BP网络学习算法改进方案的探析

Ｂ Ｐ 网络是目前应用最为广泛的神经网络，但由于 Ｂ Ｐ 网络采用的是梯度下降法，这就不可避免地会出现网络学习收敛速度慢及容易陷入局部极小等问题。此外，学习因子和惯性因子选取对网络的收敛有较大的影响，但它们只能凭 经验确定。因此， Ｂ Ｐ 网络的有 效应用受到 了一定的限 制。针对 Ｂ Ｐ网络的学习收敛速度慢这一主要缺陷，对改进激励函数、改进 误差函数、改进一般化误差、学 习因子和惯性因子的自适应调整、梯度下降法与直接搜索法相结合、全局 优化、非线性优化、拓仆修正算 法等多种改进方案按改进原理进行了分类综述，并在此基础上，通过解决 Ｘ Ｏ Ｒ 问题的仿真实验对部分改进方案进行实验性评价，分析说明了它们的优劣和特点。

The Study of Improvedment Plans of BP Neural Network Learning Algorithm

Back propagation neural network is one kind of neural networks with most wide application. Because of adopting the gradient method by BP neural network, the problems including slowly learning convergent velocity and easily converging to local minimum can not be avoided. In addition, the selection of learning factor and inertial factor affects the convergence of BP neural network, which are usually determined by experiences. Therefore the effective application of BP neural network is limited. Being aimed at the main defects of slowly learning convergent velocity of the BP neural network, lots of improved plans such as improving excitation function, improving error function, improving general error, the self-adaptation adjustment of learning factor and inertial factor, the combination of the gradient method and the direct searching method, general optimization, non-linear optimization, and topology correcting algorithm are classified and summarized beased on the principles of the improvements. Parts of them are experimentally evaluated by simulated experiment of solving XOR problem, and their virtues and defects are analyzed and explained.