Sensitivity Analysis And Application Of Orthogonal Weight Function Neural Network

The monograph"New Theory and Methods of Neural Network"presents a new type of artificial neural network and spline weight function neural network learning algorithm. This new artificial neural network algorithm overcomes the defect of traditional algorithm that the training weight is difficult to reflect the information of training samples, also avoids many other shortcomings like local minima and slow convergence. Orthogonal weight function neural network is the implementation of the weight functions neural networks, so it has the advantages of weight functions neural networks. During process of neural network training, if samples are affected by noise, the trained neural network weights and the output of the network will be changed. Sensitivity is used to analysis this kind of change and the impact to the neural network. Based on analysis of sensitivity of neural network, this paper proposes theoretical error of noise and approaching error of noise, deduces the relevant theoretical formula.In order to analyze sensitivity of orthogonal weight function neural network, this paper firstly describes some concepts including learning curve, weight function, neural network, attributes of orthogonal function and related analytical methods; Through orthogonal polynomial, samples are represented as weight of network. The coefficients of the network weights are calculated and specific sensitivity calculation formula is deducted while general calculation formula of sensitivity of the orthogonal function is applied to legendre orthogonal polynomial and first class chebyshev orthogonal polynomials. With the help of MATLAB, it is proved that theory put forward is correct and valid by the comparison and analysis of the results of simulating weight function and theoretical sensitivity.Finally, this paper designs a kind of classifier for intrusion detection. The theory of sensitivity is applied to the intrusion detection. By the way of analysis of sensitivity, test data will be filtered optimized, and detection rate of intrusion detection will be raised, which proves the correctness of this theory.