在行列可自由变换的条件下,平面上的3点结构只有6个
4点结构只有16个
3点结构按照训练集A-B矩阵的高有3s1和3s2两种顺序,同样4点结构也有4s1和4s2两种顺序。4s1可以用结构加法转化成3点结构顺序4s1-3.
(A,B)---6*n*2---(0,1)(1,0)
分类A和B,让B全是0。当收敛误差为7e-4,收敛199次取迭代次数平均值。让隐藏层节点数n分别为2,5,10,15,20,25,30,40,50,60,70,80,90,100.分别统计和计算3s1,3s2和4s1-3的顺序,并比较他们的关系,和4s1-3随n的变化规律。
得到4s1的迭代次数表格
2 | 5 | 10 | 15 | 20 | 25 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 | |
4s1 | 4s1 | 4s1 | 4s1 | 4s1 | 4s1 | 4s1 | 4s1 | 4s1 | 4s1 | 4s1 | 4s1 | 4s1 | 4s1 | |
1 | 130067.1 | 37220.82 | 13046.51 | 7873.322 | 5819.814 | 4789.221 | 4202.005 | 3535.714 | 3102.95 | 2842.568 | 2655.03 | 2499.935 | 2369.201 | 2263.151 |
2 | 137345.7 | 49718.1 | 23714.48 | 16839.98 | 13494.02 | 11621.09 | 10464.07 | 8921.352 | 8019.161 | 7368.869 | 6863.704 | 6475.08 | 6134.739 | 5822.136 |
3 | 147623.9 | 58216.74 | 28695.01 | 20334.97 | 16379.75 | 14030.89 | 12464.74 | 10686.93 | 9451.07 | 8625.945 | 8036.678 | 7526.206 | 7064.899 | 6698.91 |
4 | 213364.7 | 77940.34 | 33998.63 | 22686.07 | 17562.29 | 14828.98 | 12975.07 | 10808.76 | 9510.065 | 8593.764 | 7907.613 | 7380.94 | 6963.372 | 6589.181 |
5 | 223146.8 | 85425.66 | 46045.48 | 34847.41 | 29456.1 | 26736.76 | 24480.26 | 22250.95 | 20657.38 | 19691.63 | 18948.08 | 18383.89 | 17934.41 | 17562.79 |
6 | 227472.6 | 93281.46 | 51568.79 | 39175.95 | 33140.04 | 29703.8 | 27187.2 | 24212.2 | 22426.8 | 21138.11 | 20310.16 | 19583.37 | 18985.09 | 18563.1 |
7 | 220691.5 | 86532.45 | 47650.36 | 38132.75 | 32924.79 | 29952.65 | 27969.19 | 25676.67 | 24028.51 | 23033.83 | 22208.83 | 21674.96 | 21204.53 | 20880.33 |
8 | 232172.3 | 98035.13 | 57638.09 | 45530.14 | 40072.25 | 36286.81 | 33727.32 | 30664.38 | 28761.96 | 27434.41 | 26535.44 | 25861.95 | 25197.51 | 24694.35 |
9 | 146737.9 | 64957.43 | 38848.33 | 31045.95 | 27086.24 | 24994.58 | 23485.43 | 21604.11 | 20577.65 | 19723.06 | 19174.6 | 18725.03 | 18368.45 | 18105.69 |
10 | 219349.4 | 88791.16 | 48261.11 | 36778.5 | 31096.33 | 28159.24 | 26175.72 | 23645.35 | 22136.61 | 21192.29 | 20355.74 | 19958.72 | 19556.67 | 19195.54 |
11 | 157378 | 75064.35 | 47351.46 | 37856.79 | 32727.38 | 30238.38 | 27994.4 | 25184.79 | 23130.64 | 21552.99 | 20284.81 | 19137.51 | 18056.55 | 17047.78 |
12 | 233260.6 | 103800.9 | 58271.86 | 44946.56 | 37980.69 | 33919.7 | 31024.71 | 27406.5 | 24908.13 | 22967.06 | 21449.99 | 20174.62 | 19025.15 | 18017.59 |
13 | 234041.2 | 105014.7 | 66094.02 | 54752.43 | 48769.57 | 45373.27 | 42767.27 | 39669.3 | 37677.35 | 36231.88 | 35229.96 | 34493.62 | 33782.32 | 33263.12 |
14 | 232146.7 | 105151.7 | 61783.93 | 48237.33 | 41565.32 | 38028.71 | 35176.25 | 31848.89 | 29527.8 | 28025.49 | 26949.87 | 26048.09 | 25325.43 | 24630.64 |
