贝叶斯-高斯神经网络非线性系统辨识

Nonlinear system identification using a Bayesian Gaussian neural network

提出了一种贝叶斯－高斯神经网络用于非线性系统辨识。网络的拓扑结构和连接权值可以由训练样本直接得到；其输出是多信息合成的贝叶斯推理过程；其训练过程仅是用于学习输入因子，从而使训练速度远高于一般的反向传播神经网络。此外，网络的自组织过程根据某种最优准则实现，使得当系统动态特性发生漂移时，网络可以根据新的样本迅速优化其连接权值，而不需要重新进行学习，这对于实时应用是十分重要的。实际应用中，网络的工作过程可以是推广与自组织交替进行的过程。仿真研究表明，该网络的辨识效果可与经拓扑结构优选的反向传播网络相比，而其自组织能力则是权值不变的后者无法相比的。

A Bayesian Gaussian neural network is put forward for nonlinear system identification. This network has the following features: the topology, connection weights and thresholds can be set immediately when training samples are attainable; the output is a Bayesian reasoning process with the fusion of multi pieces of information; while the training process is only to optimize the input factors, which makes the training simpler than that of back propagation neural network. The self organizing ability of this network can also be easily achieved in an optimal way, so that when the dynamics of the system drifts, the network is able to swiftly optimize its connection weights and thresholds according to new training samples, which is very important for on line adaptation of network to systems. In applications, the working of this network can be an alternating process. Simulation shows that this network is comparable to the back propagation neural network in the prediction of a single input single output nonlinear system, and its self organizing ability outperforms the latter which has fixed connection weights and thresholds.