Pseudo-Random Recurrent Artificial Neural Networks for Control

点击翻译按钮获取中文摘要

Tang 9626655 A novel concept, the pseudo-random recurrent artificial neural networks (ANN) will be used in control applications. The Pseudo-random recurrent ANN is a novel concept derived from previous work on pseudo-random interconnection networks. Instead of using a fully connected feedforward or recurrent network, a pseudo-random ANN is connected with a pseudo-random graph that can provide almost random connections of lots of neurons with a small number of links at each neuron. The idea is to use the large number of neurons to capture the non-linear nature of the control system; and yet the small number of links makes real-time implementation feasible. The underlying pseudo-random graph, known as Cayley graphs have been the densest known graphs which means they have the largest number of nodes for a given degree and diameter. By exploiting the node-symmetric property of these graphs, simple simultaneous broadcasting their signals to all other neurons can be developed. The result is a "brain-like" ANN that have a large number of neurons (hundreds and thousands) connected in an almost random manner with a relatively small number of connections. These pseudo-random ANNs will be implemented with reinforcement control strategies such as the backpropagation of utility and the heuristic dynamic programming for some existing non-linear control applications. The investigation of ANNs that do not require complete connections of neurons will pave the way for a new form of ANN that has more resemblance to its biological counterpart.