Repeated, morphological functionality, from limbs to leaves, is widespread in nature. Pattern formation in early embryo development has shed light on how and why the same genes are expressed in different locations or at different times. Practitioners working in evolutionary computation have long regarded nature's reuse of modular functionality with admiration. But repeating nature's trick has proven difficult. To date, no one has managed to evolve the design for a car, a house or a plane. Or indeed anything where the number of interdependent parts exposed to random mutation is large. It seems that while we can use evolutionary algorithms for search-based optimisation with great success, we cannot use them to tackle large, complex designs where functional reuse is essential.
This thesis argues that the modular functionality provided by gene reuse could play an important part in evolutionary computation being able to scale, and that by expressing subsets of genes in specific contexts, successive stages of phenotype configuration can be controlled by evolutionary search. We present a conceptual model of context-specific gene expression and show how a genome representation can hold many genes, only a few of which need be expressed in a solution. As genes are expressed in different contexts, their functional role in a solution changes. By allowing gene expression to discover phenotype solutions, evolutionary search can guide itself across multiple search domains.
The work here describes the design and implementation of a prototype system to demonstrates the above features and evolve genomes that are able to use gene expression to find and deploy solutions, permitting mechanisms of dynamic control to be discovered by evolutionary computation.
Full thesis :7.31MB