The increasing availability of parallel computing architecture provides an opportunity to solve more complex problems by using fine-grained parallel scheme. The previous work on fine-grained parallel evolutionary algorithm (FGPEA) has verified that FGPEA is better than sequential evolutionary algorithm (EA) in the speed of EA and the quality of final solutions.
This thesis proposes two efficient migration methods and a complete binary tree topology in FGPEA.
The design of effective EA is to obtain a proper balance between exploration and exploitation. The spread rate of the best individual is one way to control the balance. This thesis proposes a complete binary tree topology which slows down the spread rate. Since tree topology has good exploration and poor exploitation, it has a good performance in solving such as heavily constrained problems.
And this thesis proposes two migration methods in FGPEA. The migration methods preserve that a superior individual takes almost all the subpopulations. The one is the restriction of the migration and the other is the modified individual migration.
To test the performance of the proposed migration methods and tree topology, CrayT3E was used as a parallel computing architecture and five unconstrained numerical optimization problems, five constrained numerical optimization problems and five combinatorial optimization problems were evaluated for global minima. The results indicate that FGPEAs using the proposed migration methods have better performance in unconstrained numerical optimization problems and constrained numerical optimization problems and FGPEA with tree topology has better performance on heavily constrained numerical optimization problems.
