Variation and selection: An evolutionary model of learning in neural networks

In the present work, we study the emergence of nontrivial computational capabilities in networks competing against each other in an environment where possession of such capabilities is an advantage. Our approach is to simulate a variable population of network automata. In a way directly analogous to biological evolution, the population will converge, under the influence of selective pressure, to a group of automata that will be optimally suited for solving the task at hand. We visualize the approach as comprising two separate kinds of processes - a low-level 'performance' process and a higher level 'metaperformance' process.