Adaptive Neural Network
in [Matlab]
|
From: Maitha on 10 Jul 2010 07:33 Hi all, I am trying to develop an Adaptive Neural Network, which has 4 inputs and one output. The network should predict for next year (365 days). Training data sample : 4383 Testing data : 365 I have used newfftd function as follows: net = newfftd(pn,tn,[0 1], 5) % pn : normalized input patterns, [-1, 1] % tn : normalized output [-1,1] Based on what the delay input should be sit up? Should be [0 365] because I want to predict incoming 365 days? ps : Matlab (R2008a) is used Thanks in advance Maitha
From: Greg Heath on 12 Jul 2010 04:52 On Jul 10, 7:33 am, "Maitha " <maitha_...(a)hotmail.com> wrote: > Hi all, > > I am trying to develop an Adaptive Neural Network, which has 4 inputs and one output. The network should predict for next year (365 days). > > Training data sample : 4383 > Testing data : 365 > > I have used newfftd function as follows: > net = newfftd(pn,tn,[0 1], 5) > > % pn : normalized input patterns, [-1, 1] > % tn : normalized output [-1,1] > > Based on what the delay input should be sit up? Should be [0 365] because I want to predict incoming 365 days? No. That only uses the current day (0) and the same day of the previous year (365). Maybe you meant [0:365]? Plot your training data to determine the order of a reasonable polynomial fit. fft your data to determine if there are any dominant periodicities. This can help in a trial and error search for how much past data is needed for prediction. > ps : Matlab (R2008a) is used Hope this helps. Greg
From: kk KKsingh on 12 Jul 2010 06:12 Greg Heath <heath(a)alumni.brown.edu> wrote in message <8c9af343-8751-4e98-9e33-1f1c1835397f(a)j13g2000yqj.googlegroups.com>... > On Jul 10, 7:33 am, "Maitha " <maitha_...(a)hotmail.com> wrote: > > Hi all, > > > > I am trying to develop an Adaptive Neural Network, which has 4 inputs and one output. The network should predict for next year (365 days). > > > > Training data sample : 4383 > > Testing data : 365 > > > > I have used newfftd function as follows: > > net = newfftd(pn,tn,[0 1], 5) > > > > % pn : normalized input patterns, [-1, 1] > > % tn : normalized output [-1,1] > > > > Based on what the delay input should be sit up? Should be [0 365] because I want to predict incoming 365 days? > > No. That only uses the current day (0) and the > same day of the previous year (365). > > Maybe you meant [0:365]? > > Plot your training data to determine the > order of a reasonable polynomial fit. > > fft your data to determine if there are > any dominant periodicities. > > This can help in a trial and error search for > how much past data is needed for prediction. > > > ps : Matlab (R2008a) is used > > Hope this helps. > > Greg Please check the Matlab Help on ANN tool box they gave example of same problem over there kumar
From: Maitha on 14 Jul 2010 21:15 "kk KKsingh" <akikumar1983(a)gmail.com> wrote in message <i1epq1$rih$1(a)fred.mathworks.com>... > Greg Heath <heath(a)alumni.brown.edu> wrote in message <8c9af343-8751-4e98-9e33-1f1c1835397f(a)j13g2000yqj.googlegroups.com>... > > On Jul 10, 7:33 am, "Maitha " <maitha_...(a)hotmail.com> wrote: > > > Hi all, > > > > > > I am trying to develop an Adaptive Neural Network, which has 4 inputs and one output. The network should predict for next year (365 days). > > > > > > Training data sample : 4383 > > > Testing data : 365 > > > > > > I have used newfftd function as follows: > > > net = newfftd(pn,tn,[0 1], 5) > > > > > > % pn : normalized input patterns, [-1, 1] > > > % tn : normalized output [-1,1] > > > > > > Based on what the delay input should be sit up? Should be [0 365] because I want to predict incoming 365 days? > > > > No. That only uses the current day (0) and the > > same day of the previous year (365). > > > > Maybe you meant [0:365]? > > > > Plot your training data to determine the > > order of a reasonable polynomial fit. > > > > fft your data to determine if there are > > any dominant periodicities. > > > > This can help in a trial and error search for > > how much past data is needed for prediction. > > > > > ps : Matlab (R2008a) is used > > > > Hope this helps. > > > > Greg > > Please check the Matlab Help on ANN tool box they gave example of same problem over there > > kumar Greg & Kumar thank you for replaying Maybe you meant [0:365]? YES I plot training data (4 inputs) without output, I got graph but how can I interpret it to determine the order of a reasonable polynomial fit?? I covert training data (4 inputs) to Discrete Fourier transform, I got very strange graph??!! Should I deal with each input separately?
From: Greg Heath on 19 Jul 2010 06:11 On Jul 14, 9:15 pm, "Maitha " <maitha_...(a)hotmail.com> wrote: > "kk KKsingh" <akikumar1...(a)gmail.com> wrote in message <i1epq1$ri...(a)fred..mathworks.com>... > > Greg Heath <he...(a)alumni.brown.edu> wrote in message <8c9af343-8751-4e98-9e33-1f1c18353...(a)j13g2000yqj.googlegroups.com>... > > > On Jul 10, 7:33 am, "Maitha " <maitha_...(a)hotmail.com> wrote: > > > > Hi all, > > > > > I am trying to develop anAdaptiveNeuralNetwork, which has 4 inputs and one output. The network should predict for next year (365 days). > > > > > Training data sample : 4383 > > > > Testing data : 365 > > > > > I have used newfftd function as follows: > > > > net = newfftd(pn,tn,[0 1], 5) > > > > > % pn : normalized input patterns, [-1, 1] > > > > % tn : normalized output [-1,1] > > > > > Based on what the delay input should be sit up? Should be [0 365] because I want to predict incoming 365 days? > > > > No. That only uses the current day (0) and the > > > same day of the previous year (365). > > > > Maybe you meant [0:365]? > > > > Plot your training data to determine the > > > order of a reasonable polynomial fit. > > > > fft your data to determine if there are > > > any dominant periodicities. > > > > This can help in a trial and error search for > > > how much past data is needed for prediction. > > > > > ps : Matlab (R2008a) is used > > > > Hope this helps. > > > > Greg > > > Please check the Matlab Help on ANN tool box they gave example of same problem over there > > > kumar > > Greg & Kumar thank you for replaying > > Maybe you meant [0:365]? YES > > I plot training data (4 inputs) without output, I got graph but how can I interpret it to determine the order of a reasonable polynomial fit?? > > I covert training data (4 inputs) to Discrete Fourier transform, I got very strange graph??!! Should I deal with each input separately?- Hide quoted text - Most real-world NN design involves search by trial and error guided by common sense. If you look at a plot of data cross-eyed and can visualize a smooth polynomial approximation that fits the general data trend, you can often guess at a reasonable initial value in a search for the order of that polynomial. If need be you can resort to a few illustrative fits on examples of polynomial + noise help polyfit However, since this is just a rule of thumb guide, don't get hung up on trying to optimize a polynomial fit. Since you need at least N+1 points to estimate the N+1 coefficients of an Nth order polynomial, that would be a reasonable initial value in a search for the minimum number of delays in a TDNN model. If you plot the fft of zero mean data sampled at a rate of Fs = 1/dt = 1/timespacing help zscore the sampled frequency spectrum will have a sample frequency spacing of df = Fs/N. If the lowest significant value is at frequency f0, then it is reasonable to assume that it will take at least N > Fs/f0 delays for a reasonable TDNN model. Hope this helps. Greg
|
Pages: 1 Prev: remove duplicate rows in sparse matrix Next: Reading Serial data through Simulink Input Packet |