15 | 232951.3 | 109898.7 | 67340.41 | 54332.63 | 47362.96 | 43677.86 | 41257.21 | 37758.94 | 35650.82 | 34253.57 | 33170.94 | 32613.25 | 31953.1 | 31449.1 |
16 | 462026.8 | 201191.6 | 108180.1 | 81425.93 | 68329.01 | 61127.04 | 55639.52 | 49425.21 | 45494.8 | 43077.39 | 41091.7 | 39565.35 | 38554.58 | 37576.77 |
这14组数据的顺序为
2 | 5 | 10 | 15 | 20 | 25 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 |
1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 |
9 | 3 | 3 | 3 | 3 | 3 | 3 | 3 | 3 | 4 | 4 | 4 | 4 | 4 |
3 | 9 | 4 | 4 | 4 | 4 | 4 | 4 | 4 | 3 | 3 | 3 | 3 | 3 |
11 | 11 | 9 | 9 | 9 | 9 | 9 | 9 | 9 | 5 | 5 | 5 | 5 | 11 |
4 | 4 | 5 | 5 | 5 | 5 | 5 | 5 | 5 | 9 | 9 | 9 | 11 | 5 |
10 | 5 | 11 | 10 | 10 | 10 | 10 | 10 | 10 | 6 | 11 | 11 | 9 | 12 |
7 | 7 | 7 | 11 | 11 | 6 | 6 | 6 | 6 | 10 | 6 | 6 | 6 | 9 |
5 | 10 | 10 | 7 | 7 | 7 | 7 | 11 | 11 | 11 | 10 | 10 | 12 | 6 |
6 | 6 | 6 | 6 | 6 | 11 | 11 | 7 | 7 | 12 | 12 | 12 | 10 | 10 |
14 | 8 | 8 | 12 | 12 | 12 | 12 | 12 | 12 | 7 | 7 | 7 | 7 | 7 |
8 | 12 | 12 | 8 | 8 | 8 | 8 | 8 | 8 | 8 | 8 | 8 | 8 | 14 |
15 | 13 | 14 | 14 | 14 | 14 | 14 | 14 | 14 | 14 | 14 | 14 | 14 | 8 |
12 | 14 | 13 | 15 | 15 | 15 | 15 | 15 | 15 | 15 | 15 | 15 | 15 | 15 |
13 | 15 | 15 | 13 | 13 | 13 | 13 | 13 | 13 | 13 | 13 | 13 | 13 | 13 |
16 | 16 | 16 | 16 | 16 | 16 | 16 | 16 | 16 | 16 | 16 | 16 | 16 | 16 |
因为隐藏层节点数是按照数量依次增加由2变化到100,因此也有理由猜测这14组顺序也有内在的变化规律。
与之对应3s1,3s2,4s1-3的数值分别为
2 | 5 | 10 | 15 | ||||||||||||
3s1 | 3s2 | 4s1-3 | 3s1 | 3s2 | 4s1-3 | 3s1 | 3s2 | 4s1-3 | 3s1 | 3s2 | 4s1-3 | ||||
1 | 1 | 1 | 1.49 | 1 | 1 | 2.11 | 1 | 1 | 3.14 | 1 | 1 | 3.86 | |||
2 | 0.96 | 1.84 | 1.64 | 1.46 | 0.85 | 2.31 | 1.68 | 1.01 | 3.59 | 1.8 | 1.1 | 4.49 | |||
3 | 0.95 | 0.69 | 1.49 | 1.37 | 0.82 | 2.07 | 1.57 | 0.98 | 3.26 | 1.72 | 1.09 | 4.16 | |||
4 | 0.71 | 1.02 | 1.54 | 1.14 | 1.16 | 2.37 | 1.37 | 1.34 | 3.92 | 1.51 | 1.47 | 5.09 | |||
5 | 0.98 | 0.73 | 1.54 | 1.59 | 1 | 2.31 | 2.04 | 1.32 | 3.93 | 2.31 | 1.55 | 5.28 | |||
6 | 1.46 | 2.12 | 2.92 | 2.37 | 2.38 | 4.35 | 2.74 | 2.67 | 6.58 | 2.96 | 2.82 | 8.15 | |||
20 | 25 | 30 | 40 | ||||||||||||
3s1 | 3s2 | 4s1-3 | 3s1 | 3s2 | 4s1-3 | 3s1 | 3s2 | 4s1-3 | 3s1 | 3s2 | 4s1-3 | ||||
1 | 1 | 1 | 4.34 | 1 | 1 | 4.67 | 1 | 1 | 4.84 | 1 | 1 | 5.05 | |||
2 | 1.92 | 1.17 | 5.11 | 1.97 | 1.21 | 5.55 | 2.03 | 1.23 | 5.78 | 2.14 | 1.27 | 6.12 | |||
3 | 1.85 | 1.17 | 4.83 | 1.94 | 1.24 | 5.3 | 2.02 | 1.27 | 5.58 | 2.14 | 1.34 | 5.99 | |||
4 | 1.65 | 1.57 | 5.92 | 1.73 | 1.65 | 6.57 | 1.83 | 1.7 | 6.99 | 1.97 | 1.8 | 7.52 | |||
