| || || Nonconvex programming|
| || || Competitive island cooperative coevolution for real parameter global optimization|
Author: Bali, Kavitesh Kumar
Institution: University of the South Pacific.
Subject: Nonconvex programming, Genetic algorithms
Call No.: Pac T 57 .817 .B35 2015
Copyright:Over 80% of this thesis may be copied without the authors written permission
Abstract: Cooperative Coevolution (CC) is an evolutionary algorithm that features the divide-andconquer paradigm as an efficient technique for solving global optimization problems. A major difficulty associated with CC is the choice of a good decomposition strategy, especially when applied to problems that possess interacting decision variables. Identifying an efficient problem decomposition scheme is vital such that the interacting variables are captured and grouped together into separate subcomponents ensuring that the interdependence between them is kept to a minimum. Hence, finding an optimal decomposition for solving such problems is laborious and requires extensive empirical studies. This thesis presents a competitive island cooperative coevolution (CICC) algorithm in which unique suboptimal decompositions are implemented as islands that compete and collaborate during the optimization process. Each problem decomposition has features that may be beneficial at different stages of optimization. Initially, two and three-island instances of the CICC algorithm are deployed and tested on a wide variety of global optimization benchmark functions that feature different levels of separability and multimodality. A multi-island competitive cooperative coevolution (MICCC) is also proposed, as a successor to CICC, in which several different problem decompositions (implemented as islands) are given a chance to compete, collaborate and motivate other islands while converging to a common solution. Results show that enforcing competition with a wider pool of suboptimal decompositions can significantly improve the performance during the course of the optimization phase and can yield solutions with better quality when compared to standalone benchmark counterparts.