ABSTRACT
Biologically realistic, neural-like systems were evolved in computer simulations to solve a variety of life-like dynamical problems. Problem-independent properties of the networks were found which suprisingly also prevail in real nervous systems. These include inequalities between inhibitory and excitatory parameters, and the tendency of neurons to cluster into feed-forward inhibitory and other biologically common circuits. A simple scheme for connecting the neurons, based on idealized axons and dendrites was found to have advantages over full connectivity, because of its relative sparcity in connections and parameters and superior scaling behavior with network size.
