制作两个神经网络用来分类2**0.5、3**0.5,2**0.5和e。每个无理数取3万位有效数字,每10个数字变成一张图片。用前2500张图片来训练网络,用2500-3000张图片来做测试。
比如2**0.5的第一张图片
收敛标准δ取0.5到1e-4共16个值,每个收敛标准收敛199次,统计平均分辨准确率,迭代次数。
得到的表格
训练集 | 0-2500 | ||||||||
测试集 | 2500-3000 | ||||||||
f2[0] | f2[1] | 迭代次数n | 平均准确率p-ave | δ | 耗时ms/次 | 耗时ms/199次 | 耗时 min/199 | 最大值p-max | pave标准差 |
0.501147535 | 0.499689475 | 17.51758794 | 0.499537688 | 0.5 | 4.884422111 | 972 | 0.0162 | 0.522 | 0.004812811 |
0.422210311 | 0.577346693 | 11189.55779 | 0.483246231 | 0.4 | 104.9095477 | 20877 | 0.34795 | 0.524 | 0.013828623 |
0.296495826 | 0.703493738 | 34580.47236 | 0.498226131 | 0.3 | 316.3115578 | 62946 | 1.0491 | 0.506 | 0.004294922 |
0.266027637 | 0.733965365 | 116589.8241 | 0.486874372 | 0.2 | 1056.276382 | 210215 | 3.503583333 | 0.525 | 0.01838175 |
0.441127684 | 0.558872658 | 178663.6633 | 0.493236181 | 0.1 | 1204.477387 | 239691 | 3.99485 | 0.537 | 0.019346601 |
0.527119391 | 0.472880614 | 299744.0854 | 0.491236181 | 0.01 | 2715.798995 | 540460 | 9.007666667 | 0.53 | 0.013548702 |
0.57774068 | 0.422259283 | 427566.1055 | 0.49361809 | 0.001 | 3880.844221 | 772291 | 12.87151667 | 0.541 | 0.015254454 |
0.507523048 | 0.492476938 | 423875.9045 | 0.492979899 | 9.00E-04 | 3867.422111 | 769633 | 12.82721667 | 0.534 | 0.014838142 |
0.567738564 | 0.432261394 | 452748.3869 | 0.491341709 | 8.00E-04 | 4193.211055 | 834466 | 13.90776667 | 0.531 | 0.015622231 |
0.482439484 | 0.517560534 | 452554.5729 | 0.492432161 | 7.00E-04 | 3811.422111 | 758474 | 12.64123333 | 0.537 | 0.015850881 |
0.537647222 | 0.462352754 | 459869.4724 | 0.492135678 | 6.00E-04 | 4248.542714 | 845465 | 14.09108333 | 0.54 | 0.015820815 |
0.517569942 | 0.482430045 | 466761.0905 | 0.494291457 | 5.00E-04 | 4226.030151 | 840989 | 14.01648333 | 0.539 | 0.01634968 |
0.522594814 | 0.477405195 | 489836.603 | 0.491351759 | 4.00E-04 | 4419.703518 | 879525 | 14.65875 | 0.527 | 0.015002743 |
0.542689058 | 0.457310917 | 528504.8643 | 0.491201005 | 3.00E-04 | 4827.341709 | 960648 | 16.0108 | 0.536 | 0.013936316 |
0.572835882 | 0.427164091 | 569002.3518 | 0.492668342 | 2.00E-04 | 5219.713568 | 1038733 | 17.31221667 | 0.533 | 0.013133277 |
0.552753899 | 0.447246105 | 660039.6482 | 0.493246231 | 1.00E-04 | 6018.592965 | 1197716 | 19.96193333 | 0.53 | 0.013978615 |
2**0.5 | |
2500-3000 | |
3**0.5 | e |
平均准确率p-ave | 平均准确率p-ave |
0.500452 | 0.499538 |
0.517995 | 0.483246 |
0.519864 | 0.498226 |
0.513638 | 0.486874 |
0.512447 | 0.493236 |
0.504769 | 0.491236 |
0.506643 | 0.493618 |
0.504467 | 0.49298 |
0.50492 | 0.491342 |
0.504734 | 0.492432 |
0.506146 | 0.492136 |
0.506638 | 0.494291 |
0.50705 | 0.491352 |
0.504935 | 0.491201 |
0.50692 | 0.492668 |
0.504578 | 0.493246 |
可以看到3**0.5的分类准确率显著的大于50%,而e的pave数据显著的小于50%。差异非常明显。这表明无理数的数字分布是有规律的,给出一个10位的数字序列可能存在一种方法判断这个序列属于哪个无理数。
无理数的数据来源
https://www.wolframalpha.com/input/?i=x%5E2-1
N[sqr(2),30000]