The Most Suitable Architecture of Hidden - Layer in BP Neural Networks for Function Approximation

Aim To determine the most suitable architecture of hidden-layer in an er- ror-back -propagation neural netwok for function approximation. Methods To train the typical multi-layer BP neural netwoks with different hidden layers and neurons to approximate a typical function, and analyze the results. Results The most suitable number of hidden layers in a BP neural network is 4,and the most suitable number of neurons in each hidden layer is between 10 to 20,and a BP neural network with a single hidden layer has the worst results.Conclusion For function approximation,the most suitable number of hidden layers in a BP neural network should be about 4, and there should be suitable number of neurons in each hidden layer.