f***t
发帖数: 2247 | 1 有个问题求教计算机大仙出手相救，包子酬谢！
需求如下：
[1],俺一直是搞实验的，对modeling一窍不通，所以麻烦您写code的时候加点备注:(
还有就是别笑话俺，俺读书少:(
[2],麻烦您使用神经网络算法（和老板讨论后，他决定的，俺不懂这个玩艺），而且使
用MATLAB（因为俺这个破系只有MATLAB）
[3],俺update一个工艺流程，需要监控流速随时间的变化，共6个监控点，现已测得11
个时间点上的数据。麻烦您给俺一段代码，预测一下第12个时间点的流速数据（检测时
间的间隔是相同的）。
数据如下：
时间1：1.12 1.82 0.11 0.54 4.87 0.64
时间2：1.61 2.04 0.16 0.68 4.93 0.66
时间3：1.32 1.71 0.21 0.63 4.89 0.58
时间4：1.69 2.00 0.19 0.78 4.83 0.60
时间5：1.12 2.04 0.21 0.72 4.76 0.59
时间6：1.55 1.92 0.16 0.62 4.68 0.60
时间7：1.67 1.96 0.17 0.83 4.82 0.52
时间8：1.73 1.99 0.21 0.72 4.75 0.62
时间9：1.77 2.18 0.20 0.70 4.56 0.70
时间10：1.45 1.93 0.16 0.60 4.73 0.65
时间11：1.38 2.05 0.11 0.72 4.99 0.73 |
p*q
发帖数: 59 | 2 well, as an initial model...we assume there is a function that maps a time
value to a 5-dimension (all over space of R...i.e. real number space...)
vector...
f: T -> R X R X R X R X R
the target is to obtain this function f...
could you give us a bit more information, such as whether those 5 dimensions
are all independent from each other (i.e. are they all orthogonal?) or has
a certain dependency among certain dimensions? |
p*q
发帖数: 59 | 3 BTW, if you have a really large data set (I mean, potentially), you don't
have to use MATLAB just because you only have MATLAB installed on your
machine...
There are tons of open sourced tools out there that you could use, with a
range of different learning algorithms...
Tensorflow, for instance, is innately built upon neural networks...an
example
for your case using tensor flow could be found here...
https://www.tensorflow.org/get_started/estimator |
d****o
发帖数: 32610 | 4 就你这几个点什么NN也救不了
11
【在 f***t 的大作中提到】
: 有个问题求教计算机大仙出手相救，包子酬谢！
: 需求如下：
: [1],俺一直是搞实验的，对modeling一窍不通，所以麻烦您写code的时候加点备注:(
: 还有就是别笑话俺，俺读书少:(
: [2],麻烦您使用神经网络算法（和老板讨论后，他决定的，俺不懂这个玩艺），而且使
: 用MATLAB（因为俺这个破系只有MATLAB）
: [3],俺update一个工艺流程，需要监控流速随时间的变化，共6个监控点，现已测得11
: 个时间点上的数据。麻烦您给俺一段代码，预测一下第12个时间点的流速数据（检测时
: 间的间隔是相同的）。
: 数据如下：
|
f***t
发帖数: 2247 | 5 先多谢！
另，在同一个监控时间内，这6个监控点的流速应该是有关联的，虽然这6个测试点都是
在不同的branch的节点位置上，但毕竟还是在同一个流场中。
dimensions
has
【在 p*q 的大作中提到】
: well, as an initial model...we assume there is a function that maps a time
: value to a 5-dimension (all over space of R...i.e. real number space...)
: vector...
: f: T -> R X R X R X R X R
: the target is to obtain this function f...
: could you give us a bit more information, such as whether those 5 dimensions
: are all independent from each other (i.e. are they all orthogonal?) or has
: a certain dependency among certain dimensions?
|
f***t
发帖数: 2247 | 6 我现在还在继续测流速，数据会越来越多。麻烦你们先给我个模型，数据多了之后，我
修改一下参数就可以了，谢谢啦。
【在 p*q 的大作中提到】
: BTW, if you have a really large data set (I mean, potentially), you don't
: have to use MATLAB just because you only have MATLAB installed on your
: machine...
: There are tons of open sourced tools out there that you could use, with a
: range of different learning algorithms...
