基于知识的阶层型神经网络结构及参数的一种确定方法

神经网络具有优秀的学习能力，但神经网络的权值及阈值却无法解释与理解，给进一步的应用带来了困难。国内外学者就这一问题进行了各种探讨，研究怎样从神经网络中抽取规则，但算法较复杂，规则的可理解性较差。从不同的视角出发，提出一种从知识的角度来考察阶层型神经网络的结构及参数的思路。利用从样本数据中获得的知识（模糊规则），来确定网络的大小，即中间层的结点数目，以及网络的参数，即网络的双重及结点的阈值。该方法的特点是不用精简网络的结构，也不用改变网络以往的BP学习算法。按照这种方法构造出的神经网络，即使不学习，其输出也会大致地跟踪样本，网络的学习时间将会缩短。与此同时，网络的参数，即权重及阈值的意义可以解释，为直接从神经网络中提取知识提供了依据。仿真算例验证了所提出的方法的有效性。

A Method to Determine the Structure and Parameters of BP Neural Network from Knowledge

Neural networks have been widely used in many fields of science and engineering for its high level learning abilities But the further application is difficult because the weights and the thresholds of the neural network cannot be explained and understood at all Many research works on extracting rules from the neural network have been carried out for these issues However, the methods remain more complicated themselves and have the inferior understandabilities for extracted rules To overcome the drawbacks of the previous methods, a new method for determining the structure and the parameters of a general BP neural network from the knowledge is proposed Using the knowledge derived from the samples, the size and the parameters of the network can be determined The advantage of the proposed method is both the structure of the network and the traditional BP learning algorithm employed in the network need not to be changed Using this approuch, the network can approximately track the outputs without learning As a result, the learning time of the network can be shortened to some degree, and the significance of the weights and the thresholds can then be explained This is very helpful to extract fuzzy rules from the neural network directly The simulation results prove the validity of the proposed method