Development of Methods for Solving Bilevel Optimization Problems

Download files
Access & Terms of Use
open access
Copyright: Islam, Md Monjurul
Altmetric
Abstract
Bilevel optimization, also referred to as bilevel programming, involves solving an upper level problem subject to the optimality of a corresponding lower level problem. The upper and lower level problems are also referred to as the leader and follower problems, respectively. Both levels have their associated objective(s), variable(s) and constraint(s). Such problems model real-life scenarios of cases where the performance of an upper level authority is realizable/sustainable only if the corresponding lower level objective is optimum. A number of practical applications in the field of engineering, logistics, economics and transportation have inherent nested structure that are suited to this type of modelling. The range of applications as well as a rapid increase in the size and complexity of such problems has prompted active interest in the design of efficient algorithms for bilevel optimization. Bilevel optimization problems present a number of unique and interesting challenges to algorithm design. The nested nature of the problem requires optimization of a lower level problem to evaluate each upper level solution, which makes it computationally exorbitant. Theoretically, an upper level solution is considered valid/feasible only if the corresponding lower level variables are the true global optimum of the lower level problem. Global optimality can be reliably asserted in very limited cases, for example convex and linear problems. In deceptive cases, an inaccurate lower level optimum may result in an objective value better than true optimum at the upper level, which poses a severe challenge for ranking/selection strategies used within any optimization technique. In turn, this also makes the performance evaluation very difficult since the performance cannot be judged based on the objective values alone. While the area of bilevel (or more generally, multilevel) programming itself is not very new, most reports in this direction up until about a decade ago considered solving linear or at most quadratic problems at both levels. Correspondingly, the focus on was on development of exact methods to solve such problems. However, such methods typically require assumptions on mathematical properties, which may not always hold in practical applications. With increasing use of computer simulation-based evaluations in a number of disciplines in science and engineering, there is more need than ever to handle problems that are highly nonlinear or even black-box in nature. Metaheuristic algorithms, such as evolutionary algorithms are more suited to this emerging paradigm. The foray of evolutionary algorithms in bilevel programming is relatively recent and there remains scope of substantial development in the field in terms of addressing the aforementioned challenges. The work presented in this thesis is directed towards improving evolutionary techniques to enable them solve generic bilevel problems more accurately using lower number of function evaluations compared to the existing methods. Three key approaches are investigated towards accomplishing this: (a) e active hybridization of global and local search methods during dierent stages of the overall search; (b) use of surrogate models to guide the search using approximations in lieu of true function evaluations, and (c) use of a non-nested re-formulation of the problem. While most of the work is focused on single-objective problems, preliminary studies are also presented on multi-objective bilevel problems. The performance of the proposed approaches is evaluated on a comprehensive suite of mathematical test problems available in the literature, as well as some practical problems. The proposed approaches are observed to achieve a favourable balance between accuracy and computational expense for solving bilevel optimization problems, and thus exhibit suitability for use in real-life applications.
Persistent link to this record
Link to Publisher Version
Link to Open Access Version
Additional Link
Author(s)
Islam, Md Monjurul
Supervisor(s)
Ray, Tapabrata
Singh, Hemant Kumar
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
2018
Resource Type
Thesis
Degree Type
PhD Doctorate
UNSW Faculty
Files
download public version.pdf 1.23 MB Adobe Portable Document Format
Related dataset(s)