中文摘要：

为了进一步提高BP（Back Propagation）网络中LM（Levenberg-Marquardt）算法的收敛速度,针对传统LM算法通过误差矩阵求逆来更新网络参数的计算量较大的缺点,提出一种分治优化算法：基于矩阵分块的递归Cholesky分解算法.并对算法进行了时间复杂度的理论分析对比,最后通过Matlab软件实际仿真实验进行对比验证,理论分析与实验仿真基本吻合.结果表明：对于LMBP算法中更新网络参数值的计算,与传统的矩阵求逆方法以及已有的改进算法相比,基于矩阵分块的递归Cholesky分解算法计算量更少、运算速度更快.

英文摘要：

In order to improve the convergence speed of LM（Levenberg-Marquardt） algorithm based on BP（Back Propagation） neural network further,a divide and conquer optimization algorithm： the recursive Cholesky decomposition algorithm based on a partitioned matrix is proposed,directing against the high computation drawback of conventional LM algorithm which updates the network parameters through calculating the inverse one of error matrix. And the theoretical analysis with a comparison of algorithms time complexity is made,and a comparison verification of the practical simulation experiment by Matlab is given at last,which confirms that the experiment simulation dovetails with the theoretical analysis approximately. The fact turns out thatas for the calculation of updating the network parameters in LMBP algorithm,the recursive Cholesky decomposition algorithm based on matrix blocks is less computation required and faster,compared with the traditional and existing algorithms during updating the network parameters.