基于点衍射干涉仪的分布式随机并行梯度下降算法

Distributed stochastic parallel gradient descent based on point diffraction interferometer

分析了一种光强解耦合的分布式随机并行梯度下降算法,此算法借助近场的波前传感器得到性能指标来得到算法的更新参数,这种性能指标解耦了随机并行梯度下降算法使用的耦合的全场光强,使得算法性能得到提升。分析了一种马赫泽得形式的自参考点衍射干涉仪作为波前传感器。建立两种仿真模型对算法进行了分析,结果表明分布式的随机并行梯度下降算法比原算法在收敛速度上有了数量级的提升,在127仿真结果单元模型上,收敛速度是原算法的10倍以上且收敛结果几乎相同。仿真研究针对不同光束之间的平移相差和同光束的高阶像差,显示了算法应用在光束相干合成的前景。

Distributed stochastic parallel gradient descent algorithm（DSPGD） was analyzed which utilized decoupled light intensity which improved the performance of the algorithm.The algorithm acquired the performance metric by means of near field wave-front sensor.The self-referencing point diffraction interferometer（PDI） like Mach-Zehnder interferometer was analyzed as the wave-front sensor.Two simulation models were constructed and the simulation results show the DSPGD algorithm has order of magnitudes upgrading on convergence rate.Especially in simulation of 127 corrected units,the convergence rate is ten times faster than SPGD algorithm and the corrected results are almost identical.Piston errors of the different light beams and higher orders phase aberration were corrected in the simulation.The results and convergence speed show the higher convergence rates compared with SPGD algorithm and reveal prospects of SPGD algorithm applied on coherent beams combination.