University of Sussex
Browse
Tomko,_Nicholas.pdf (1.77 MB)

Separating what is evaluated from what is selected in artificial evolution

Download (1.77 MB)
thesis
posted on 2023-06-08, 16:18 authored by Nicholas Tomko
In artificial evolution, selection and evaluation are separate and distinct steps. This distinction is rather different in natural evolution, where fitness (corresponding to evaluation) is a direct consequence of selection rather than a precursor to it. This thesis presents a new way of thinking about artificial evolution that separates evaluation and selection and consequently opens up the space of potential evolutionary algorithms beyond the limitations imposed by ignoring this distinction. In Part I of the thesis we explore how varying the level of evaluation and selection impacts evolution. Using novel genetic algorithms (GAs) we show how group level evaluation allows evolution to find solutions to problems that require niching or a division of labour amongst component parts, something that cannot be accomplished using a standard GA. One of the inspirations for testing GAs with group-level evaluation was recent research into bacterial evolution which shows in bacterial colonies, distinguishing between the individual and group is very difficult because of the symbiotic relationship between different bacteria. We find that depending on the task it sometimes makes sense to select the individual while in other cases simply selecting groups is the best choice. Finally, we present a method for evolving the group size in these types of GAs that has the benefit of avoiding the need to know the optimal division of labour ahead of time. In Part II we move away from studying the relationship between evaluation and selection to show how our novel view of evolution can be used to develop GAs that implement horizontal gene transfer which was again inspired by looking at bacterial evolution. By testing these GAs on a variety of different tasks we show how this promiscuous gene swapping is often beneficial to evolution because it can reduce the probability of the population getting stuck on a sub-optimal solution. The thesis demonstrates the benefits of of looking at artificial evolution in terms of both evaluation and selection when it comes to algorithm development, and thus provides the GA community with a new context in which they can choose different algorithms appropriate to different tasks.

History

File Version

  • Published version

Pages

172.0

Department affiliated with

  • Informatics Theses

Qualification level

  • doctoral

Qualification name

  • phd

Language

  • eng

Institution

University of Sussex

Full text available

  • Yes

Legacy Posted Date

2013-11-18

Usage metrics

    University of Sussex (Theses)

    Categories

    No categories selected

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC