An evolving neural network with the amorphous structure

Evolutionary transparent cellular neural networks (ETCNNs) are designed and developed based on non-uniform cellular automata (CA). The phylogenetic level of the ETCNNs shows fractal numbers of synaptic layers in their topological morphology, and an extended backpropagation (BP) learning algorithm is induced through a transparent signal propagation mechanism at the epigenetic level. Therefore, an ETCNN deserves to be considered as a new paradigm of complex adaptive systems in artificial life. In simulations, ETCNNs have been applied to various function approximation problems with successful results for a relatively large number of training sets, and have shown convergent behaviors, both in the neural network structure and in the synaptic weights.