基于3种训练神经网络方法解决异或问题的研究

分别运用误差反向传播(error back propagation)算法、扩展卡尔曼滤波(extended kalman filter)和数值积分卡尔曼滤波(cubature kalman filter)算法对多层神经网络模型进行逐次状态估计,并将其用于解决异或的分类问题。从仿真实验结果来看,利用BP,EKF和CKF算法训练的神经网络的输出信号的均方误差曲线的收敛速度依次加快,这使得神经网络的实际输出值越来越逼近其期望输出值,同时针对异或问题3种算法都得到了良好的分类结果.

Three neural-network algorithms for XOR problems

Respectively,this paper uses the error back- propagation algorithm,the extended Kalman filter and the cubature Kalman filter to evaluate sequentially the state of the multi- layer neural network model,and applies those algorithms to solving the classification problem of XOR. The simulation shows that the convergence speed of the mean square error curve of the output to the neural network trained by EKF is faster than that of BP,the convergence speed of the mean square error curve of the output to the neural network trained by CKF is faster than that of EKF,which makes the real output of the neural network oriented more to the expected output,meanwhile all the three algorithms can gain good results of classification for the XOR problems.