Development of optimization methods to deal with current challenges in engineering design optimization

Download files
Access & Terms of Use
open access
Copyright: Singh, Hemant
Altmetric
Abstract
In engineering design, optimization is customary, and often indispensable. Typical cases include minimization of drag for vehicles, minimization of weight for structures like buildings and bridges, maximization of power and lift for aircraft and rockets, minimization of fuel consumption for engines, etc. Therefore it comes as no surprise that development of fast and efficient optimization algorithms for engineering design is an actively pursued research area. In recent decades, metaheuristic algorithms have proven to be efficient, robust and versatile methods for numerical optimization. However, they usually need to evaluate a large number of candidate designs to find the optimum. This becomes prohibitive for engineering optimization problems in which each design evaluation may require computationally expensive analysis and, consequently, the optimization process may take much longer time than affordable. Considering that the number of design evaluations is a critical factor in overall optimization time, it is imperative to develop techniques to search for the optimum design using fewest evaluations possible. With this singular goal, this thesis investigates a range of domains in which existing approaches can be improved. Engineering problems are often highly non-linear, discontinuous, and non-differentiable, which rules out (or restricts) the applicability of analytical techniques for solving them. However, they exhibit additional attributes that prove challenging even to the existing metaheuristic techniques, thus making the search difficult and, consequently, creating a necessity for carrying out large numbers of evaluations. These include: (a) Constraints - constraints render a fraction (possibly large fraction) of the search space infeasible, making it hard to find the optimum and at times even a feasible design; (b) Large number of objectives - Pareto-dominance sorting, a commonly used technique in multi-objective optimization algorithms, is inadequate to solve problems with large numbers of objectives, a fact well reported in literature; (c) Large number of variables - the search space grows exponentially with the number of variables, which results in a corresponding increase in computational effort; and (d) Multiple models - For certain problems, there may be multiple candidate models to choose a solution from, with none of them being an obviously preferred one. In such a case, one may need to explore each one of them to find the global best. In this thesis, studies are conducted on each of these domains individually. Shortcomings of the existing methods are analyzed, and novel techniques are developed for efficiently handling constraints, large number of objectives/variables and multiple models. For effective constraint handling, conventional evolutionary algorithm is enhanced using a novel infeasibility driven ranking technique, while conventional simulated annealing algorithm is enhanced using an approximate descent direction coupled with dominance-based acceptance criteria. For many-objective problems, improved secondary-ranking methods and a novel dimensionality reduction technique based on Pareto corner search are proposed. For many-variable problems, conventional cooperative coevolutionary algorithm is enhanced using a correlation based partitioning strategy which enables it to deliver competitive performances across a variety of separable and non-separable problems. Lastly, for trans-dimensional problems, a simulated annealing based algorithm which searches through model space and variable space simultaneously is presented. The proposed methods are able to achieve competitive results using markedly fewer numbers of design evaluations compared to conventional optimizers, which is demonstrated through rigorous numerical experiments on benchmark problems. Finally, the proposed algorithms are applied to a number of engineering design problems in order to accentuate their functionality and viability for solving real-life problems.
Persistent link to this record
Link to Publisher Version
Link to Open Access Version
Additional Link
Author(s)
Singh, Hemant
Supervisor(s)
Ray, Tapabrata
Smith, Warren
Creator(s)
Editor(s)
Translator(s)
Curator(s)
Designer(s)
Arranger(s)
Composer(s)
Recordist(s)
Conference Proceedings Editor(s)
Other Contributor(s)
Corporate/Industry Contributor(s)
Publication Year
2011
Resource Type
Thesis
Degree Type
PhD Doctorate
UNSW Faculty
Files
download whole.pdf 3.13 MB Adobe Portable Document Format
Related dataset(s)