5 | 2.55 | 1.72 | 6.27 | 2.74 | 1.85 | 7.03 | 2.89 | 1.94 | 7.52 | 3.15 | 2.13 | 8.24 | |||
6 | 3.12 | 2.93 | 9.22 | 3.25 | 3.02 | 10 | 3.37 | 3.07 | 10.4 | 3.54 | 3.18 | 10.9 | |||
50 | 60 | 70 | 80 | ||||||||||||
3s1 | 3s2 | 4s1-3 | 3s1 | 3s2 | 4s1-3 | 3s1 | 3s2 | 4s1-3 | 3s1 | 3s2 | 4s1-3 | ||||
1 | 1 | 1 | 5.23 | 1 | 1 | 5.29 | 1 | 1 | 5.32 | 1 | 1 | 5.36 | |||
2 | 2.22 | 1.27 | 6.44 | 2.29 | 1.25 | 6.62 | 2.35 | 1.23 | 6.79 | 2.4 | 1.2 | 6.95 | |||
3 | 2.24 | 1.36 | 6.34 | 2.35 | 1.37 | 6.56 | 2.42 | 1.38 | 6.73 | 2.49 | 1.37 | 6.91 | |||
4 | 2.08 | 1.86 | 8.01 | 2.17 | 1.91 | 8.33 | 2.25 | 1.94 | 8.57 | 2.33 | 1.96 | 8.85 | |||
5 | 3.32 | 2.22 | 8.88 | 3.49 | 2.31 | 9.3 | 3.61 | 2.37 | 9.65 | 3.75 | 2.41 | 10 | |||
6 | 3.67 | 3.23 | 11.5 | 3.81 | 3.25 | 11.8 | 3.93 | 3.29 | 12.1 | 4.02 | 3.29 | 12.4 | |||
90 | 100 | ||||||||||||||
3s1 | 3s2 | 4s1-3 | 3s1 | 3s2 | 4s1-3 | ||||||||||
1 | 1 | 1 | 5.38 | 1 | 1 | 5.38 | |||||||||
2 | 2.47 | 1.16 | 7.1 | 2.5 | 1.11 | 7.25 | |||||||||
3 | 2.56 | 1.35 | 7.06 | 2.62 | 1.32 | 7.2 | |||||||||
4 | 2.42 | 1.97 | 9.09 | 2.49 | 1.97 | 9.29 | |||||||||
5 | 3.88 | 2.43 | 10.3 | 3.97 | 2.44 | 10.6 | |||||||||
6 | 4.15 | 3.29 | 12.7 | 4.23 | 3.28 | 13 |
因为3点结构中2,3,5有3行,因此这些数据分成1,4,6和2,3,5两部分比较
如当n=40的时候
40 | |||
3s1 | 3s2 | 4s1-3 | |
2 | 2.136763 | 1.269147 | 6.117891 |
3 | 2.144655 | 1.336655 | 5.992435 |
5 | 3.149348 | 2.126441 | 8.238479 |
二次归一化 | |||
1.00003 | 1.000037 | 1.001292 | |
1.003723 | 1.053231 | 0.980759 | |
1.473931 | 1.675551 | 1.34836 |
3s1和3s2通过二次归一化的方法几乎重合,这一规律对所有这14组数据都成立,隐藏层节点数n对ns1和ns2之间的关系没有影响。
4s1-3-2,3,5的数据
2 | 5 | 10 | 15 | 20 | 25 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 | |
2 | 1.64 | 2.31 | 3.59 | 4.49 | 5.11 | 5.55 | 5.78 | 6.12 | 6.44 | 6.62 | 6.79 | 6.95 | 7.1 | 7.25 |
3 | 1.49 | 2.07 | 3.26 | 4.16 | 4.83 | 5.3 | 5.58 | 5.99 | 6.34 | 6.56 | 6.73 | 6.91 | 7.06 | 7.2 |
5 | 1.54 | 2.31 | 3.93 | 5.28 | 6.27 | 7.03 | 7.52 | 8.24 | 8.88 | 9.3 | 9.65 | 10 | 10.3 | 10.6 |
4s1-3-2,3,5的数据随着n的增加变大。
统计1,4,6的数据
40 | |||
3s1 | 3s2 | 4s1-3 | |
1 | 1 | 1 | 5.05 |
4 | 1.97 | 1.8 | 7.52 |
6 | 3.54 | 3.18 | 10.9 |
3s1和3s2的数据几乎是重合的,但4s1-3的数据很接近一条直线
2 | 5 | 10 | 15 | 20 | 25 | 30 | 40 | 50 | 60 | 70 | 80 | 90 | 100 | |
1 | 1.49 | 2.11 | 3.14 | 3.86 | 4.34 | 4.67 | 4.84 | 5.05 | 5.23 | 5.29 | 5.32 | 5.36 | 5.38 | 5.38 |
4 | 1.54 | 2.37 | 3.92 | 5.09 | 5.92 | 6.57 | 6.99 | 7.52 | 8.01 | 8.33 | 8.57 | 8.85 | 9.09 | 9.29 |
6 | 2.92 | 4.35 | 6.58 | 8.15 | 9.22 | 10 | 10.4 | 10.9 | 11.5 | 11.8 | 12.1 | 12.4 | 12.7 | 13 |
这个表中当n>=40以后曲线非常接近直线,这就意味这改变n就是改变了4s1-3-1,4,6与x轴的夹角,n越大夹角越大,呈明显的线性关系。