关键词: 正交基函数，隐神经元，权值，学习算法，函数逼近

Abstract: Based on polynomial curve-fitting theory, an orthogonal basis feed-forward neural network is constructed.The model is adopted by a three-layer structure, where the hidden-layer neurons are activated by orthogonal polynomial functions. In view of the network, an algorithm is proposed that a kind of hidden layer activation function is the orthogonal polynomial and the number of neurons can be ctuickly determined. Through mathematical proof , the validity of the algorithm is theoretically proved. The algorithm is verified by computer simulations, comparing with the conventional BP algorithm. The results show that this algorithm not only breaks through the traditional BP neural network limitalions, such as slow convergence rate, optimal number of hidden neurons that difficult to be determined, but also can achievc higher precision. The effectiveness of the designed algorithm is validated.

Key words: Orthogonal base function, Hidden neuron, Weight, Learning algorithm, Function approximation