| View previous topic :: View next topic |
| Author |
Message |
Guest
|
 Posted: Tue Aug 26, 2008 12:33 pm Post subject: forecasting:validation set |
 |
|
hi folks,
i am doing forecasting using neural network. i have some
confusion.suppose, in my training set, i give 2D input x(t-1) and
x(t-2) and my desired response is x(t). My question is how can I make
sure that my output is coming x(t) if I will give input x(t-1) and
x(t-2)?
Another question please. for checking this network with the validation
set, will i give the same desired response x(t) of training set or I
have to give new desired response (it's also x(t), but for validation
set x(t)) ?
Thanks. I will be very grateful if you will reply my mail.
ashraf |
|
| |
|
 Back to top |
Greg Heath
Guest
|
| Posted: Tue Aug 26, 2008 12:59 pm Post subject: Re: forecasting:validation set |
|
|
On Aug 26, 8:33 am, ashraf.p...@gmail.com wrote:
| Quote: | hi folks,
i am doing forecasting using neural network. i have some
confusion.suppose, in my training set, i give 2D input x(t-1) and
x(t-2) and my desired response is x(t). My question is how can I make
sure that my output is coming x(t) if I will give input x(t-1) and
x(t-2)?
Another question please. for checking this network with the validation
set, will i give the same desired response x(t) of training set or I
have to give new desired response (it's also x(t), but for validation
set x(t)) ?
Thanks. I will be very grateful if you will reply my mail.
ashraf
|
See my answer to a similar request.
http://groups.google.com/group/comp.soft-sys.matlab/msg/bdd8160316ad1174?hl=en
Hope this helps.
Greg |
|
| |
|
| Back to top |
Guest
|
| Posted: Tue Aug 26, 2008 9:46 pm Post subject: Re: forecasting:validation set |
|
|
On Aug 26, 9:59 am, Greg Heath <he...@alumni.brown.edu> wrote:
| Quote: | On Aug 26, 8:33 am, ashraf.p...@gmail.com wrote:
hi folks,
i am doing forecasting using neural network. i have some
confusion.suppose, in my training set, i give 2D input x(t-1) and
x(t-2) and my desired response is x(t). My question is how can I make
sure that my output is coming x(t) if I will give input x(t-1) and
x(t-2)?
Another question please. for checking this network with the validation
set, will i give the same desired response x(t) of training set or I
have to give new desired response (it's also x(t), but for validation
set x(t)) ?
Thanks. I will be very grateful if you will reply my mail.
ashraf
See my answer to a similar request.
http://groups.google.com/group/comp.soft-sys.matlab/msg/bdd8160316ad1...
Hope this helps.
Greg
|
Dear Greg,
Many many thanks for the link as it is very helpful for my
understanding regarding training set and validation set for
prediction. But could you kindly give me a vision regarding test set?
As I know, for prediction, training set is used for weight
selection,validation set is used for tuning different model
parameters(e.g., number of neuron in hidden layer) and test set is
used for prediction and final error measurement(in my case). what
should be the m and n days for testing input and testing output?
Should have there any relation with learning(training + validation)
data or it is independent?
Your cooperation is highly appreciable. Looking forward to reply your
mail.
ashraf |
|
| |
|
| Back to top |
Greg Heath
Guest
|
| Posted: Wed Aug 27, 2008 3:55 am Post subject: Re: forecasting:validation set |
|
|
On Aug 26, 5:46 pm, ashraf.p...@gmail.com wrote:
| Quote: | On Aug 26, 9:59 am, Greg Heath <he...@alumni.brown.edu> wrote:
On Aug 26, 8:33 am, ashraf.p...@gmail.com wrote:
hi folks,
i am doing forecasting using neural network. i have some
confusion.suppose, in my training set, i give 2D input x(t-1) and
x(t-2) and my desired response is x(t). My question is how can I make
sure that my output is coming x(t) if I will give input x(t-1) and
x(t-2)?
Another question please. for checking this network with the validation
set, will i give the same desired response x(t) of training set or I
have to give new desired response (it's also x(t), but for validation
set x(t)) ?
Thanks. I will be very grateful if you will reply my mail.
ashraf
See my answer to a similar request.
http://groups.google.com/group/comp.soft-sys.matlab/msg/bdd8160316ad1...
Hope this helps.
Greg
Dear Greg,
Many many thanks for the link as it is very helpful for my
understanding regarding training set and validation set for
prediction. But could you kindly give me a vision regarding
test set?
|
If you assume the Ntst errors are i.i.d. zero-mean Gaussian,
with variance s0^2, then SSEtst has a CHISQUARE distribution
http://en.wikipedia.org/wiki/Chi-square_distribution
with mean = Ntst*s0^2, variance var = 2*Ntst*s0^4 and standard error
se = sqrt(var/Ntst). For a precise estimate of SSEtst
choose
se/mean = sqrt(2)/Ntst << 1
| Quote: | As I know, for prediction, training set is used for weight
selection,
|
To ESTIMATE the WEIGHTS of the candidate models.
| Quote: | validation set is used for tuning different model
parameters(e.g., number of neuron in hidden layer)
|
.... H, but also Nd, m and n in order to SELECT the
BEST candidate MODEL(s).
| Quote: | and test set is used for prediction and final error . measurement(in my case).
|
To ESTIMATE the generalization ERROR (i.e., estimate
the expected error on NONDESIGN data)
| Quote: | what should be the m and n days for testing input and
testing output?
|
You have already specified n = 2, m =1 in your original
post. If you choose another value for m you can use the
validation set to estimate a good value for n.
| Quote: | Should have there any relation with learning(training +
validation) data or it is independent?
|
You partition the data into N = Nd + Ntst. There will be
a correlation distance d which will make the last several
design points and first several test points correlated.
