Research Of Ensemble Incremental Learning Based On RBF

The traditional neural network can learn only one time, can not achieve second time learning based on the first study. In the real life could not get all training data for the neural network, so many scholars proposed incremental learning of the neural network. The paper studies new class incremental learning methods, which can learn new knowledge from new sample, based on keeping have learned konwledge.In order to solve plasticity and stability problems of the incremental learning, this paper uses integrated method to incremental learning. Aim at network growing too fast, this paper put forward integrating incremental learning method based on RBF(radial basis function). A RBF sub network will be trained every time when learning a new class, then put the sub network into the integrated system to form a large system. Useing the nearest center method to determine the winning sub network, which can determine the output results and has nothing to do with the other non winning sub network.In order to solve the different sub network center similar problems in the learning process,this paper introduces the SOM(self-organizing map) neural network, using trained codebook vectors to construct PNN(probabilistic neural network), and by maximum probability determining the winning sub network. The combination methodwas put out base on nearest center method and maximum probability method. This paper solves the problem of network growing too fast, because one sub network been put into integerted network. Winning sub network can learn one class, and also can learn many classes, thus avoiding tediouslearning of one by one class.Finally, useing Statlog(Landsat Satellite) datasetof UCI machine learning datasetto shows that three methods are validity, which can increment learn not more than one class based on keeping have learned knowledge.