Method and apparatus for training a neural network using evolutionary

Abstract

A method and apparatus for training neural networks using evolutionary programming. A network is adjusted to operate in a weighted configuration defined by a set of weight values and a plurality of training patterns are input to the network to generate evaluations of the training patterns as network outputs. Each evaluation is compared to a desired output to obtain a corresponding error. From all of the errors, an overall error value corresponding to the set of weight values is determined. The above steps are repeated with different weighted configurations to obtain a plurality of overall error values. Then, for each set of weight values, a score is determined by selecting error comparison values from a predetermined variable probability distribution and comparing them to the corresponding overall error value. A predetermined number of the sets of weight values determined to have the best scores are selected and copies are made. The copies are mutated by adding random numbers to their weights and the above steps are repeated with the best sets and the mutated copies defining the weighted configurations. This procedure is repeated until the overall error values diminish to below an acceptable threshold. The random numbers added to the weight values of copies are obtained from a continuous random distribution of numbers having zero mean and variance determined such that it would be expected to converge to zero as the different sets of weight values in successive iterations converge toward sets of weight values yielding the desired neural network performance.