Cooperative evolution by heterogeneous mix of differential evolution variants in a distributed framework for unconstrained global optimization

Abstract

Differential Evolution (DE) is a simple yet powerful stochastic real-parameter population based optimization algorithm. The conceptual and algorithmic simplicity, ease of implementation and high convergence characteristics of DE is attracting many researchers who are working on its newlinevarious aspects. Despite the active research on DE during the last decade, there are still many newlineopen problems. One such problem is the difficulty in choosing the right DE variant given an optimization problem, as the efficacy of DE variants differs. In an effort towards this direction, this thesis proposes to mix competitive DE variants with diverse characteristics in a distributed framework to co-operatively enhance the efficacy of the system as a whole with robust optimization characteristics. Consequently in an effort to identify competitive variants 14 classical DE variants and 14 Dynamic Differential Evolution (DDE) variants have been analyzed. The extensive empirical comparative performance analyses indentified four DE newlinevariants and their dynamic counterparts to be competitive. A theoretical analysis on the identified competitive DE variants suggested that mixing of perturbation schemes may contribute to exploration-exploitation balance. Subsequently, this thesis mixed all possible combinations of 2 out of 4 DE variants in each island restricting the island size to 4. This resulted in a class of newlinealgorithms called distributed mixed variants Differential Evolution (dmvDE). The thesis also attempted mixing DE and DDE variants in a distributed framework resulting in yet another class of algorithms called distributed mixed variants (Dynamic) Differential Evolution (dmvD2E). Both dmvDE and dmvD2E have been benchmarked against five state-of-the-art distributed DE algorithms. The simulation results showed that dmvDE and dmvD2E outperformed the compared algorithms in most of the cases displaying robust optimization characteristics.