Abstract
Meta-evolution, the ability to generate novel ways of generating novelty, is one of the major goals of Artificial Life research. Biological systems can be seen to perform open-ended meta-evolution, but unfortunately this has proved very difficult to replicate within computer programs. This thesis defines a theoretical framework for thinking about the issues in this area and charting out possible solutions. We use this framework to take some tentative steps towards meta-evolutionary algorithms.
We begin with a review of how meta-evolution has developed historically, through different novelty-generation algorithms. We describe previous algorithms in terms of their biological model and their computational model, to distinguish their biological inspiration from their computational implementation details. We conclude that the route to meta-evolution lies in enriching the biological models of current algorithms, rather than their computational ones.
We use the theoretical idea of embodiment to analyse both biological and com- putational systems, and decompose the problem of achieving meta-evolution. We present a new definition of embodiment, allowing the concept to be applied to biological systems. We conclude that biological systems achieve meta-evolution by their embodiment within the physical world. We notice that current computa- tional systems have poor embodiment within their virtual worlds. This leads to the hypothesis that improving the embodiment of computational systems, within virtual worlds, is a route to improving their meta-evolution.
In discussing how embodiment can be realised in virtual worlds, we use Artificial Chemistries as a language in which to program embodiment. We present a new definition of Artificial Chemistries, mapping their components onto our definition of embodiment. We end by presenting a new Artificial Chemistry, GraphMol, designed with embodiment in mind. We use GraphMol to embody the process of copying a string, and analyse its meta-evolution via a range of different experiments
Full thesis : PDF 30.2MB