%A Krung Sinapiromsaran
%T Practical Optimization of Simulation: Computation and Tools
%D December 2000
%R 00-07
%I COMPUTER SCIENCES DEPARTMENT, UNIVERSITY OF WISCONSIN
%C MADISON, WI
%X
The main topic of this thesis is solving simulation optimizations using a
deterministic nonlinear solver based on the sample-path optimization concept.
The simulation function is considered as a black box that deterministically
returns the exact output for the same input value.
The gradient-based nonlinear solver finds a local optimal based on the
function and gradient evaluation of the sample path simulation function.
The simulation output is used for the function evaluation while
the derivative of a quadratic model is returned for the gradient evaluation.
We locally build a quadratic model from the surrounding simulation points
using a least squares approximation.
This scheme does not require the modification of the original simulation
source code and can be carried out automatically.
Due to the large number of simulation runs, the high-throughput
computing environment, CONDOR, is used.
Simulation computations are then distributed over heterogeneous machines which
can be executed on entirely different computer architectures within the network.
Additionally, a resource failure is automatically handled using
the checkpoint and migration feature in this CONDOR environment.
We implement a Master-Worker condor PVM server to avoid CONDOR
scheduler waiting period overhead.
At the end of the thesis, we show how to solve a nonlinear programming problem
using a primal-dual formulation.
The first and second order derivative models of the Lagrangian function are
built using an automatic differentiation that is supplied by a
modeling language.
In addition, the original nonlinear objective function is incorporated
into a Fischer-Burmeister merit function to guide the mixed complementarity
solver to find the optimal solution of a nonlinear program.
\noindent {\bf Keywords:}
Simulation optimization, gradient based optimization,
sample-path assumption, modeling language, nonlinear program,
high-throughput computing environment,
Complementarity problem, automatic differentiation.