一种基于U-D分解卡尔曼滤波多层感知器学习算法

在研究多层感知器结构后，提出一种利用U－D分解卡尔曼滤波训练多层网的新算法．仿真结果表明：与BP算法比较，此算法有着学习速度快、数值稳定性好、对学习参数不敏感、能避免局部极小点等特点。

A learning algorithm based on the U-D factor Kalman filter for training multilayer percetrons

After studying the structure of multilayer perceptrons this article presents a new algorithm for train ing the multilayer network at each layer using the U-D factor Kalman filter.Simulation results mdicate that the algorithm has advantages of fast learning speed,good numeric stability,msensitivity of learning parameters,aviod ance of local minimum compared with the BP algorithm.