Properties of Genetic Representations of Neural Architectures.

Genetic algorithms and related evolutionary techniques o er a promising approach for automatically exploring the design space of neural architectures for arti cial intelligence and cognitive modeling. Central to this process of evolutionary design of neural architectures (EDNA) is the choice of the representation scheme that is used to encode a neural architecture in the form of a gene string (genotype) and to decode a genotype into the corresponding neural architecture (phenotype). The representation scheme used not only constrains the class of neural architectures that are representable (evolvable) in the system, but also determines the e ciency and the time-space complexity of the evolutionary design procedure as a whole. This paper identi es and discusses a set of properties that can be used to characterize di erent representations used in EDNA and to design or select representations with the necessary properties for particular classes of applications.