGURE 5
 = {4.0, 0.3, 0.1, 0.9}.
Sensitivity coefficients for Z

FIGURE 6
Angular domain discretization.

“HTHP” — “HTHP-1069” — 2009/7/31 — 14:05 — page 146 — #10

 Radiative Properties Estimation with the PCA based on a Sensitivity Analysis 147

Parameters  PCA-INI
Z  PCA-SA
Z  LM
Z
τ0 2.19 4.10 4.00
ω 0.33 0.31 0.30
ρ1 0.23 0.13 0.10
ρ2 0.91 0.89 0.90

Q(Z) 5.25E-4 1.0E-6 0
TABLE 1
 = {4.0, 0.3, 0.1, 0.9}.
Estimations obtained with noiseless experimental data for Z

Run  PCA-INI
Z 
Q(Z)  PCA-SA
Z Q(z)  LM
Z Q(z)
1 {2.16,0.34,0.23,0.92} 5E-4 {3.92,0.29,0.05,0.90} 2E-5 {3.43,0.30,0.13,0.87} 7E-6
2 {1.86,0.34,0.33,0.95} 6E-4 {3.27,0.34,0.21,0.95} 8E-5 {3.35,0.32,0.17,0.95} 3E-6
3 {2.82,0.42,0.46,0.91} 2E-4 {4.04,0.28,0.26,0.90} 3E-3 {3.92,0.28,0.08,0.90} 2E-6
4 {3.76,0.40,0.34,0.47} 1E-3 {4.68,0.28,0.39,0.79} 8E-3 {5.00,0.34,0.32,0.75} 3E-5
5 {3.01,0.38,0.33,0.80} 5E-4 {4.37,0.28,0.02,0.85} 1E-5 {3.15,0.25,0.01,0.95} 2E-5
z̄ {2.39,0.38,0.34,0.81} {4.06,0.29,0.19,0.88} {3.77,0.30,0.14,0.88}
σ {0.75,0.04,0.08,0.20} {0.53,0.03,0.15,0.06} {0.74,0.03,0.12,0.08}

TABLE 2
=
Estimations obtained in five runs with noisy experimental data with up to 5% error.Z
{4.0, 0.3, 0.1, 0.9}.

Run  PCA-INI
Z 
Q(Z)  PCA-SA
Z Q(z)  LM
Z Q(z)
1 {3.80,0.37,0.27,0.79} 4E-4 {4.88,0.34,0.27,0.78} 4E-5 {3.50,0.29,0.04,0.94} 1E-5
2 {3.64,0.50,0.52,0.74} 6E-4 {4.80,0.28,0.02,0.75} 4E-5 {3.15,0.32,0.18,0.96} 5E-6
3 {3.78,0.35,0.33,0.65} 6E-4 {4.91,0.29,0.47,0.73} 9E-3 {4.23,0.24,0.01,0.86} 2E-5
4 {4.46,0.44,0.51,0.40} 4E-4 {4.85,0.29,0.05,0.76} 1E-5 {3.13,0.31,0.15,0.96} 9E-6
5 {3.61,0.51,0.56,0.80} 6E-4 {3.45,0.28,0.01,0.95} 1E-5 {3.40,0.31,0.15,0.95} 7E-6
z̄ {3.86,0.43,0.44,0.68} {4.58,0.30,0.16,0;79} {3.48,0.29,0.11,0.93}
σ {0.35,0.07,0.13,0.17} {0.63,0.03,0.20,0.09} {0.45,0.03,0.08,0.04}

TABLE 3
=
Estimations obtained in five runs with noisy experimental data with up to 8% error.Z
{4.0, 0.3, 0.1, 0.9}.

Run  PCA-INI
Z 
Q(Z)  PCA-SA
Z Q(z)  LM
Z Q(z)
1 {4.25,0.37,0.35,0.51} 5E-4 {4.23,0.28,0.34,0.87} 6E-3 {3.52,0.32,0.08,0.98} 5E-5
2 {3.37,0.38,0.29,0.70} 9E-4 {3.12,0.28,0.01,0.95} 6E-5 {3.11,0.34,0.01,0.94} 9E-5
3 {4.25,0.50,0.55,0.41} 9E-4 {4.81,0.27,0.51,0.73} 9E-3 {5.67,0.26,0.43,0.81} 3E-5
4 {4.04,0.43,0.49,0.86} 1E-4 {4.77,0.41,0.02,0.77} 9E-3 {5.01,0.37,0.30,0.72} 4E-5
5 {3.83,0.42,0.41,0.50} 8E-4 {4.67,0.28,0.05,0.97} 5E-5 Did not converge –
z̄ {3.95,0.42,0.42,0.60} {4.32,0.30,0.19,0.86} {4.33,0.32,0.21,0.86}
σ {0.37,0.05,0.11,0.18} {0.71,0.06,0.23,0.11} {1.21,0.05,0.19,0.12}

TABLE 4
Estimations obtained in five runs with noisy experimental data with up to 11% error
 = {4.0, 0.3, 0.1, 0.9}.
for Z

“HTHP” — “HTHP-1069” — 2009/7/31 — 14:05 — page 147 — #11

148 D. C. Knupp et al.

FIGURE 7
Estimates of ω, considering noisy experimental data with up to 5% error.

