二进制映射神经网络的几何学习算法及其应用

A GEOMETRICAL LEARNING ALGORITHM FOR BINARY MAPPING NEURAL NETWORKS AND ITS APPLICATION

提出一种一般二进制映射问题的前馈网络学习算法 .给出一种求解超平面以几何分割训练点的新方法 ,不仅相应地构造了隐层神经网络 ,而且使得只需再构造一个输出层网络便可实现训练样本所描述的映射 .该算法在学习收敛速度方面优于 BP算法和 SC算法 ,对样本数据的分布和密集程度变化适应性强 。

A learning algorithm is presented for the general binary mapping problem. The algorithm employs a new method to compute hyper planes to divide the training points into distinct areas so that the hidden layer of neural networks is correspondingly constructed. This makes it possible that only one output layer needs construction to implement the mapping determined by the training points. The algorithm is better than BP and SC algorithms in respect of the convergence velocity, is more adaptive to the change of the distribution and density of the training points, and has a better error tolerant ability.