Tuning the structure and parameters of a neural network by a new network model based on genetic algorithms

In this paper, according to the disadvantage of the tradition method tuning the structure and parameter of neural network (i.e., network growth or network pruning), a new neural network model has been proposed. In new model, the hidden nodes selection layer is inserted between output layer and hidden layer of the traditional feedforward neural network, and the switch, whose state is off, the hidden node is selected, otherwise given up, is set to each link between the hidden layer and hidden nodes selection layer. Using an application example on the sunspot numbers, it is proved that the network using the proposed network model trained with the genetic algorithms to tune, is the best appropriate. By the comparison with tradition BP neural network, it is again proved that the prediction result, using the proposed neural network, is better and more stable than using the tradition BP neural network. So, the proposed method, tuning the structure and parameters of a neural network by using genetic algorithms, is Reasonable and effective.