FIGURE 8
Estimates of ω, considering noisy experimental data with up to 8% error.

experimental data with higher sensitivity coefficients, the estimates of ω,

the only property with reasonable sensitivity coefficients, are much better
than the estimates obtained when all properties were estimated simultane-
ously using PCA. In Figs. 7–9 are shown the same results, comparing the
estimates of ω obtained by the standard version of PCA (PCA-INI) and the
version using only the experimental data with higher sensitivity coefficients
(PCA-SA).
It can also be noticed that the Levenberg-Marquadt method (LM) converged
in most cases and obtained reasonable estimates for all properties, especially

“HTHP” — “HTHP-1069” — 2009/7/31 — 14:05 — page 148 — #12

 Radiative Properties Estimation with the PCA based on a Sensitivity Analysis 149

FIGURE 9
Estimates of ω, considering noisy experimental data with up to 11% error.

ω, as it was expected. It is important to stress once more that the test case
considered here is a very complicated one and a good initial guess is needed
for LM to converge.
Some tests using the first column estimates (standard PCA) directly as the
initial guesses for LM have also been performed. Although LM converged
and yielded good estimates in most runs, it was not able to converge more
frequently than when it was used the sensitivity analysis approach presented
in this work.
Some tests considering radiative properties for which the sensitivity coef-
ficients were higher have also been performed using the same approach
shown in Fig. 4, but in these cases the estimates did not present significant
improvement.

7 CONCLUSIONS

The results obtained in the present work suggest the feasibility of the estimation
of radiative properties separetely using only the data with better sensitivity
coefficients. It has been intentionally considered a very difficult test case and
the proposed approach was able to yield good estimates for the only property
with reasonable sensitivity.
A version of Levenberg-Marquardt (LM) estimating simultaneously the
properties using a set of data with sensitivity to all unknown radiative proper-
ties has been tried, but no significant improvement has been verified. Further
investigations should be performed on this subject in order to evaluate the
influence of estimating the properties separately.

“HTHP” — “HTHP-1069” — 2009/7/31 — 14:05 — page 149 — #13

150 D. C. Knupp et al.

ACKNOWLEDGEMENTS

The authors acknowledge the financial support provided by CNPq, Con-

selho Nacional de Desenvolvimento Científico e Tecnológico, and FAPERJ,
Fundação Carlos Chagas Filho de Amparo à Pesquisa do Estado do Rio de
Janeiro.

REFERENCES
[1] Kim, H. K., and Charette, A. A sensitivity function – based conjugate gradient method for
optical tomography with the frequency domain equation of radiative transfer. Journal of
Quantitative Spectroscopy & Radiative Transfer 104 (2007), 24–39.
[2] Carita Montero, R. F., Roberty, N. C. and Silva Neto, A. J. Reconstruction of a combination
of the absorption and scattering coefficients with a discrete ordinates method consistent with
the source-detector system. Inverse Problems in Engineering 12(1) (2004), 81–101.
[3] Chalhoub, E. S. and Campos Velho, H. F. Simultaneous estimation of radiation phase
function and albedo in natural waters. Journal of Quantitative Spectroscopy & Radiative
Transfer 69 (2001), 137–149.
[4] Nenarokomov, A. and Titov, D. Optimal experiment design to estimate the radiative prop-
erties of materials. Journal of Quantitative Spectroscopy & Radiative Transfer 93 (2005),
313–323.
[5] Hespel, L., Mainguy, S. and Greffet, J.-J. Radiative properties of scattering and absorbing
dense media: Theory and experimental study. Journal of Quantitative Spectroscopy &
Radiative Transfer 77 (2003), 193–210.
[6] Tinel, C., Testud, J., Guyot, A. and Caillault, K. Cloud parameter retrieval from combined
remote sensing observations. Phys. Chem. Earth (B) 25(10–12) (2000), 1063–1067.
[7] Milandri, A., Asllanaj, F. and Jeandel, G. Determination of radiative properties of fibrous
media by an inverse method – comparison with the Mie theory. Journal of Quantitative
Spectroscopy & Radiative Transfer 74 (2002), 637–653.
[8] Zhou, H.-C., Hou, Y.-B., Chen, D.-L. and Zheng, C.-G. An inverse radiative transfer prob-
lem of simultaneously estimating profiles of temperature and radiative parameters from
boundary intensity and temperature measurements. Journal of Quantitative Spectroscopy &
Radiative Transfer 74 (2002), 605–620.
[9] Kudo, K., Kuroda, A., Fujikane, T., Saido, S. and Oguma, M. Development of a method
to estimate profiles of equivalent absorption coefficient for gray analysis. Journal of
Quantitative Spectroscopy & Radiative Transfer 73 (2002), 385–395.
[10] Souto, R. P., Stephany, S., Becceneri, J. C., Campos Velho, H.F. and Silva Neto, A.J.
Reconstruction of a Spatial Dependent Scattering Albedo in a Radiative Transfer Problem
Using a Hybrid Ant Colony System Implementation and a Pre-Regularization Scheme,
Proceedings of the 6th World Congress of Structural and Multidisciplinary Optimization,
Rio de Janeiro, Brazil, 2005.
[11] Souto, R. P., Campos Velho, H. F. and Stephany, S. Reconstruction of vertical profiles of
the absorption and scattering coefficients from multispectral radiances. Mathematics and
Computers in Simulation 73 (2006), 255–267.
[12] An, W., Ruan, L. M. and Qi, H. Inverse radiation problem in one-dimensional slab by
time-resolved reflected and transmitted signals. Journal of Quantitative Spectroscopy &
Radiative Transfer 107 (2007), 47–60.
[13] Silva Neto, A.J. Explicit and Implicit Formulations for Inverse Radiative Transfer Problems,
Proceedings of the 5th World Congress on Computational Mechanics, Mini-Symposium MS
125 – Computational Treatment of Inverse Problems in Mechanics, Vienna, Austria, 2002.

