Search All of the Math Forum:
Views expressed in these public forums are not endorsed by
NCTM or The Math Forum.



Re: neural network
Posted:
Jan 15, 2013 5:12 AM


"Greg Heath" <heath@alumni.brown.edu> wrote in message <kd0hoe$p0s$1@newscl01ah.mathworks.com>... > "Jamaa Ambarak" <jamaa73@yahoo.com> wrote in message <kcqpbo$4dq$1@newscl01ah.mathworks.com>... > > First of all , I have deal with real data that used for lane predication and I have collected my data and now I want to use it to predict lane deviation of a car, I have used time series for prediction and I have selected 3 features for that and they" lateral position, steerangle and speed velocity." the data is about 3 .5 hours of driving , I have collected data from 15 drivers and the drivers are drowses some of them they draft 200 , 360 > >times of the lane,
Do you mean
" the drivers are drowsy and some of them drift out of the lane as much as 200 to 360 times in ~ a 14 minute period"?
200 to 360 consecutive times?
> >so I have selected time windowing from lane deviation cases(out of the lane ) and the > >normal cases(in the lane) for training .the window size is 1second =50 samples
So the 200 to 360 times means 4 to ~7 consecutive seconds in ~ a 14 minute period?
> >and the 3 features are represented in matrix as "150x400" as example 50 samples lateral position, 50 samples speed velocity, and 50 samples steerangle. The 400 are the other cases in one matrix. > > For testing data I want to use sliding window to test whole driver. I use the sliding window for 3 features, the data is big it is about 500.000 for one driver and I want to test sample by sample for example [1 2 3 4 5 6..........150] Transpose. The second one will be > >[2 3 4 5 6.....151] Transpose. And so on ? the matrix will be as [150X405316]. > Have you considered having a IHO net (O=1, I = 3*n, n= 1,2,...) where any combination of position, angle and velocity would correspond to a target of 0 or 1?
I would start there and then increase the input dimension by factors of 3.
If you don't use a validation stopping subset or regularization (trainbr/msereg), N inputoutput training pairs of dimension 3n and 1, respectively, yields
Neq = N*O = N % Number of training equations Nw = (I+1)*H + (H+1)*O % Number of unknown weights to estimate Then Neq > Nw or H < Hub is required where
Hub = floor( (NeqO)/(I+O+1)) = floor((N1)/(3*n+2))
However, it desired that H<< Hub or at least H < Hub/2 Have you considered the significant crosscorrelation lags between the inputs and output? What about the significant crosscorrelation lags between the inputs and the significant autocorrelation lags of each input?
> It is hard for me to believe that you really need 50 samples a second to determine lane deviation. How much subsampling have you investigated? > Greg



