Transcript Document

A Hybrid Optimization Approach
for Automated Parameter Estimation Problems
Carlos A. Quintero1
Miguel Argáez1, Hector Klie2, Leticia Velázquez1 and Mary Wheeler2
2 ICES-The
Center for Subsurface Modeling
The University of Texas at Austin
1Department
of Mathematical Sciences
The University of Texas at El Paso
Local Search: NKIP
We consider the optimization problem in the form:
minimize f S ( x)
Test Case
subject to a  x  b
We test the algorithm on the Rastrigin’s Problems [Ref 6]:
Abstract
n
We present a hybrid optimization approach for solving
automated parameter estimation problems that is based on the
coupling of the Simultaneous Perturbation Stochastic
Approximation (SPSA)[Ref.1] and a globalized Newton-Krylov
Interior Point algorithm (NKIP) presented by Argáez et
al.[Refs.2,3]. The procedure generates a surrogate model that
yield to use efficiently first order information and applies NKIP
algorithm to find an optimal solution. We implement the hybrid
optimization algorithm on a simple test case, and present some
preliminary numerical results.
where a and b are determined by the sampled points given by SPSA.
NKIP is a globalized and derivative dependent optimization method. This
method calculates the directions using the conjugate gradient algorithm, and a
linesearch is implemented to guarantee a sufficient decrease of the objective
function as described in Argáez and Tapia [Ref.4]. This algorithm has been
developed for obtaining an optimal solution for large scale and degenerate
problems.
NKIP algorithm apply to the surrogate model find an optimal solution at
x*=(1.94, 0.01) and fs(x*)= 6.26.
min f ( x1 , x2 ,..., xn )  10n   ( x  10 cos(2xi )),
i 1
2
i
 5.12  xi  5.12
3-D view, n=2
Sampling Data from SPSA
Future Work
Further research and numerical experimentation are needed to demonstrate the
effectiveness of the hybrid optimization scheme being proposed, especially for
solving large application problems of interest of the Department of Defense.
Surrogate Model
Problem Formulation
We find the surrogate model fs(x) using an interpolation method
( xk , f ( xk )),k ,provided
1,..., p
with the data,
by SPSA.
This can be performed in different ways, e.g., radial basis
functions, kriging, regression analysis, or using artificial neural
networks.
We consider the global optimization problem in the form:
minimize f x, f : R n  R (1)
In our test case, we optimize the surrogate function
n
References
m
f s ( x)   w j h j ( x),
We are interested in problem (1) that have many local minima.
j 1
where the multiquadric basis functions are given by
Hybrid Optimization Scheme
n
h j ( x)  1  
The scheme is to use SPSA as the sampling device to perform a
global search of the parameter space and switch to NKIP to perform
the local search via a surrogate model.
i 1
( xi  cij )
rij
2
2
.
The interpolation algorithm [Ref.4] characterizes the uncertainty
parameters cij , rij , w j , i  1,...,n, j  1,..,m.
Multi-Start
We plot the original model function and the surrogate function:
Global Search
Via SPSA
Stop
x*=(0,0) global solution
f(x*)=0
Yes
Explore
Parameter
space
No
Sampling
Filtering Data
Interpolate
Response surface
Optimal Solution found for
Original model?
Improved Solution
fs (x)
Surrogate Model
Sensitivity analysis
To add a multi-start on SPSA
Filter data from SPSA
If x* is such that f(x*) does not satisfy an upper bound given by
 , i.e. f (x*)   , then we use x* as an initial point for SPSA.
Front view
where the global solution x* is such that
f x *  f x for all x  R
1.
2.
3.
Local Search
Via NKIP
fs(x)
SPSA data points
Global Search: SPSA
SPSA is a global derivative free optimization method that uses
only objective function measurements. In contrast with most
algorithms which requires the gradient of the objective function
that is often difficult or impossible to obtain. Further, SPSA is
especially efficient in high-dimensional problems in terms of
providing a good approximate solution for a relatively small
number of measurements of the objective function [Ref. 5].
The parameter estimation is first carried out by means of SPSA
algorithm. This process may be performed by starting with
different initial guesses (multistart). This increases the chances
for finding a global solution, and yields to find a vast sampling
of the parameter space.
[1] J. C. Spall. Introduction to stochastic search and optimization: Estimation,
simulation and control. John Wiley & Sons, Inc., New Jersey, 2003.
[2] M. Argáez, R. Sáenz, and L. Velázquez. A trust–region interior–point
algorithm for nonnegative constrained minimization. Technical report,
Department of Mathematics, The University of Texas at El Paso, 2004.
[3] M. Argáez and R.A. Tapia. On the global convergence of a modified
augmented Lagrangian linesearch interior-point Newton method for nonlinear
programming. J. Optim. Theory Appl., 114:1–25, 2002.
[4]Mark J. L. Orr. Matlab Functions for Radial Basis Function Networks, 1999.
[5] H. Klie and W. Bangerth and M. F. Wheeler and M. Parashar and V.
Matossian. Parallel well location optimization using stochastic algorithms on the
grid computational framework. 9th European Conf. on the Mathematics of Oil
Recovery (ECMOR), August, 2004.
[6] A.J. Keane and P.B. Nair. Computational Approaches for Aerospace Design:
The Pursuit of Excellence. Wiley, England, 2005.
Contact Information
f(x)
Carlos A. Quintero, Graduate Student
The University of Texas at El Paso
Department of Mathematical Sciences
500 W. University Avenue
El Paso, Texas 79968-0514
Email: [email protected]
Phone: (915) 747-6858
Fax: (915) 747-6502
Acknowledgments
The authors were partially supported
by DOD PET project EQM KY06-03 .
The authors thank IMA for the travel
support to attend the Blackwell-Tapia
Conference, November 3-4, 2006.