However, this is usually ignored.
| Quote: | Your cooperation is highly appreciable.
|
appreciated
| Quote: | Looking forward to reply your mail.
|
Hope this helps.
Greg |
|
| |
|
| Back to top |
marshal
Guest
|
| Posted: Thu Aug 28, 2008 4:40 pm Post subject: Re: forecasting:validation set |
|
|
On Aug 26, 9:55 pm, Greg Heath <he...@alumni.brown.edu> wrote:
| Quote: | On Aug 26, 5:46 pm, ashraf.p...@gmail.com wrote:
On Aug 26, 9:59 am, Greg Heath <he...@alumni.brown.edu> wrote:
On Aug 26, 8:33 am, ashraf.p...@gmail.com wrote:
hi folks,
i am doing forecasting using neural network. i have some
confusion.suppose, in my training set, i give 2D input x(t-1) and
x(t-2) and my desired response is x(t). My question is how can I make
sure that my output is coming x(t) if I will give input x(t-1) and
x(t-2)?
Another question please. for checking this network with the validation
set, will i give the same desired response x(t) of training set or I
have to give new desired response (it's also x(t), but for validation
set x(t)) ?
Thanks. I will be very grateful if you will reply my mail.
ashraf
See my answer to a similar request.
http://groups.google.com/group/comp.soft-sys.matlab/msg/bdd8160316ad1....
Hope this helps.
Greg
Dear Greg,
Many many thanks for the link as it is very helpful for my
understanding regarding training set and validation set for
prediction. But could you kindly give me a vision regarding
test set?
If you assume the Ntst errors are i.i.d. zero-mean Gaussian,
with variance s0^2, then SSEtst has a CHISQUARE distribution
http://en.wikipedia.org/wiki/Chi-square_distribution
with mean = Ntst*s0^2, variance var = 2*Ntst*s0^4 and standard error
se = sqrt(var/Ntst). For a precise estimate of SSEtst
choose
se/mean = sqrt(2)/Ntst << 1
As I know, for prediction, training set is used for weight
selection,
To ESTIMATE the WEIGHTS of the candidate models.
validation set is used for tuning different model
parameters(e.g., number of neuron in hidden layer)
... H, but also Nd, m and n in order to SELECT the
BEST candidate MODEL(s).
and test set is used for prediction and final error . measurement(in my case).
To ESTIMATE the generalization ERROR (i.e., estimate
the expected error on NONDESIGN data)
what should be the m and n days for testing input and
testing output?
You have already specified n = 2, m =1 in your original
post. If you choose another value for m you can use the
validation set to estimate a good value for n.
Should have there any relation with learning(training +
validation) data or it is independent?
You partition the data into N = Nd + Ntst. There will be
a correlation distance d which will make the last several
design points and first several test points correlated.
However, this is usually ignored.
Your cooperation is highly appreciable.
appreciated
Looking forward to reply your mail.
Hope this helps.
Greg
|
Hi there,
Thats a nice info. Thanks a ton. I was thinking on similar lines to
determine optical flow. But just to get started, I thought of tackling
a relatively simple problem. I want to determine the position of pulse
at time t given that I know where it is in 2-dimensional space at time
t-1 and t-2.
I want to resolve it with NN and I was wondering what could be the
elements of my input data. I am thinking about what can be my input
and output data. I am still a novice in the field of NN. I just have a
problem here or rather I should say I just thought of making a
problem.
I will be thankful if anybody could suggest me what all can be inputs
and outputs for this problem if I want to predict the position of a
pulse signal at time t.
Thanks,
Anup |
|
| |
|
| Back to top |
Greg Heath
Guest
|
| Posted: Thu Aug 28, 2008 6:56 pm Post subject: Re: forecasting:validation set |
|
|
On Aug 28, 12:40 pm, marshal <anupsatishsaw...@gmail.com> wrote:
------SNIP
| Quote: | Thats a nice info. Thanks a ton. I was thinking on similar lines to
determine optical flow. But just to get started, I thought of tackling
a relatively simple problem. I want to determine the position of pulse
at time t given that I know where it is in 2-dimensional space at time
t-1 and t-2.
I want to resolve it with NN and I was wondering what could be the
elements of my input data. I am thinking about what can be my input
and output data. I am still a novice in the field of NN. I just have a
problem here or rather I should say I just thought of making a
problem.
I will be thankful if anybody could suggest me what all can be inputs
and outputs for this problem if I want to predict the position of a
pulse signal at time t.
|
A pulse of what?
What caused it?
What causes it to move?
Is there a differential equation that describes the movement?
Greg |
|
| |
|
| Back to top |
|
