Convergence Analysis of the Learning Algorithms for Multi-Valued Neurons

In this paper, based on the idea of gradient descent, the learning algorithms are derived for the multi-valued neurons (MVNs) which assure the minimum of the estimation errors. In addition, finite convergence of the proposed learning algorithms is discussed. As an extension of the MVNs, it is finally proved that such kind of learning algorithms can also be effectively applied to universal binary neurons and multi-valued neurons with periodic activation factions, and the relating algorithms converge in finite steps.