平面内的3个点只可能有6种结构
1 | - | - | - | - | 4 | - | - | - | - | |
- | - | - | - | - | - | - | - | |||
- | - | - | - | - | - | - | - | |||
- | - | - | 1 | - | - | - | 1 | |||
- | - | 1 | 1 | - | 1 | 1 | - | |||
2 | - | - | - | - | 5 | - | - | - | - | |
- | - | - | - | - | - | - | - | |||
- | - | - | 1 | - | - | - | 1 | |||
- | - | - | 1 | - | - | 1 | - | |||
- | - | - | 1 | - | 1 | - | - | |||
3 | - | - | - | - | 6 | - | - | - | - | |
- | - | - | - | - | - | - | - | |||
- | - | - | 1 | - | - | - | - | |||
- | - | - | 1 | - | - | - | - | |||
- | - | 1 | - | - | 1 | 1 | 1 |
向一个平面内扔3个石子,只有这6种可能,
现在向这个三角形内随机的扔3个石子,统计6个结构的占比,从9个点中选3个点有84种可能,分别占比为
84 | ||
1 | 24 | 0.285714 |
2 | 11 | 0.130952 |
3 | 34 | 0.404762 |
4 | 8 | 0.095238 |
5 | 6 | 0.071429 |
6 | 1 | 0.011905 |
顺序是3>1>2>4>5>6
用网络分类这6个结构
3 | - | - | - | - | 50827.11 |
- | - | - | - | 50827.11 | |
- | - | - | 1 | 50827.11 | |
- | - | - | 1 | 50827.11 | |
- | - | 1 | - | 50827.11 | |
1 | - | - | - | - | 50914.58 |
- | - | - | - | 50914.58 | |
- | - | - | - | 50914.58 | |
- | - | - | 1 | 50914.58 | |
- | - | 1 | 1 | 50914.58 | |
2 | - | - | - | - | 51308.42 |
- | - | - | - | 51308.42 | |
- | - | - | 1 | 51308.42 | |
- | - | - | 1 | 51308.42 | |
- | - | - | 1 | 51308.42 | |
4 | - | - | - | - | 69065.23 |
- | - | - | - | 69065.23 | |
- | - | - | - | 69065.23 | |
- | - | - | 1 | 69065.23 | |
- | 1 | 1 | - | 69065.23 | |
5 | - | - | - | - | 69478.73 |
- | - | - | - | 69478.73 | |
- | - | - | 1 | 69478.73 | |
- | - | 1 | - | 69478.73 | |
- | 1 | - | - | 69478.73 | |
6 | - | - | - | - | 131879.6 |
- | - | - | - | 131879.6 | |
- | - | - | - | 131879.6 | |
- | - | - | - | 131879.6 | |
- | 1 | 1 | 1 | 131879.6 |
( A, B )---4*11*2---( 1, 0 )( 0, 1 )
当收敛误差为7e-4,B全为0,隐藏层节点数为11个的时候,迭代次数的顺序就是3<1<2<4<5<6
改变三角形的形状
这个三角形有14个点,取3个点有364种可能,
364 | ||
1 | 33 | 0.090659 |
2 | 165 | 0.453297 |
3 | 144 | 0.395604 |
4 | 12 | 0.032967 |
5 | 9 | 0.024725 |
6 | 1 | 0.002747 |
6个结构的占比为2>3>1>4>5>6
( A, B )---4*105*2---( 1, 0 )( 0, 1 )
其他参数不变,把隐藏层节点数调整为105个点时,
2 | - | - | - | - | 15442.24 |
- | - | - | - | 15442.24 | |
- | - | - | 1 | 15442.24 | |
- | - | - | 1 | 15442.24 | |
- | - | - | 1 | 15442.24 | |
3 | - | - | - | - | 19619.02 |
- | - | - | - | 19619.02 | |
- | - | - | 1 | 19619.02 | |
- | - | - | 1 | 19619.02 | |
- | - | 1 | - | 19619.02 | |
1 | - | - | - | - | 19751.36 |
- | - | - | - | 19751.36 | |
- | - | - | - | 19751.36 | |
- | - | - | 1 | 19751.36 | |
- | - | 1 | 1 | 19751.36 | |
4 | - | - | - | - | 32694.38 |
- | - | - | - | 32694.38 | |
- | - | - | - | 32694.38 | |
- | - | - | 1 | 32694.38 | |
- | 1 | 1 | - | 32694.38 | |
5 | - | - | - | - | 40501.47 |
- | - | - | - | 40501.47 | |
- | - | - | 1 | 40501.47 | |
- | - | 1 | - | 40501.47 | |
- | 1 | - | - | 40501.47 | |
6 | - | - | - | - | 52162.32 |
- | - | - | - | 52162.32 | |
- | - | - | - | 52162.32 | |
- | - | - | - | 52162.32 | |
- | 1 | 1 | 1 | 52162.32 |
迭代次数的顺序就是2<3<1<4<5<6.因为搜索区域内2的结构更多因此更容易找到,所以迭代次数更小。而结构6的数量占比更少,所以迭代次数更大。
综合比较这两组数据,暗示了一种可能。增加隐藏层节点数就是在拉伸这个三角形,因为搜索形状的变化改变了6个结构的占比,并导致搜索难度的改变,最终决定了迭代次数的顺序。
计算机验算两组数据,都随机扔了1000次。
1000 | 1000 | |||||
1 | 303 | 0.3 | 1 | 93 | 0.09 | |
2 | 126 | 0.13 | 2 | 475 | 0.48 | |
3 | 383 | 0.38 | 3 | 374 | 0.37 | |
4 | 110 | 0.11 | 4 | 40 | 0.04 | |
5 | 70 | 0.07 | 5 | 14 | 0.01 | |
6 | 8 | 0.01 | 6 | 4 | 0 | |
3>1>2>4>5>6 | 2>3>1>4>5>6 |