Research Of Neural Network Data Fusion Based On Improvement PSO

Data fusion, as a new forefront discipline combining traditional disciplines with newly developed engineering areas, has been extensively applied in many other areas than military ones to which it has been originally applied. Artificial Neural Network (ANN) is the most active branches of computational intelligence and machine learning, it research mechanism about human intelligence activities based on the physical structure of the human brain. The research of data fusion based on the ANN is hot issue these years. This thesis reviews the development of data fusion and specific application of the back propagation (BP) neural network in the data fusion. Although the BP neural networks have been widely used, it also has some shortcomings, just like the long time require in net training and easily to be plunged into local optima.This thesis researches the Neural Network for Data Fusion use the improved PSO as training algorithm. It could speed up the convergence and improve the accuracy of nonlinear optimization. PSO, which stems from the simulation of birds flock's looking for food, has been paid attention and researched wildly. Relative to Genetic Algorithm, PSO can be implemented easily because it hasn't crossover and mutation operation and many parameters to be adjusted. Meanwhile, PSO's convergence speed is generally faster than Genetic Algorithm.A new PSO, called evolutionary grad included niche PSO , which is presented based on the "prematurity" in global optimization by the normal particle swarm optimization. The new algorithm maintain particle swarm's diversity by dividing it to small niches, the evolutionary gradient speed up the convergence, while enhancing the capabilities of jumping out of the local optimal solution. In recent years, there have been increasing interests in using PSO in dealing with ANN optimization. In this thesis, the new PSO is introduced to the ANN weight optimization. The new PSO is able to search highly dimensional weight space particularly along every dimension and make convergence precision more accurate.After the test of five typical benchmark functions, the results showed that compared to the basic PSO, the performance of improved PSO is significant. Performance on all test functions got much improvement, especially on the multimodal functions.