Evolution of neural net architectures by a hierarchical grammar-based genetic system

Publication Type:

Conference Paper

Source:

ICANN'94: International Conference on Artificial Neural Networks, Springer, Sorrento, Italy (1994)

Abstract:

We present a hierarchically structured system for the evolution of connectionist systems. Our approach is exemplified by evolution paradigms for neural network topologies and weights. Our descriptions of a network's connectivity are based on context-free grammars which are used to characterize signal flow from input to output neurons. Evolution of a simple task gives a first impression about the capabilities of this approach.