DYNAMICAL GENOMIC NETWORK APPLIED TO ARTIFICIAL NEUROGENESIS

Optimisation of the structure of artificial neural network using evolu tionary techniques has been investigated by a number of authors using various approaches. In this paper, we claim that such a process requir es the design of a complex neurogenesis model featuring a set of funda mental properties such as modularity and the possibility of complexity adaptation. Developmental and molecular biology might be an interesti ng source of inspiration for designing such powerful artificial neurog enesis systems allowing the generation of complex modular neural struc tures. This paper provides a description of a neurogenesis model based on a modelling of a natural genomic network and associated with an ev olutionary process. Experimental results demonstrate some basic capabi lities of the proposed neurogenesis model to produce multi-layered neu ral networks. An application to learning in the control of a mobile ro bot lead to unexpected results, giving hints for continuing the resear ch towards the automatic generation of more complex adaptive neural ne tworks.