A genetic method for optimization of asynchronous random neural networks and its application to action control

A genetic method is proposed to optimize random neural networks composed of asynchronous thresholding neural units. Each unit belongs to one of three categories, input units, hidden units and output units, and any kinds of connections among units including feedforward, feedback and mutual connections are allowable except connections to input units. Several virtual living things whose genotype are the connections among neural units are randomly generated, and generation iteration is repeated in order to optimize them. In the generation iteration, individuals adequate to a given problem make their children and inferior ones are removed from the population. Optimized neural networks are obtained as evolved individuals. An action control problem for a computer game is treated as an application of this method.