Optimality-Oriented Stabilization for Recurrent Neural Networks

Inspired by biological neuronal systems, artificial neural networks have demonstrated superior characteristics of learning, adaptation, classification and function-approximation. They have been successfully applied to many areas. However, conventional neural networks are feed-forward networks that move the information in only one direction, forward, from the inputs to the outputs. There are no cycles or loops in the network and no feedback connections are present. Therefore, there are some drawbacks in using those networks. Since the Hopfield’s paper (Hopfield, 1984), recurrent neural networks, also called dynamic neural networks, have emerged to be a highly effective methodology. Because these networks have feedback connections that represent dynamic memory, which make them more suitable for modeling biological intelligence than purely feedforward neural networks. In addition, within the structure of recurrent neural networks, they have bi-directional data flow. Therefore, they have been used to solve various difficult problems in many ABSTRACT