STUDY OF IMPROVED BACKPROPAGATION ALGORITHMS WITH COUPLED NEURON THROUGH SIMULATION

Abstract

Due to generalization property and capability, to process large amount of information, neural networks have drawn attraction from various fields. Backpropagation (BP) is one of the most commonly used systematic method to train multilayered neural network. However, i.t requires long duration in learning with no guarantee that it will finally converge. BP algorithm fails to train a network with.nondifferntiable activation function. Coupled neuron possesses two kinds of •activation functions. There are three learning rules with three types of coupled neurons namely, analog coupled neuron, rule Ca-CNR), saturating linear coupled neuron rule CL-CNRD, and dig taL coupled neuron rule Cd-CNR.). These algorithms require much less time to converge to a solution compared to 1W algorithm. The possibility of not converging is too less. It is possible to train a network with nondifferentiable activation function. In this dissertation work a-CNR, sl-CNR, d-CNR and BP algorithms have been simulated. Higher convergence speed of -these algorithms compared to BP algorithm has been shown by considering three problems, XOR E:.)rO1.)lem, medircal d:i.aynosLice system, and (7,4.) cyclic code. a-CNR and s.l.-CNR algori:thms are also capable of approximating non--linear .functions. This capability has been shown by approximating a sine function.