An Investigation Of Topologies and migration schemes for asynchronous distributed evolutionary algorithms

Distributed evolutionary algorithms are of increasing interest and importance for three main reasons: (i) a well designed distributed evolutionary algorithm (dEA) can outperform a `standard' EA in terms of reliability, solution quality, and speed; (ii) they can (of course) be implemented on parallel hardware, and hence combine efficient utilization of parallel resources with very fast and reliable optimization; (iii) parallel hardware resources are increasingly common. A dEA operates as separate evolving populations with occasional interaction between them via `migration'. A specific dEA is characterized by the topology and nature of these interactions. Although the field is sizeable, there is still relatively little exploration of the performance of alternative topologies and interaction mechanisms. In this paper we compare some simple, novel dEA topologies with the cube-based topology that forms the basis of Alba et al's GD-RCGA (a state of the art dEA). We find the best results (when topologies are compared on a like for like basis in terms of number of processors) emerge from a three-level tree-based topology.