A Samples Self-learning System Of Neural Network Based On Clustering

The generalization ability of neural network has a great influence on its usability. How to improve this ability is getting more and more attention. There is a common situation that network will lead great error in actual application even it has been traind with high accuracy. In practical application, people always pay more attention on network's reflection ability about unknown inputs and outputs instead of the modeling capacity about known samples.Several common methods used to improve generalization ability are introduced and analyzed in this paper. In view of these methods'drawbacks, samples self-learning method of neural network based on clustering has been proposed to improve the completeness of samples set. The method of clustering is introduced in this system to analyze network's actual inputs in system's practical use. On the premise of normal real-time prediction of network, new representative data are found out by clustering. Analyze those representative data and get possible new samples. After some treatment, new samples composed of those data can be added to samples set to retrain network. By using this method, network will be more suitable for practical situation and meanwhile, prediction accuracy can also be improved.Take Visual Studio 2005 as platform of development to test samples self-learning method of neural network based on clustering. Taking this method into fire risk monitoring system, experiment results show that compared with the system that uses common algorithm, new samples can be collected according to actual conditions to improve generalization ability of neural network and reduce prediction error by using this optimized method.