基于多准则的多层模糊神经网络学习算法

A multi-criteria learning method for multi-layer fuzzy neural networks

模糊算子函数丢失信息量过大,并且在某些点不存在导数,由此导致在采用传统的误差平方和准则优化网络参数时,有些参数无法得到调整,而且网络容易陷入局部极小,甚至发散.本文提出了一种基于模糊熵准则和误差平方和准则的多准则多层模糊神经网络学习算法,在一定程度上克服了单准则学习算法的局限性.

In this paper, a new approach, the muhicriteria learning （MCL） algorithm based on a composite criterion including both fuzzy entropy and mean - squared error criterion, is proposed for training multi layer max - min neural networks. The new algorithm overcomes the limitations in mono - criterion learning, that is, the mean -squared error criterion will easily converging to a local minimum value and slow converging speed. Compared with the traditional fuzzy back -propagation （FBP） algorithm, it is found that the proposed MCL algorithm provides a faster learning speed and higher stability.