Evolutionary Algorithm For Large-Scale Optimization

In many real-world decision-making processes, a decision is made by solving an optimization problem. However, the performance of an optimization algorithm deteriorates significantly with the size of the problems, more complex relationships among the variables, and non-standard properties of the fitness function. To deal with these issues, several Evolutionary Algorithms (EAs) based on the divide-and-conquer concept have been proposed. A popular approach in this area is Cooperative Co-evolution (CC) which usually divides a problem into a number of smaller sub-problems that can be optimized cooperatively. One of the challenging issues, in such problem solving, is the decomposition of the problems and computational tasks. Thus, to obtain a high-quality solution in the optimization phase, the decomposition must ensure that (i) each sub-problem contains a group of highly dependent variables and (ii) the interdependencies among sub-problems are minimal which is difficult in the current research context. The second challenge is designing an appropriate optimization algorithm suitable for decomposed problems. Existing optimization approaches use a single optimizer to solve all the sub-problems and place equal emphasis on them during the optimization process despite their characteristics being different and their contributions varying significantly. The main objective of this research is to develop an evolutionary framework capable of handling large and complex optimization problems. In this thesis, an algorithmic framework for solving Large-scale Global Optimization (LSGO) problems, which contains decomposition and optimization approaches, is proposed. Firstly, a set of problems that can be decomposed into a number of sub-problems in such a way that each variable will appear in only one sub-problem is considered. The objective of this method for decomposing large-scale unconstrained problems, called Enhanced Differential Grouping (EDG), is to observe the interactions between the decision variables and ensure that there are no interdependencies among the sub-problems. Then, for problems that cannot be decomposed without leaving some common variables in multiple sub-problems, a novel decomposition method, the objective of which is to minimize the number of decision variables shared among sub-problems, is proposed. These subproblems are solved during the optimization process based on a round-robin strategy in a CC framework. The next optimization algorithm is based on a CC strategy in which multiple algorithms work adaptively depending on their success in solving the sub-problems. This process incorporates three population-based algorithms, namely, Differential Evolution (DE), Covariance Matrix Adaptation Evolution Strategy (CMA-ES) and Particle Swarm Optimization (PSO), with a novel fuzzy rule-based adaptation scheme using them to solve the sub-problems during the optimization process. The effectiveness of each sub-problem and optimizer pair is assessed using two criteria, namely, the fitness improvement and population diversity. Finally, for possible reduction of computational duration using parallel processing, particularly with a Graphics Processing Unit (GPU), an adaptive parallel DE algorithm is implemented for solving an application problem. The proposed decomposition techniques, optimization process and overall algorithmic framework are analyzed by solving well-known benchmark suites, and their results compared with those of state-of-the-art algorithms. All the proposed algorithms achieve significant improvements over the others, with the first, second, third and fourth enhancing the fitness values by up to 99.68%, 99.11%, 99.99% and 99.00%, respectively, and the last being up to 374.7 times faster.

Grantor dc:publisher

UNSW, Sydney

Year dc:date

2020

Author dc:creator

• Radwan, Mohamed

• Evolutionary algorithms
• large-scale optimization
• cooperative co-evolution
• overlapping functions
• fuzzy logic
• EEG signals

Statement dc:rights

• open access
• CC BY 4.0
• free_to_read

Licence

• https://purl.org/coar/access_right/c_abf2
• https://creativecommons.org/licenses/by/4.0/

Language dc:language

en

Identifier

https://doi.org/10.26190/unsworks/30256

OAI identifier oai:identifier

oai:unsworks.library.unsw.edu.au:1959.4/102399