Hi,
Don't know if anybody answered your question but it does appear that Matlab can train a NARX net with multiple trials (or sequences) recorded at the same time although there are a few caveats. The main point is that you have to format your data correctly in the cell format however. It's a terrible shame that Matlab hasn't really addressed this in any documentation as it is not clear to anybody apparently and the examples do not really address this at all.
Step 1: So for example if you have a 1x10 cell array, each of the cells (1 through 10) represents the timesteps of your data. If you wanted 20 timesteps you would have a 1x20 cell array and so on and so forth. This is the case for both your inputs and targets.
Step 2: Ok now to look at each cell in the 1x10 cell array (still using 10 timesteps). Let's say I have 10 inputs corresponding to 1 target of double data and I have 20 trials or 20 sets of data recorded at one time. Each individual cell in the Input array will be 10x20 double. Each cell in the Target array will be 1x20 double. Obviously doesn't have to be double but giving an example.
Step 3: So you have a 1x10 cell array. Each of these cells contain either 10x20 double or 1x20 double data for the input or target array.
Step 4: Now each cell must be 10x20 double or 1x20 in size. If your inputs or targets change in size then you will need to pad with nans to make each cell the same size.
To pad the targetseries and inputseries arrays you can use the code below:
%Pad nans to targetseries and inputseries
maxSize = max(cellfun(@numel,targetSeries)); %# Get the maximum vector size
fcn = @(x) [x nan(1,maxSize-numel(x))]; %# Create an anonymous function
targetSeriesNN = cellfun(fcn,targetSeries,'UniformOutput',false); %# Pad each cell with NaNs
SizeofInputArray=size(inputSeries{1,1});
SizeofInputArray=SizeofInputArray(1,1);
maxSize = max(cellfun(@length,inputSeries)); %# Get the maximum vector size
fcn = @(x) [x nan(SizeofInputArray,maxSize-size(x,2))]; %# Create an anonymous function
inputSeriesNN = cellfun(fcn,inputSeries,'UniformOutput',false); %# Pad each cell with NaNs
Step 5: So now you are ready to train your data. The best way to do this intuititvely is to use the usual split 70,15,15 in the input/target data both at each timestep and also in time. So for example the first cell (and the first timestep data) may be found in the 1x1 cell. So the simple way would be to split your first timestep data 70,15,15 for both the input and target data and then subsequent timesteps in the same way. Your first timesteps has 20 trials of data and so the first 14 inputs would be used for training and then the next 3 would be used for test and the next 3 for validation. This also needs to be done in time which makes this tricky (ideally the first 70% of the timestep data would be used for training and then the last 30% would be used for test and validation. So it seems divideblock needs to be combined with a divide mode of sample for this. The best I have come up with is net.divideMode = 'sampletime'. Although maybe somebody could correct me on whether that is true or whether I am wrong. I would appreciate any feedback anybody has?
Step 6: Train away.
One issue I have is that my inputs are not always of the same size and thus I get a nan when doing the prediction or training output. The issue stems from the fact that when using the data from previous delays it uses the same exact index in the array and so the nan causes issues. I naively thought the training used all the inputs at every index.
So if my first input is 1x15 although padded to 1x20 using nans
My second input is 1x20.
My third input is 1x20.
My output target will be nan from 1x15 to 1x20 will be nan because of the nan padded in the first input.
