相干合成中的随机并行梯度下降算法性能研究  

Performance of Stochastic Parallel Gradient Descent Algorithm in Coherent Combination

作　　者：李兴珂 何云涛[1] 

机构地区：[1]北京航空航天大学电子信息工程学院,北京100191

出　　处：《光学学报》2016年第12期1-8,共8页Acta Optica Sinica

基　　金：上海航天创新基金(SAST2015090);国家自然科学基金(11673079)

摘　　要：随机并行梯度下降算法(SPGD)是一种基于直接性能指标优化的相位控制方法,在自适应光学中有较好的适用性。该算法主要包含增益系数和随机扰动幅度两个可变参数,其取值对算法收敛性有很大的影响。对双边SPGD算法实现收敛时参数的取值要求进行研究,结合算法原理分析了算法参数的取值范围,并通过大量仿真实验找出所有使双边SPGD算法收敛的增益系数和随机扰动幅度值;得到随机扰动幅度的取值下限,理论和仿真分析了下限存在的原因及取值;在相干合成中存在相位噪声,研究了不同相位校正器参数的情况下可使算法收敛的参数的取值范围。Stochastic parallel gradient descent algorithm（SPGD）,aphase control method based on directly optimizing performance index,has good applicability in adaptive optics.The algorithm involves two variable parameters：gain coefficient and random perturbation amplitude,whose values have great influence on the convergence of the algorithm.The requirements of the parameter values for the convergence of SPGD algorithm are discussed,and the value range of parameters are analyzed by combining with the principle of the algorithm.Furthermore,a large number of simulations are conducted to analyze all the gain coefficients and the amplitude of random disturbance,which can ensure the convergency of bilateral SPGD.Meanwhile,the lower limit of the random disturbance amplitude is obtained,and the reason for its existence and the lower limit value are also analyzed theoretically and experimentally.With the existence of phase noise in coherent combination,the range of parameters for the algorithm to be convergent is analyzed with different phase corrector parameters.

关 键 词：自适应光学 随机并行梯度下降算法 相位控制 梯度估计 相干合成 

分 类 号：O437[机械工程—光学工程]