“HTHP” — “HTHP-1069” — 2009/7/31 — 14:05 — page 150 — #14

 Radiative Properties Estimation with the PCA based on a Sensitivity Analysis 151

[14] Silva Neto, A. J., Roberty, N. C., Pinheiro, R. P. F. and Acevedo, N.I.A. Inverse prob-
lems explicit and implicit formulations with applications in engineering, biophysics and
biotechnology. Inverse Problems in Science and Engineering 15(4) (2007), 373–411.
[15] Sacco, W. F., Oliveira, C. R. E. and Pereira, C. M. N. A. Two stochastic optimization
algorithms applied to nuclear reactor core design. Progress in Nuclear Energy 48(6) (2006),
525–539.
[16] Knupp, D. C., Silva Neto, A. J. and Sacco, W. F. Estimation of Radiative Properties with
the Particle Collision Algorithm, Inverse Problems, Design and Optimization Symposium,
Miami, USA, 2007.
[17] Goldberg D. E. Genetic Algorithms in Search, Optimization & Machine Learning, Addison-
Wesley, Reading, MA, 1987.
[18] Sacco, W. F., Lapa, CMF, Pereira, CMNA and Alves Filho, H. A metropolis algorithm
applied to a nuclear power plant auxiliary feedwater system surveillance tests policy
optimization. Progress in Nuclear Energy 50(1) (2008), 15–21.
[19] Resende, M. G. C. and Werneck, R. F. A hybrid multistart heuristic for the uncapacitated
facility location problem. European Journal of Operational Research 174 (2006), 54–68.
[20] Menon, P. P., Kim, J., Bates, D. G., and Postlethwaite, I. Clearance of nonlinear flight
control laws using hybrid evolutionary optimization. IEEE Transactions on Evolutionary
Computation 10 (2006), 689–699.
[21] Liao, G. C. and Tsao, T. P. Application of a fuzzy neural network combined with a
chaos genetic algorithm and simulated annealing to short-term load forecasting. IEEE
Transactions on Evolutionary Computation 10 (2006), 330–340.
[22] Sacco, W. F., Alves Filho, H., Henderson and Oliveira, C. R. A metropolis algorithm
combined with Nelder-Mead Simplex applied to nuclear reactor core design. Annals of
Nuclear Energy 35(5) (2008), 861–867.
[23] Nelder, J. A. and Mead, R. A simplex method for minimization. Comput. J. (7) (1965),
308–313.
[24] Marquardt, D. W. An Algorithm for least-squares estimation of nonlinear parameters. J. Soc.
Industr. Appl. Math. (11) (1963), 431–441.
[25] Silva Neto, A. J. and Moura Neto, F. D. Inverse Problems – Fundamental Concepts and
Applications, EdUERJ, Rio de Janeiro, Brazil. (in Portuguese) 2005.
[26] Knupp, D. C., Sacco, W. F. and Silva Neto, A.J. Radiative Properties Estimation with a
Combination of the Particle Collision Algorithm and the Levenberg-Marquardt Method,
Inverse Problems Symposium, East Lansing, USA, 2007.
[27] Özisik, M. N. Radiative Transfer and Interactions with Conduction and Convection, John
Wiley, New York, USA, 1973.
[28] Chandrasekhar, S. Radiative Transfer, Dover Publications, Inc., New York, USA, 1960.
[29] Aarts, E. and Korst, J. Simulated Annealing and Boltzmann Machines, John Wiley and Sons,
New York, 1989.
[30] Metropolis, N., Rosenbluth, A. W., Rosenbluth, M. N., Teller, H. A. and Teller, E. Equation
of state calculation by fast computing machines. Journal of Chemical Physics (21) (1953),
1087–1092.
[31] Beck, J. V., Blackwell, B. and St. Clair Jr., C. R. Inverse Heat Conduction – Ill-Posed
Problems, John Wiley & Sons, New York, USA, 1985.

“HTHP” — “HTHP-1069” — 2009/7/31 — 14:05 — page 151 — #15

 “HTHP” — “HTHP-1069” — 2009/7/31 — 14:05 — page 152 — #16

View publication stats