基于分段随机扰动幅值的随机并行梯度下降算法研究  

Research of Stochastic Parallel Gradient Descent Algorithm Based on Segmentation Random Disturbance Amplitude

作　　者：吴健[1] 杨慧珍[2] 龚成龙[2] 

机构地区：[1]中国矿业大学信息与电气工程学院,江苏徐州221116 [2]淮海工学院电子工程学院,江苏连云港222005

出　　处：《中国激光》2014年第7期222-227,共6页Chinese Journal of Lasers

基　　金：中国科学院自适应光学重点实验室基金(LAOF201302)

摘　　要：为了提高随机并行梯度下降(SPGD)算法的收敛速度,提出了分段随机扰动幅值的改进方法。从理论上分析了固定增益系数时,随机扰动幅值对SPGD算法收敛速度的影响;提出了分段随机扰动幅值的改进方法;基于61单元变形镜,建立无波前探测自适应光学系统模型,对前65阶Zernike多项式模拟的满足Kolmogorov谱的大气湍流畸变波前进行校正。结果表明,采用分段随机扰动幅值的SPGD算法比固定最佳随机扰动幅值时传统SPGD算法的收敛速度提高了近1.6倍,证明了该改进算法的可行性。The improved method of random perturbance amplitude section is proposed to increase the convergence speed of stochastic parallel gradient descent（SPGD）algorithm.The SPGD algorithm convergence rate,which can be effected by the random disturbance amplitude,is analyzed when the gain coefficient is fixed.The segmentation random perturbance amplitude method is put forward.The adaptive optics system without wavefront sensor is built with a 61-element deformation mirror to correct the wavefront aberrations,which is simulated by the 65-order Zernike polynomials and the aberrations meet the Kolmogorov spectrum.Compared with the best fixed initial perturbance amplitude SPGD algorithm,the convergence speed increases 1.6 times by adopting the SPGD algorithm based on the segmentation random perturbance amplitude.The improved algorithm is verified to be feasible.

关 键 词：自适应光学 波前校正 随机并行梯度下降 分段随机扰动幅值 

分 类 号：TP273.2[自动化与计算机技术—检测技术与自动化装置]