How to calculate correctly Ntrneq, Nw, Hub, Hmax, MSEgoal for MULTIPLE SEQUENCES in neural networks?
2 次查看(过去 30 天)
显示 更早的评论
I am trying to find the minimum number of nodes in hidden layer that satisfies my MSEgoal.
I have my input and output cell arrays of {1x491} with [2x33] values. Technically, I have 33 time series of length 491 with padded NaNs with 2-inputs and 2-outputs.
If I do:
x = cell2mat(X);
t = cell2mat(T);
[ I N ] = size(x); % [ 2 16203] % Including the NaN values
[ O N ] = size(t);
N = the total number of point that will be use for training, including the NaNs.
However, I assume the I do not want to count the NaNs to calculate Ntrneq so I removed them:
[ I N ] = size(x(:,any(~isnan(x)))); % [ 2 12116] % Removed NaN values from the count
[ O N ] = size(t(:,any(~isnan(t))));
However, Ntreq is huge = 24202, with a HUB of 967 and Hmax of 96. Which is also huge compared to h=10 that have been using in a pilot test where error are in the order of 1/10000.
But if I do:
[ I N ] = size(X); % [ 1 491]
[ O N ] = size(T);
Am I only calculating Ntrneq, Nw, Hub, Hmax as if I only had 1 time series but still I&O =2?
This is my complete code:
x = cell2mat(X);
t = cell2mat(T);
%[ I N ] = size(x); % [ 2 16203] % Including the NaN values
% [ O N ] = size(t);
[ I N ] = size(x(:,any(~isnan(x)))); % [ 2 12116] % Removed NaN values from the count
[ O N ] = size(t(:,any(~isnan(t))));
% [ I N ] = size(X); % [ 1 491] % Removed NaN values from the count
% [ O N ] = size(T);
Ntrn = N % 12116 use all data set only for finding h
rng('default')
FD = 10:15; % Determine with pilot chracterization. It will be updated with auto-correlation of targets
ID = 10:15; % Determine with pilot chracterization. It will be updated with cross-correlation of targets and inputs
steps=15; % Lag same es MXFD or MXID = max(FD) or max(ID)
NFD = length(FD); NID = length(ID);
Ntrneq = (Ntrn-steps)*O %24202, I deleted the lags
Hub = -1+ceil( (Ntrneq-O) / ((NID*I)+(NFD*O)+1)) % 967
Hmax = floor(Hub/10) % 96
Hmin = 0; dH = 1; Ntrials = 10; j=0;
rng(0)
for h = Hmin:dH:Hmax
j = j+1;
if h == 0
net = narxnet( ID, FD, [] );
Nw = ( NID*I + NFD*O + 1)*O;
else
net = narxnet( ID, FD, h );
Nw = ( NID*I + NFD*O + 1)*h + ( h + 1)*O;
end
Ndof = Ntrn-Nw;
[ Xs Xi Ai Ts ] = preparets(net,X,{},T);
ts = cell2mat(Ts);
xs = cell2mat(Xs);
MSE00s = mean(nanvar(ts',1)); %I omitted the NaNs
MSE00as = mean(nanvar(ts',0));
MSEgoal = 0.01*Ndof*MSE00as/Ntrneq;
MinGrad = MSEgoal/10; % or MSEgoal/100;
net.trainParam.goal = MSEgoal;
net.trainParam.min_grad = MinGrad;
net.divideFcn = 'dividetrain';
net.performParam.normalization = 'standard'; % Does this affect the performance because my MSEgoal is in the normal range.
for i = 1:Ntrials
net = configure(net,Xs,Ts);
[ net tr Ys ] = train(net,Xs,Ts,Xi,Ai);
ys = cell2mat(Ys);
stopcrit{i,j} = tr.stop;
bestepoch(i,j) = tr.best_epoch;
MSE = mse(ts-ys);
MSEa = Ntrneq*MSE/Ndof;
R2(i,j) = 1-MSE/MSE00s; %? different ranges?
R2a(i,j) = 1-MSEa/MSE00as; %? different ranges?
end
stopcrit = stopcrit
bestepoch = bestepoch
R2 = R2
R2a = R2a
%Totaltime = toc
Lastly, in my code I am not normalizing my inputs and outputs but I used:
net.performParam.normalization = 'standard';
becasue my outputs have different ranges. Does that mean that my
net.trainParam.goal = MSEgoal;
net.trainParam.min_grad = MinGrad;
are on different ranges (data vs training). so my R2, too?
I have posted this question about using zscore on my data set:
because I couldn't make it work.
I have attached my X and T.
Thank you,
0 个评论
回答(0 个)
另请参阅
类别
在 Help Center 和 File Exchange 中查找有关 Modeling and Prediction with NARX and Time-Delay Networks 的更多信息
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!