( A, B )---1*30*2---( 1, 0 )( 0, 1 )
让网络的输入只有1个节点,AB各由11张二值化的图片组成,让A中有3个0,8个1.B中全是0,排列组合A的所有可能,统计迭代次数的顺序。在前面实验中得到了8-3分布的数据
A-B | 迭代次数 | 标准差 | |||||
11 | "00011111111" | 3+f1 | 20470.21 | 0 | 0 | 8 | 3.771 |
11 | "01001111111" | 3+f2 | 21099.48 | 1 | 0 | 7 | 3.091 |
11 | "00101111111" | 3+f3 | 21759.81 | 0 | 1 | 7 | 3.091 |
11 | "01100111111" | 3+f4 | 21810.68 | 2 | 0 | 6 | 2.494 |
11 | "01110011111" | 3+f5 | 22313.81 | 3 | 0 | 5 | 2.055 |
11 | "01010111111" | 3+f6 | 22402.91 | 1 | 1 | 6 | 2.357 |
11 | "01111001111" | 3+f7 | 22920.43 | 4 | 0 | 4 | 1.886 |
11 | "00110111111" | 3+f8 | 22920.3 | 0 | 2 | 6 | 2.494 |
11 | "01101011111" | 3+f9 | 23066.35 | 2 | 1 | 5 | 1.7 |
11 | "00111011111" | 3+f10 | 23183.93 | 0 | 3 | 5 | 2.055 |
11 | "01110101111" | 3+f11 | 23570.92 | 3 | 1 | 4 | 1.247 |
11 | "01011011111" | 3+f12 | 23729.11 | 1 | 2 | 5 | 1.7 |
11 | "01011101111" | 3+f13 | 24086.94 | 1 | 3 | 4 | 1.247 |
11 | "01101101111" | 3+f14 | 24342.91 | 2 | 2 | 4 | 0.943 |
11 | "01110111011" | 3+f15 | 24833.37 | 3 | 3 | 2 | 0.471 |
也得到了9-2分布的数据
A-B | 迭代次数 | 标准差 | |||||
11 | 0*0*1*1*1*1*1*1*1*1*1 | 2+g1 | 18792.34 | 0 | 9 | 4.5 | |
11 | 0*1*0*1*1*1*1*1*1*1*1 | 2+g2 | 19501.1 | 1 | 8 | 3.5 | |
11 | 0*1*1*0*1*1*1*1*1*1*1 | 2+g3 | 19969.63 | 2 | 7 | 2.5 | |
11 | 0*1*1*1*0*1*1*1*1*1*1 | 2+g4 | 20556.24 | 3 | 6 | 1.5 | |
11 | 0*1*1*1*1*0*1*1*1*1*1 | 2+g5 | 21103.48 | 4 | 5 | 0.5 |
收敛误差7e-4,每组收敛199次。统计平均值
这次用列排斥力的方法把这两组数据放在一起去比较,观察迭代次数的大小变化顺序。
如计算3+f2的"01001111111"
0 | 1 | 0 | 0 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | |
1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 | |
1 | |||||||||||
2 | 0.11 | 0.06 | 0.04 | 0.03 | 0.02 | 0.02 | 0.01 | ||||
3 | |||||||||||
4 | |||||||||||
5 | 1 | 0.25 | 0.11 | 0.06 | 0.04 | 0.03 | |||||
6 | 1 | 0.25 | 0.11 | 0.06 | 0.04 | ||||||
7 | 1 | 0.25 | 0.11 | 0.06 | |||||||
8 | 1 | 0.25 | 0.11 | ||||||||
9 | 1 | 0.25 | |||||||||
10 | 1 | ||||||||||
11 | |||||||||||
F=8.2794
其余各组数据为
A-B | 迭代次数 | F | 标准差 | |||||
11 | 0*0*1*1*1*1*1*1*1*1*1 | 2+g1 | 18792.34 | 11.02894 | 0 | 9 | 4.5 | |
11 | 0*1*0*1*1*1*1*1*1*1*1 | 2+g2 | 19501.1 | 10.04129 | 1 | 8 | 3.5 | |
11 | 0*1*1*0*1*1*1*1*1*1*1 | 2+g3 | 19969.63 | 9.806912 | 2 | 7 | 2.5 | |
11 | "00011111111" | 3+f1 | 20470.21 | 9.501519 | 0 | 0 | 8 | 3.77124 |
11 | 0*1*1*1*0*1*1*1*1*1*1 | 2+g4 | 20556.24 | 9.716209 | 3 | 6 | 1.5 | |
11 | "01001111111" | 3+f2 | 21099.48 | 8.27949 | 1 | 0 | 7 | 3.09121 |
11 | 0*1*1*1*1*0*1*1*1*1*1 | 2+g5 | 21103.48 | 9.681487 | 4 | 5 | 0.5 | |
11 | "00101111111" | 3+f3 | 21759.81 | 8.517144 | 0 | 1 | 7 | 3.09121 |
11 | "01100111111" | 3+f4 | 21810.68 | 7.954412 | 2 | 0 | 6 | 2.49444 |
11 | "01110011111" | 3+f5 | 22313.81 | 7.828987 | 3 | 0 | 5 | 2.0548 |
11 | "01010111111" | 3+f6 | 22402.91 | 7.438787 | 1 | 1 | 6 | 2.35702 |
11 | "00110111111" | 3+f8 | 22920.3 | 8.287552 | 0 | 2 | 6 | 2.49444 |
11 | "01111001111" | 3+f7 | 22920.43 | 7.794265 | 4 | 0 | 4 | 1.88562 |
11 | "01101011111" | 3+f9 | 23066.35 | 7.16969 | 2 | 1 | 5 | 1.69967 |
11 | "00111011111" | 3+f10 | 23183.93 | 8.204219 | 0 | 3 | 5 | 2.0548 |
11 | "01110101111" | 3+f11 | 23570.92 | 7.078987 | 3 | 1 | 4 | 1.24722 |
11 | "01011011111" | 3+f12 | 23729.11 | 7.265176 | 1 | 2 | 5 | 1.69967 |
11 | "01011101111" | 3+f13 | 24086.94 | 7.216565 | 1 | 3 | 4 | 1.24722 |
11 | "01101101111" | 3+f14 | 24342.91 | 7.030801 | 2 | 2 | 4 | 0.94281 |
11 | "01110111011" | 3+f15 | 24833.37 | 7.016912 | 3 | 3 | 2 | 0.4714 |
比较标准差曲线
数值波动很大,再比较列排斥力曲线
数据要平滑很多。力与迭代次数的反比关系清晰。