5 Comments on Previous Distributed Models

The goals of distributed models and techniques such as niching is to maintain a population that occupies different areas of the search space. There is, however, no guarantee that these methods will perform as expected. Methods may fail to apply the right amount or kind of pressure to keep a population distributed or allow fitness improvement. The previous work described in Section 7.1 highlighted the variations of models for genetic algorithms and the more specific focus on efficiency in genetic programming models.

Distributed models commonly use panmictic-like subpopulations that perform the standard evolutionary algorithm within each island, migrating fit individuals between islands to replace the worse fit. Fitness-based migrant selection and insertion can be considered an added selection pressure, especially when migrants are copied. However, deme models can reduce selection pressure by only conducting selection within demes. Previous studies also highlighted the importance of connectedness of the topology, rather than the topology itself [Cantú-Paz, 1999]. As diversity controlling methods are, in a sense, largely countering the effects of selection, the type of selection pressure a model defines is likely to be extremely important in determining its effectiveness.

Lastly, in distributed populations, migrants are likely to be, at least initially, genetically different from their new subpopulation. This is particularly true if islands converge toward different regions of the search space. However, previous work does not typically address the actual role of migrants. It is questionable if migrants effectively add new genetic material, takeover the new subpopulation or are selected at all in their new subpopulation. The following study examines the role of dissimilar individuals, relevant to both standard populations and distributed models.