The aim of this research is to investigate the use of a divide-and-conquer approach for solving continuous large-scale global optimization problems using evolutionary methods. The curse of dimensionality is a major hindrance to the efficient optimization of large-scale problems. Problem decomposition is an intuitive way of improving the scalability of optimization algorithms. However, the black-box nature of many real-world problems makes problem decomposition a difficult task due to the unknown interdependence pattern of the input variables. A good decomposition is one that minimizes the interdependence of the identified subproblems. In this thesis, we propose a differential grouping algorithm which is mathematically derived from the definition of partial separability, and is used to automatically identify independent components of black-box optimization problems with high accuracy. The subproblems formed by differential grouping are then optimized in a round-robin fashion using cooperative co-evolution. The advent of differential grouping makes it possible to estimate the contribution of individual components of a problem towards improving the overall solution quality. This is a precursor to the development of a contribution-based cooperative co-evolution that uses the estimated contribution information to allocate computational resources to components with a dominant effect on the overall solution quality. The existing large-scale benchmark problems confirm the efficacy of both contribution-based cooperative co-evolution as well as differential grouping. However, the shortcomings of existing benchmark problems limit the depth of our investigations on the proposed algorithms. To fill this gap, a set of challenging large-scale problems is proposed for analyzing the reliability and robustness of differential grouping and the contribution-based framework. In the light of the findings based on the new benchmark suite, a parameter-free differential grouping is proposed that outperforms its predecessor on the new and the old benchmark suites. An improved contribution-based framework with a better exploration/exploitation balance is also proposed that outperforms state-of-the-art algorithms on the new large-scale benchmark problems.
展开▼
