【摘要】 为解决EM算法对初始值比较敏感,易陷入局部最优和当模型缺失变量或隐变量的比例较高时收敛速度慢等问题,本文首先采用经典的K-means聚类算法进行初始值选取,使得初始值更加接近EM序列的稳定点,在一定程度上能避免EM序列陷入局部极值,然后采用ε-加速算法进行加速,即K-means+ε-加速EM算法,并比较不同情况加速的EM算法在有限混合模型参数估计中的时间成本和精度大小。数值结果表明:在不降低精度的情况下K-means+ε-加速EM算法在各种条件下均能够大大加快EM算法的收敛速度。更多还原

【Abstract】 In order to solve the problem that EM algorithm is sensitive to initial value, easy to fall into local optimum and slow convergence speed when the ratio of missing variables or hidden variables is high, this paper firstly uses the classical K-means clustering algorithm to select the initial value. The initial value is closer to the stable point of the EM sequence, and the EM sequence can be prevented from falling into the local extremum to a certain extent. Then the acceleration algorithm is adopted to accelerate EM sequence, that is, the K-means+ accelerated EM algorithm, and the time cost and precision of the accelerated EM algorithm has been compared in the parameter estimation of the finite mixed model. Numerical results show that K-means + accelerated EM algorithm can greatly speed up the convergence of EM algorithm under various conditions without reducing the accuracy.更多还原