基于混合空洞注意力机制的图卷积强化学习

宋莉,宛袁玉,宋明黎

Graph convolutional reinforcement learning via hybrid dilated attention mechanism

Li SONG,Yuanyu WAN,Mingli SONG

表 1 所提算法与基准算法收敛时的实验结果

Tab.1 Experimental results when proposed algorithm and benchmark algorithms converge

算法 损失值 KL-loss atten q 成功率 传统GCRL 6.15×1016 3.99×1016 –3.93×1015 8.80 0.68 GCRL-MHA 7.54×1013 5.02×1016 –4.87×1015 9.21 0.71 GCRL-MSDA 6.82×1010 4.58×1016 –4.89×1015 11.59 0.89 GCRL-EDA 6.88×107 4.54×1016 –4.93×1015 12.02 0.92