Artificial Development and Evolution using Common Developmental Genomes

Abstract

Developmental biology seeks to understand how organisms are constructed. Development is a set of very complex processes involving a coded program as the organisms’ genome. This genome describes how to build the organism, but not how the organism will look like. The zygote (the initial single cell), will eventually develop into a trillion cell organism. This extraordinary phenomenon has been an inspiration to the world of Computer Science and Artificial Life and has led to the creation of Artificial Embryogeny (AE). Artificial Embryogeny is a sub-discipline of evolutionary computation (EC) in which a phenotype undergoes a developmental phase. The number of AE systems currently being developed investigate mainly how principal biological processes and mechanisms can be exploited in the artificial world. One approach that utilize the phase of biological development in artificial systems is called Artificial Development (AD) where the genotype (genetic representation) contain a similar set of instructions - as in the biological organisms case - called generative program or developmental encoding. Therefore, the process of development comprise to actually execute those instructions and deal with the highly parallel interactions between them and the structure they create. On the other hand, nature uses the same fundamental machinery and almost the same genetic information to create vastly different creatures. A study reveals that about 99% of mouse genomes have direct counterparts in humans with cats having 90% of their homologous genes identical to humans. How is it possible for nature to use a vast majority of the same genetic representation in the DNA but still be able to develop such distant species? It was found that a common regulator gene can control the formation of many of the internal organs in both nematodes and vertebrates. Therefore, the very same gene can initiate the process of formation and define its outcome, for example, an intestine or a muscle cell. This thesis investigates how to design an Artificial Embryogeny by using the same genetic information to develop a class of computational architectures or different computational architectures. The result of this investigation has given rise to the Common Developmental Genomes (CDG). The computational architectures targeted, have a common characteristic of being sparsely-connected networks, with each node acting as a simple computational unit. Such computational architectures are cellular automata and boolean networks, artificial neural networks and cellular neural networks. The approach followed includes the following steps: a. investigate which architectures are suitable for development with such a model, b. describe a common developmental approach that can handle the targeted architectures, c. define how genetic information can be exploited by the developmental process, so as to develop these architectures and d. identify a suitable genome representation to ensure that different structures can be developed and achieved. The target architectures chosen throughout this thesis were cellular automata and random boolean networks. The reason for choosing those particular architectures is that they have similar structural and functional properties; in addition, random boolean networks are considered a generalization of cellular automata. Core work and design principles are given in Paper I. Though this paper it was able to show how genetic information can be represented and how targeted architectures can be integrated in the genotype. It is also shown how target architectures can be evolved and different structures achieved. Paper II studies the ability of CDG to evolve a simple financial market model in problems of varying complexity. CDG was shown to evolve better for some architecture sizes. In addition, CDGs evolvability were studied in case of limited resources (Paper III) with very promising results in certain cases. Paper IV focuses on how genetic operators affect evolution of CDG and studies their developmental dynamics under more complex and random environments. Paper V studies the ability of CDG to adapt when the target goal changes over evolutionary time. CDG were able to find very good solutions with rather simplified structure than anticipated. Paper VI focuses on how CDG exploit the underlying architectures during development and build final structure (network morphology). It was shown that during evolution, CDG exploit a larger number of nodes/cells and manage to maintain only a few neutral and static cells/nodes of the final structure.