Gordon, T.G.W.; (2005) Exploiting development to enhance the scalability of hardware evolution. Doctoral thesis , University of London.

PDF U592084.pdf Download (13MB)

Abstract

Evolutionary algorithms do not scale well to the large, complex circuit design problems typical of the real world. Although techniques based on traditional design decomposition have been proposed to enhance hardware evolution's scalability, they often rely on traditional domain knowledge that may not be appropriate for evolutionary search and might limit evolution's opportunity to innovate. It has been proposed that reliance on such knowledge can be avoided by introducing a model of biological development to the evolutionary algorithm, but this approach has not yet achieved its potential. Prior demonstrations of how development can enhance scalability used toy problems that are not indicative of evolving hardware. Prior attempts to apply development to hardware evolution have rarely been successful and have never explored its effect on scalability in detail. This thesis demonstrates that development can enhance scalability in hardware evolution, primarily through a statistical comparison of hardware evolution's performance with and without development using circuit design problems of various sizes. This is reinforced by proposing and demonstrating three key mechanisms that development uses to enhance scalability: the creation of modules, the reuse of modules, and the discovery of design abstractions. The thesis includes several minor contributions: hardware is evolved using a common reconfigurable architecture at a lower level of abstraction than reported elsewhere. It is shown that this can allow evolution to exploit the architecture more efficiently and perhaps search more effectively. Also the benefits of several features of developmental models are explored through the biases they impose on the evolutionary search. Features that are explored include the type of environmental context development uses and the constraints on symmetry and information transmission they impose, genetic operators that may improve the robustness of gene networks, and how development is mapped to hardware. Also performance is compared against contemporary developmental models.

Download activity - last month

Export as JSONCSVXML

Download activity - last 12 months

Export as CSVJSONXML

Downloads by country - last 12 months

Export as XMLJSONCSV

Archive Staff Only

View Item