: Tensorflow, for instance, is innately built upon neural networks...an
: example
: for your case using tensor flow could be found here...
: https://www.tensorflow.org/get_started/estimator
|
f***t
发帖数: 2247 | 7 我现在还在继续测流速，这需要时间，数据会越来越多。麻烦你们先给我个模型，数据
多了之后，我修改一下参数就可以了，谢谢啦。
【在 d****o 的大作中提到】
: 就你这几个点什么NN也救不了
:
: 11
|
p*q
发帖数: 59 | 8 I don't have MATLAB here on my machine (no direct and fast chinese input
either -- sorry about that)...
But I found a youtube tutorial on prediction with MATLAB neural network
package, which seems to be targeting on a similar problem as yours...and
quite intuitive ( roughly 5mins)...
(forgot to post the youtube link...here it is...)
update:
https://www.youtube.com/watch?v=4tP1JA6aH5o
【在 f***t 的大作中提到】
: 先多谢！
: 另，在同一个监控时间内，这6个监控点的流速应该是有关联的，虽然这6个测试点都是
: 在不同的branch的节点位置上，但毕竟还是在同一个流场中。
:
: dimensions
: has
|
d****o
发帖数: 32610 | 9 你这个要用RNN，现在好像流行LSTM
自己搜个教程研究研究吧
【在 f***t 的大作中提到】
: 我现在还在继续测流速，这需要时间，数据会越来越多。麻烦你们先给我个模型，数据
: 多了之后，我修改一下参数就可以了，谢谢啦。
|
f***t
发帖数: 2247 | 10 好的，正在看这个Tensorflow，先表示感谢！
【在 p*q 的大作中提到】
: BTW, if you have a really large data set (I mean, potentially), you don't
: have to use MATLAB just because you only have MATLAB installed on your
: machine...
: There are tons of open sourced tools out there that you could use, with a
: range of different learning algorithms...
: Tensorflow, for instance, is innately built upon neural networks...an
: example
: for your case using tensor flow could be found here...
: https://www.tensorflow.org/get_started/estimator
|
|
|
p*q
发帖数: 59 | 11 and here is a youtube tutorial (roughly 5mins long) on predicting using
MATLAB
neural network package...
no programming background needed...just some quite intuitive keyboard/mouse
operations...
https://www.youtube.com/watch?v=4tP1JA6aH5o
You could try to play with the first 10 or 9 vectors to predict the 10th or
11th vector that you already have...and see how that would work...
Good luck!
【在 f***t 的大作中提到】
: 好的，正在看这个Tensorflow，先表示感谢！
|
f***t
发帖数: 2247 | 12 好的，多谢！我正在看Tensorflow，好像这个Tensorflow在神经网络方面比matlab更牛
X。对吗？如果是这样，俺就先试试这个Tensorflow。能早点出结果，就是省钱了，做
实验的苦逼啊，整天耳朵里每时每刻都会听到一个单词，money,money,money,....
【在 p*q 的大作中提到】
: I don't have MATLAB here on my machine (no direct and fast chinese input
: either -- sorry about that)...
: But I found a youtube tutorial on prediction with MATLAB neural network
: package, which seems to be targeting on a similar problem as yours...and
: quite intuitive ( roughly 5mins)...
: (forgot to post the youtube link...here it is...)
: update:
: https://www.youtube.com/watch?v=4tP1JA6aH5o
|
p*q
发帖数: 59 | 13 sorry...That previous tensorflow link was a tutorial on classifier (i.e.
predict to
which category an element belongs)....
This one below is on regressor (i.e. predict the next value), which suits
your need more directly...
https://www.tensorflow.org/get_started/input_fn#a_neural_network_model_for_
boston_house_values (tensorflow neural network regressor)
that matlab link is also on future value prediction as well...
https://www.youtube.com/watch?v=4tP1JA6aH5o (matlab neural network
regressor)
【在 f***t 的大作中提到】
: 好的，正在看这个Tensorflow，先表示感谢！
|
f***t
发帖数: 2247 | 14 多谢兄弟，一点心意，别嫌少。
寄信人: deliver (自动发信系统)
标 题: 本站转帐通知单
发信站: BBS 未名空间站 (Thu Sep 28 00:30:30 2017)
来 源: mitbbs.com
ftmit，您好：
您转给 pvq,现金(伪币):100,收取手续费:1
同时附加了如下留言给 pvq.
