A Multi-agent Approach to the Adaptation of Migration Topology in Island Model Evolutionary Algorithms

The Island Model (IM) is an important multi-population approach for improving the performance of Evolutionary Algorithms (EAs) when solving complex problems. One of the critical parameters for defining a suitable IM is the migration topology, which determines the migratory flows between sub-populations of the model. Despite the importance of this parameter, the majority of topologies tend to be naive and fail to take into account the underlying optimization process. To deal with the problem of adequately setting a migration topology, we propose an approach in which the Island Model is transformed into a Multi-Agent System (MAS) capable of learning and adapting the inter-island links based on the experience obtained during the evolutionary process. This approach is compared against two other traditional topologies applied to island versions of two different EAs, and to their usual implementations. The preliminary results strongly suggest an advantage of the IM versions over the original algorithms, and the competitiveness of the proposed approach.