Brain Building with GenNets

This paper shows how the Genetic Algorithm can be used to evolve time-dependent neural network modules. These functional and control modules can be assembled into increasingly complex structures, thus allowing the possibility of “brain building”, using a new methodology known as Genetic Programming. The major advantage of GenNets (Genetically Programmed Neural Nets) as compared with traditional neural net learning techniques is that they need only a time-independent scalar performance measure (of the dynamical process they are controlling) to steer their evolution. GenNet modules can be evolved to perform incredible varieties of dynamic and control behaviours. Future technologies will allow Genetic Programming to be performed directly in hardware in real time, thus introducing the concept of the Darwin Machine. To illustrate the power of Genetic Programming, this paper shows how a 12 neuron, fully (self) connected, time dependent GenNet was evolved which taught a pair of stick legs to walk.