Artificial Neural Network Models And Applications In Combinatorial Optimization And Information Processing

In this paper we present some novel neural network algorithms for combinatorial optimization and some applications of information processing. An artificial neural networt is an information processing system that has certain performance characteristics in common with biological neural networks. Neural network possesses many of advantages : simple structure, easy implementation in hardware, the basic parallel computational architecture, etc. In the past decade, a lot of neural networks algorithms were proposed for combinatorial optimization.A brief overview about neural networks and a review of approximate algorithms for combinatorial optimizations are given. We turn the model of decreasing energy functions about the 8-queens problem in combinatorial optimization, which is based on Hopfield neural networks models. By simulating at computer , we acquire all solutions about the 8-queens problem.In this paper we gain the weight between a city and its converging and non-converging nodes when the elastic net gets a stability state. We also introduce a gradient ascent learning algorithm of the original elastic net and the optimal elastic net for the TSP. Once the networks get stuck in local minima, the gradient ascent algorithm attempts to fill up the valley by modifying parameters in a gradient ascent direction of the energy function. Two phase are repeated until the algorithm gets out of local minima and produces the shorted or better tour through cities. We test the algorithm on a set of TSP. For all instances, the algorithm is showed to be capable of escaping from the local minima and producing more meaningful tour than the original elastic net or the optimal elastic net.This thesis also presents a modified Kohonen Self-Organizing Map(SOM) for TSP. There are many types of SOM algorithms to solve the TSP found in the literature, whereas the purpose of this paper is to looks for the incorporation of an efficient initialization methods and the definition of a parameters adaptation law to achieve betterresults and a faster convergence. Aspects of parameters adaptation,' selecting the number of nodes of neurons, index of winner neurons and effect of the initial ordering of the cities, as well as the initial synaptic weights architecture of the modified SOM algorithm are discussed. The complexity of the modified SOM algorithm is analyzed. Some experiments are performed by the modified SOM algorithm for the same set of test problems, and solutions are compared to the solutions of the existing heuristic. The results show that the modified SOM algorithm produced better solutions than those of the existing heuristic.In the last part of the paper, the method of predicting of BP neural networks is used for SARS epidemic to improve the existing computational methods, and better accuracy of prediction is achieved. A suitable momentum term is added to BP algorithm to accelerate the convergence speed. An online prediction strategy is applied to monitor the training and predicting process. We have achieved a series of predicting results of SARS epidemic about BEIJING and SHANXI in China. On the other hand, We study quantitative eddy detection by scanning drive magnetism in changed frequency to collect signal of crack. We give the multi-layer back propagation network model, and simulate experiment date by computer. The simulating result indicate that the algorithm is rapid and the result is precision.