Back to search

University of Patras

Τεχνικές βελτίωσης της αποδοτικότητας αλγορίθμων επίλυσης προβλημάτων βελτιστοποίησης χωρίς περιορισμούς

Abstract

dc:description

Optimization is widely used in the field of mathematics but also in many other fields of science, such as business administration, economics, biology, etc. There are a lot of times that companies or the state act to optimize the use of information, technology, human resources management, personnel, services, and processes. The range of optimization applications makes it an area that research is still very important, despite the fact that many methods have already been proposed.The subject of this dissertation is to find the optimum solutions for unconstrained optimization problems using iterative methods. A major problem with many iterative optimization algoristhms is the fact that their success depends directly on the initial approach of the solution. In this dissertation, we propose techniques that solve this problem and present a new family of unconstrained optimization algorithms. In the presented techniques, well-known optimization algorithms are incorporated into the proposed techniques, a fact that enhances the well-known optimization algorithms' good features and improves their efficiency.The first chapter of this thesis is an introductory chapter in which basic ideas are presentedto better understand the next chapters.In the second chapter, we first give the main idea of algorithms of the present thesis, according to which known optimization methods are combined in such a way that thedirections given by them create a new direction on which the optimum is sought. Forthe implementation of the proposed idea we set out general principles and following theseprinciples the approximated problems of the given come out. The point used as the initialone for optimizing the given problem by an iterative method results from a preprocessingstep, the stages of which are as follows: (a) the construction of approximate problems whoseobjective functions are referred to as Approximated objective Functions (b) prioritizationof approximate problems and (c) the approximate problems' sequential and non-exhaustiveoptimization.In the same chapter, we present the proposed Lexicographic Optimization (LexOpt) algorithm, where the involved Approximated Objective Functions (AOFs)are extracted using one of the proposed principles, the cumulative principle. By comparing the results of the LexOpt algorithm with the results of the optimization methods it uses, different convergence behavior is observed, either in the computational cost or the number of convergence minima, or both. It is worth to notice that the results from the most of the test functions indicate that there is a way with which the LexOpt algorithm converges, starting from the same initial points, into a larger number of minima than the methods it uses, and even with such computational cost that leads us to an affirmative answer to the question whether there is a need to pursue further research into the LexOpt algorithm behavior.In the third chapter of the dissertation, we present a proposed technique named TaylorLexicographic Sequential Optimization (TLSO) algorithm, where the Approximated Objectivefunctions (AOFs) are extracted using another proposed principle, the sequential principle.This algorithm utilizes the information contained in the non-infinity terms of Taylor'sexpansion, avoiding their calculation. Like LexOpt, the TLSO algorithm, if applied by selectingrandom initial points, converge to more points than the utilized optimization method, although it starts from the same initial points. The main difference between LexOpt and TLSO algorithm is that in the LexOpt algorithm the objective functions of the preprocessing step problems are made by the formula of the objective function of the given problem and on the TLSO algorithm they are constructed by the approximation utilizing Taylor's expansion of the objective function of the given problem.The Trust Lexicographic Optimization (TrustLexOpt) algorithm technique has been proposedto reduce the computational cost of the algorithms already mentioned. The TrustLexOptalgorithm is based on the LexOpt algorithm and is presented in the fourth chapter of thisPh.D. dissertation. This algorithm reduces the number of problems used in the preprocessingstep, and it is suggested to use a combination of trust region and line search methods. TheTrustLexOpt algorithm uses a sub-problem, that is the minimum number of utilized subproblems, and combines trust region and line search methods in order to end up to theutilized initial point of the given optimization problem. In the same chapter, a method ofcalculating the trust region radius is proposed.The conclusions drawn in the fourth chapter and in the Ph.D. dissertation, in general, havebeen formulated by adopting the principle that when comparing two algorithms, the finallypreferred algorithm is the one that converges to the greatest number of minima. Furthermore, if two algorithms converge to the same number of minima, then we think that the algorithm with the smallest computational cost is the one that prevails over these two algorithms. From this point of view, we have concluded that there is at least one way to give priorities on the utilized elements that arise by applying the cumulative principle, where TrustLexOpt has generally excelled over TLSO and LexOpt algorithms. According to the numerical results of the Ph.D. dissertation, it is suggested that more study and research should be made on the appropriate combination of elements that if used by the TrustLexOpt algorithm it then improves the algorithm's performance. It is also suggested to further research why, when the TrustLexOpt algorithm uses the proposed trust region radius it gives better results compared to utilizing the usual trust region radius.In the last chapter of the PhD dissertation, Modified Lexicographic Optimization (MLX),Relaxation Lexicographic Optimization (RLX) and Modified Relaxation Lexicographic Optimization (MRLX) algorithms are proposed. The reason why we propose these algorithms is to study the effect of weights on parts of the objective functions used in the LexOpt algorithm preprocessing step. They are also proposed to study the effect of the fact that the objective functions of the problems of the preprocessing step have no common terms.The use of weights was implemented by using parameters that we called relaxationparameters. We then calculated the ratio of the absolute value of the greater to the smallesteigenvalue of the objective function of the given problem at the points resulting from thepreprocessing step. From a first study of the above, it appears that if relaxation parametersare used, a superiority of the LexOpt and MLX algorithms against RLX and MRLX is achieved.The results and the study presented in this chapter have not been published to date.In conclusion, the present thesis presents techniques for improving the efficiency of unconstrained optimization algorithms. While studying the numerical results of applying thesetechniques to well-known test functions, it has emerged that the way that priorities are givenin the preprocessing step of the algorithms plays an important role in the results of theproposed algorithms. Thus, giving priorities is considered an issue for further research. Onepoint that is also worth studying is the ideal number of steps that an iterative method shouldbe applied in the preprocessing step. A field for further research is the implementation of theother general principles proposed in the second chapter of this thesis. It is also proposed tostudy the behavior of the algorithms presented in this thesis with the dynamic redefinitionof the problems involved. Finally, an interesting research topic is the application of what isproposed in this dissertation for solving single-objective constrained optimization problems.

Degree

thesis:*
Grantor dc:publisher
University of Patras
Year dc:date
2018

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Nikolakakou, Georgia-Christina
  • Νικολακάκου, Γεωργία-Χριστίνα

Subjects

dc:subject × 14

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/44723

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Nikolakakou, Georgia-Christina; Νικολακάκου, Γεωργία-Χριστίνα. Τεχνικές βελτίωσης της αποδοτικότητας αλγορίθμων επίλυσης προβλημάτων βελτιστοποίησης χωρίς περιορισμούς. University of Patras, 2018. http://hdl.handle.net/10442/hedi/44723