随机方差缩减的递归动量策略梯度方法

辛垚辉,李永强,冯宇,胡磊

Stochastic variance-reduced recursive momentum policy gradient method

Yaohui XIN,Yongqiang LI,Yu FENG,Lei HU

表 4 5种实验算法在不同阶段的方差

Tab.4 Variance of five experimental algorithms at different stages

算法 Acrobot Cartpole Hopper Walker 0~30% 30%~100% 0~30% 30%~100% 0~30% 30%~100% 0~30% 30%~100% GPOMDP 6383.36 762.24 3726.25 83.80 73250.86 37910.80 7108.04 3110.71 STORM-PG 5577.74 296.93 2025.44 67.35 80832.25 20085.55 4739.30 2500.16 PAGE-PG 5178.15 247.35 3472.70 104.60 57135.13 25513.84 5098.69 2666.30 SHARP 4634.38 308.80 2542.07 152.41 45235.57 4477.66 4100.44 2607.13 SVRRM-PG 3531.65 163.94 1817.99 53.90 38120.92 2744.55 3205.00 1836.84