Evolvability, Complexity and Scalability of Cellular Evolutionary and Developmental Systems

Abstract

Man-made systems, such as supercomputers and software, IT-infrastructures and networks of any kind, are continuously growing in size and complexity. As conventional top-down engineering techniques may have reached the limit of applicability, biological organisms have been able to evolve increasing levels of complexity. Such inherent biological complexity may be said to be open ended or unbounded. This is a result of a bottom-up emergent process which produced an astounding diversity of living organisms with remarkable abilities, such as adaptation to different environments or perturbations, and reproduction, being able to survive. Even though a lot of work has been done towards a synthesis, it is still not completely clear how to unleash the full potential of biological properties into artificial systems. This thesis tackles the problem of better understanding the developmental process between genotype and phenotype and the evolution of complex systems made of large sets of elements interacting locally and giving rise to collective behaviour. In a traditional Evolutionary Algorithm approach, the genotype maps to a phenotype directly, i.e. direct 1-to-1 encoding. If one wants to scale-up the phenotype complexity, indirect encodings, e.g. developmental or generative mappings, are a necessity. In the experimental work, the chosen computational platform is Cellular Automata (CA). The biological metaphor can be applied to the physical structure similarities between artificial cellular systems and biological multi-cellular organisms. A CA can be considered as a developing organism, where the genome specification and the gene regulation information control the growth and differentiation of the cells. Such a dynamic developmental system can show adaptation, self-modification, plasticity, and self-replication properties. In this thesis, four challenges of designing Evolutionary and Developmental (EvoDevo) systems are identified and studied further, each related to a specific research question: RQ1. What kind of information must be present in the genome in order to produce computation in any of the computational classes? RQ2. How to quantify developmental complexity, i.e. emergent phenotypic complexity? RQ3. Do genome parameters give any information on the evolvability of the system? And if yes, can genome information be used to guide evolutionary search in favourable areas of the search space where the wanted emergent behaviour is more likely to be found? RQ4. How can scalability of artificial EvoDevo systems be improved towards achieving systems that can fully unleash their inherent complexity, potentially at the levels of complexity found in nature? The results in this thesis show that abstract measures of phenotypic complexity may be suited to characterize emergent cellular organisms. Genome information may be related to emergent complexity and such knowledge may be used to guide evolutionary search. For scaled-up systems, it may be possible to allow indirect encodings with genome representation growth. A framework for the evolutionary growth of genomes is proposed.