多谢兄弟，一点心意，别嫌少。
站务
mouse
or
【在 p*q 的大作中提到】
: and here is a youtube tutorial (roughly 5mins long) on predicting using
: MATLAB
: neural network package...
: no programming background needed...just some quite intuitive keyboard/mouse
: operations...
: https://www.youtube.com/watch?v=4tP1JA6aH5o
: You could try to play with the first 10 or 9 vectors to predict the 10th or
: 11th vector that you already have...and see how that would work...
: Good luck!
|
f***t
发帖数: 2247 | 15 好的，我仔细看看先！
【在 p*q 的大作中提到】
: sorry...That previous tensorflow link was a tutorial on classifier (i.e.
: predict to
: which category an element belongs)....
: This one below is on regressor (i.e. predict the next value), which suits
: your need more directly...
: https://www.tensorflow.org/get_started/input_fn#a_neural_network_model_for_
: boston_house_values (tensorflow neural network regressor)
: that matlab link is also on future value prediction as well...
: https://www.youtube.com/watch?v=4tP1JA6aH5o (matlab neural network
: regressor)
|
p*q
发帖数: 59 | 16 Ooops...the forum system breaks the link into two lines...
the "a_neural_network_model_for_boston_house_values" behind that "#" is one
whole part of the link tail...
Or you can just search for "A Neural Network Model for Boston House Values"
section on the linked page....
【在 p*q 的大作中提到】
: sorry...That previous tensorflow link was a tutorial on classifier (i.e.
: predict to
: which category an element belongs)....
: This one below is on regressor (i.e. predict the next value), which suits
: your need more directly...
: https://www.tensorflow.org/get_started/input_fn#a_neural_network_model_for_
: boston_house_values (tensorflow neural network regressor)
: that matlab link is also on future value prediction as well...
: https://www.youtube.com/watch?v=4tP1JA6aH5o (matlab neural network
: regressor)
|
p*q
发帖数: 59 | 17 Thank you so much, bro...that's more than 8 times of all the asset I
previously held!
【在 f***t 的大作中提到】
: 多谢兄弟，一点心意，别嫌少。
: 寄信人: deliver (自动发信系统)
: 标 题: 本站转帐通知单
: 发信站: BBS 未名空间站 (Thu Sep 28 00:30:30 2017)
: 来 源: mitbbs.com
: ftmit，您好：
: 您转给 pvq,现金(伪币):100,收取手续费:1
: 同时附加了如下留言给 pvq.
: 多谢兄弟，一点心意，别嫌少。
: 站务
|
p*q
发帖数: 59 | 18 Tensorflow is a dedicated neural network machine learning environment (built
on python, and hence with all
the other python libraries for all kinds of applications and research works
*)...so it is designed to make many common neural network related tasks easy
(through one or a few library function calls)..TensorFlow is quite scalable
to large distributed computing network (and actually it is a library
originally used by google to compute your every day google search results;
and some part of AlphaGo is also built with TensorFlow )
MATLAB is a more generic environment (whose core idea is based on the
abstraction of vector space) that could do more than machine learning...
MATLAB also has a many packages, but its main focus is not neural network,
unlike TensorFlow...
NOTE:
If you are working in the field of fluid mechanics, here are some python
packages that are current available that might be helpful to ease some
common computing tasks in the field...
+ AeroPython
- http://lorenabarba.com/blog/announcing-aeropython/
- https://github.com/AeroPython
+ fluidlab
- https://pypi.python.org/pypi/fluidlab/
+ fluiddyn
- https://pypi.python.org/pypi/fluiddyn/
+ Scientific Python Computational Fluid Dynamics
- https://www.archer.ac.uk/training/course-material/2014/07/SciPython_
Cranfield/Exercises/CFD_practical.pdf
or in case the above link is broken into two again, here is an equivalent
shortened link,
goo.gl/BjjZGQ
【在 f***t 的大作中提到】
: 好的，多谢！我正在看Tensorflow，好像这个Tensorflow在神经网络方面比matlab更牛
: X。对吗？如果是这样，俺就先试试这个Tensorflow。能早点出结果，就是省钱了，做
: 实验的苦逼啊，整天耳朵里每时每刻都会听到一个单词，money,money,money,....
|
S*****e
发帖数: 6676 | 19 放狗，网上好多MATLAB现成的code，你整理数据跑一下看看 |
H********g
发帖数: 43926 | 20 或曰什么结果都可以预测出来
【在 d****o 的大作中提到】
: 就你这几个点什么NN也救不了
:
: 11
|
|
|
H********g
发帖数: 43926 | |
F*******y
发帖数: 885 | 22 搞笑吗？这么几个数据，就用神经网络？就用Matlab？还 money, money的.
