Test trained NARX neural network with new external input

I have a time series NARX neural network that is already trained on an external input, x(t), as well as the output's value at certain timesteps before (d past values of y(t)). The network architecture is illustrated in the document I've uploaded.

Now I would like to take the trained network, input a new external input x(t), and get a time series prediction output. To prepare the inputs for getting the outputs I would need to use preparets to get the correct form of x(t), inputStates and layerStates, but it seems from the documentation that a target series is required for this. Since this is a test, I don't have a target series, what should I do?

Furthermore, since I have no clue what the output y(t) would be (since this a test), would it make sense for me to remove all feedback (past values of y(t)) into the hidden layers?

Thanks,

Mei

Answer 1

[Greg Heath]

If the new data immediately follows the data used to design and test the net, the following syntax should have been used

[ net tr Ys Es Xsf Asf ] = train(net,Xs,Ts,Xi,Ai);

Xinew = Xsf; Ainew = Asf;

Ysnew = net(Xsnew,Xinew,Ainew);

Otherwise

Xinew = Xnew(:,1:d); Xsnew = Xnew(:,d+1:end)

but Ainew is not known.

I would try the mean of the previously used test target data rather than use zeros. Perhaps several designs using values in the interval [mean-stdv,mean+stdv] would be useful.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Tim]

Hey Greg, sorry I still don't get it. The inputs of the net() function to predict values are:

-XsNew: I think this is the new, unseen data. This data consists of a cell array with two rows. Row two contains the classlabels. BUT I dont know them !? How can I understand this?

I give you an example of what I mean:

1. Item one
2. Item twoLoading the training classifier (NARX) with closed loop => netc
3. 2. New data (feature Values) from a time series comes in
4. 3. netc(A,B,C)with

• Item one
• Item twoA: Cell Array with two rows: ROW1: Feature Values (thats ok); ROW2: Class Values (????? How should I know them)

Here is my problem with your

Xinew = Xnew(:,1:d); Xsnew = Xnew(:,d+1:end)

explanation. I need the information about the classees, otherwise there are not enough inputs (the dimension does not match)

Thanks!

Answer 2

[mehdi asgharzadeh]

I honestly dont think these poor MATLAB guys know what they are talking about here. As everyone esle has pointed out, we don't have the output target array when we are in pridicting mode (closed loop). Training is OK but predicting this way only works when you have future values of exogenous time series but not the target itself (for some reasons) and want to estimate them. If you are trying to predict stock's future values and things like that, sorry this is not going to help you.

[musa butt]

so What is the procedure to predict from pretrained neural network.I have trained one now what is the approriate way to genrate prediction from that network.Thanks

[Philipp John]

This article helped me, to understand how to make a prediction from a trained NARXnet: https://towardsdatascience.com/building-a-narx-in-matlab-to-forecast-time-series-data-f60561864874

In this article you find a Tutorial on how to train a NARXnet and how to forecast time series data with it. There is also a link to the csv data, which he used for training and prediction.

Maybe it also helps you

Answer 3

[Jonathan LeSage]

Here is an example from the Neural Network toolbox documentation that you might find helpful:

• http://www.mathworks.com/help/nnet/ug/design-time-series-narx-feedback-neural-networks.html

To simulate an output from your NARX neural network, you need both the initial input delay states and the initial layer delay states. The preparets helps to simplify this task. You can use the inputs and outputs that you are training your neural network model with to get you started. Additional useful documentation:

• http://www.mathworks.com/help/nnet/ref/preparets.html
• https://www.mathworks.com/help/nnet/ref/sim.html

In the second link above, you might find the section, entitled "Simulate NARX Time Series Networks", useful as well.

Hope this helps to get you started!

[Mei]

In all of the documentation/examples, the NARX is tested on the same exogenous input it is trained on.

In my case, I have a completely different set of exogenous inputs (but in the same format as when the NARX network was trained of course) that I would like to test.

I attempted to take the trained NARX network and remove all the feedback connections and retain all the layer and input weights calculated from the previous training. Effectively I now have a input-output network with layer and input weights from before. I then put in the new exogenous inputs, but it didn't work - all the predictions are one and the same (0.5 at all timepoints).

This is what I did:

inputSeries = UsableConcs_Ternary; % My new exogenous input

neto = net; % net is the previously trained neural network (with feedback into both hidden layers as well as a different set of exogenous inputs)

neto.layerConnect=[false false false; true false false; false true false]; % I removed the feedback connections since I don't have a target series for this new set of exogenous inputs anyway (I want to simulate,not train)

view(neto)

[xc,xic,aic,~,~,shift] = preparets(neto,inputSeries,{}); % Again, I don't have a target series here and preparets seems to require it yc = neto(xc,xic,aic); % Simulate the trained neural network without feedback

[Russell Grm]

Hi everyone,

This question has been left unanswered for so long and so the discussion is sill open. Could any expert please prvide a proper answer or piece of code that would assisst us?

Thanks in advance,

[Greg Heath]

What EXACTLY do you not understand ?

What EXACTLY do you want to do ?

I do not understand what I have written is not perfectly clear.

Greg

[Muhammmad Ali]

Hi,

I have a similar problem. I used the nn toolbox to train a NARX neural network. I want my network to predict the output at the next time interval. Lets take the well know example of the magnetic leviation from matlab. For this example we have one input (current) and one output (leviated magnetic position). I want to deploy the trained network such that when I give in the input (current) and the delayed input and output states, it should give me the output (leviated magnetic position) at the next time interval.

But when I deploy my network using the toolbox, it generates a function with different input arguments like,

x1 input 1: from my understanding this is the input to our system (current in out case)

x2 input 2: from my understanding this is the output to our system (we only have one input, so what is this???)

xi1 delayed input state

xi2 delayed output state

What I don't understand is, what exactly is x2? Why should I give the output? I don't have this output, that is why I am using the network to predict the output.

Thank you!

Answer 4

[Mohsen Zabihi]

I have the same question. anyone can help?

[Christopher Stokely]

MATLAB has fundamental errors in their NARX implementation, despite the commentary by GH. I will ask their developers about it immediately.

In the documentation MAGLEV example on forecasting, i adjusted a single target state that was not to be used as an input since it is considered unknown. the prediction immediately went unstable.

[Christoph]

Did you figured something out?

I have a simple SISO system. It work's well when i train the network directly in closed loop and then simulate on new unseen input data. But when i want to train it in open-loop as intended and simulate in closed loop form, it went unstable.

[salma ben ftima]

Hi, I have the same issue discussed here but still don't get solution, please how do you test your narx nn only with new input since it requiers a target while doing the preparets