Text

theses

This thesis addresses the optimization of systems whose behavior is described by noisy input-output data instead of model equations. Process models may not exist, as in the case of emergent technologies, or may be inaccessible if they are embedded within a legacy computational code. When a functional form for the input-output relationships is unavailable, the process behavior is symbolically described using black-box models. Two cases motivate the need to address problems containing black-box models: 1) building a case for obtaining continued research funding during the early product life cycle, when the system information is limited to a sparse sampling set, and 2) process train optimization for systems that have been retrofitted or exhibit behavior which results in suboptimal performance. The challenge is to determine the best operating conditions which satisfy some objective, such as maximizing reaction yield or minimizing utilities costs, based on a limited amount of additional sampling that can be performed.
Surrogate data-driven models can be alternatively generated, but many substitute models may need to be built, especially in the case of process synthesis problems. Although model reliability can be improved using additional information, resource constraints can limit the number of additional experiments allowed. Since it may not be possible to a priori estimate the problem cost in terms of the number of experiments required, there is a need for strategies targeted at the generation of sufficiently accurate surrogate models at low resource cost. The problem addressed in this work focuses on the development of model-based optimization algorithms targeted at obtaining the best solutions based on limited sampling. A centroid-based sampling algorithm for global modeling has also been developed to accelerate accurate global model generation and improve subsequent local optimization. The developed algorithms enable the superior local solutions of problems containing black-box models and noisy input-output data to be obtained when the problem contains both continuous and integer variables and is defined by an arbitrary convex feasible region. The proposed algorithms are applied to many numerical examples and industrial case studies to demonstrate the improved optima attained when surrogate models are built prior to optimization.

Ph.D.

Includes bibliographical references (p. 268-270).

http://hdl.rutgers.edu/1782.2/rucore10001600001.ETD.17455

ETD_1125

doi:10.7282/T3D21XX5

ETD doctoral

The author owns the copyright to this work.

application/x-tar

3398144

0330be2fc4246eb0e0d4b5d4efa68770e596cb56

ETD

other

windows xp

application/x-tar