【作者】 莫富强； 王浩； 姚宏亮；

【Author】 Mo Fuqiang, Wang Hao, Yao Hongliang(School of Computer and Information, Hefei University of Technology, Hefei 230009, China)

【机构】 合肥工业大学计算机与信息学院；

【摘要】 针对动态贝叶斯网络EM参数学习算法中计算量大和收敛速度慢问题,通过将大规模时序数据集划分为较小的数据块,提出一种基于部分E步的加速算法(DA-EM)。DA-EM在块间循环迭代,每一次迭代执行部分E步,增量式地更新似然函数和网络参数,以缩短E步执行时间。该算法在不损失精度情况下能显著加快EM算法的收敛速度,有效地提高EM算法的计算效率。实验表明,DA-EM算法有较高的计算效率。更多还原

【Abstract】 The algorithm based on EM for DBN parameter learning often needs great computational costs and the convergence is slow. To overcome the drawbacks, large temporal data set is divided into small blocks, then an accelerating EM algorithm (DA-EM ) based on partial E-step is proposed. By iterating in the blocks cyclically, DA-EM updates likelihood function and Bayesian parameters incrementally, so the computational time of E-step is shorten. DA-EM can guarantee the precision of EM, make it converge faster and compute more efficiently. The experimental results justify the computational efficiency of the method.更多还原