Leaning of neural networks using extended genetic algorithm with neutral mutation-application to ill-posed medical diagnostic system [for leaning read learning]

For most multimodal deceptive functions it is difficult to train networks with genetic algorithms. This is because a GA changes its search direction based on the building block hypothesis. For such problems, we propose a learning method of neural networks using an extended genetic algorithm with neutral mutations. Since the extended GA with neutral mutations adopts a redundant string representation, most genetic operations become neutral for an objective function. The proposed method is also an effective approach to nonstationary classification problems. It is observed that it overcomes the difficulty by the emergent property of finding deceptive hyperplanes and escaping the population from them by using the large genetic transitions to their complements. The learning ability of the proposed method is examined successfully with an ill-posed medical diagnostic system for hepatobiliary disorders.