Try Basic exponential smoothing, or triple Exponential smoothing forecasting.
MS Excel 就可以搞定. |
H********g
发帖数: 43926 | 23 这个的相关程序在这里
https://www.mathworks.com/matlabcentral/fileexchange/28684-electricity-load-
and-price-forecasting-webinar-case-study?focused=6789428&tab=function
看起来像是用 时间（日期，小时，是否节假日），温度（drybulb），露点（dewpoint
）这几个数据作为输入，供电负载（sysload）作为目标，训练神经网络。训练的目的
是要能够通过前几个参数来预测sysload。因此只要模拟出来的曲线跟实测曲线吻合度
比较高就行了。
你的目的是什么呢？是要找你6个测点之间的函数关系，还是要找这6个测点的时间周期
性或者在时间序列上的因果性呢？（例如用它们随时间变化的快慢来辅助PID？）
【在 H********g 的大作中提到】
: 这个视频好像跟你问的问题差不多
: https://www.youtube.com/watch?v=P6Re7bIH-gQ
|
p*q
发帖数: 59 | 24 不好意思， 刚看到你的pm...
那个youtube的视频里面的例子(即https://www.youtube.com/watch?v=4tP1JA6aH5o)是
这样的：
matlab里面的input 和 target这两个变量(i.e. 两个矩阵, input是4x26, target是
1x26, )是属于training data...training data的目的是为了生成用于预测的神经网络
的具体参数.
你可以这样想: input矩阵有26个column，每个column有4个点, 每个column对应着
target矩阵里面的一个值...这样input的26个column对应着target里面的26个值...
我们可以想象需要生成这样的一个四变量的方程 fn(arg_1, arg_2, arg_3, arg_4):
现在我们知道的映射(即用于training的input和target变量)是：
fn(input_1x1, input_2x1, input_3x1, input_4x1) = target_1x1
fn(input_1x2, input_2x2, input_3x2, input_4x2) = target_1x2
fn(input_1x3, input_2x3, input_3x3, input_4x3) = target_1x3
fn(input_1x4, input_2x4, input_3x4, input_4x4) = target_1x4
fn(input_1x5, input_2x5, input_3x5, input_4x5) = target_1x5
...
...
...
fn(input_1x26, input_2x26, input_3x26, input_4x26) = target_1x26
上面的input_ixj 表示input矩阵的第i行(即 i-th row)第j列(即 j-th column)的元素
例如 input_2x3表示 input矩阵的第2行第3列的元素...
matlab里面的那个sample (4x4的矩阵)是实际需要用前面training生成的神经网络来预
测的数据, sample里面有4组值, 每组值需要做一个预测...所以一共需要做4个预测...
实际上就是用上面生成的四变量的方程fn,
(fn表示的是生成的神经网络)
来计算 4个prediction
fn (sample_1x1, sample_2x1, sample_3x1, sample_4x1) = prediction_1x1
fn (sample_1x2, sample_2x2, sample_3x2, sample_4x2) = prediction_1x2
fn (sample_1x3, sample_2x3, sample_3x3, sample_4x3) = prediction_1x3
fn (sample_1x4, sample_2x4, sample_3x4, sample_4x4) = prediction_1x4
还有没有什么地方需要表述的更详细些?
【在 p*q 的大作中提到】
: I don't have MATLAB here on my machine (no direct and fast chinese input
: either -- sorry about that)...
: But I found a youtube tutorial on prediction with MATLAB neural network
: package, which seems to be targeting on a similar problem as yours...and
: quite intuitive ( roughly 5mins)...
: (forgot to post the youtube link...here it is...)
: update:
: https://www.youtube.com/watch?v=4tP1JA6aH5o
|
