Comparisons of four learning algorithms for training the multilayer feedforward neural networks with hard-limiting neurons

In this paper, two kinds of learning algorithms that have been developed for training multilayer feedforward neural networks with hard-limiting neurons are reviewed. For the modified backpropagation algorithms, their numerical performances of convergence speed and training efficiency are compared; for the architecture generating methods, the architecture sizes of the neural networks generated are compared and their generalization ability are discussed. For any given application problem, some criteria for selecting a suitable training algorithm are also discussed