Use of target data in closed loop NARX network

Hello,

I have a question regarding NARX network.

I have created a NARX network (see below) and that works fine. Also if I provide new data that has not been use for training I get good results. The network has one input X and one output Y.

inputDelays = 1:15; %20 feedbackDelays = 1:15; %20 hiddenLayerSize = 25; %25

net1 = narxnet(inputDelays,feedbackDelays,hiddenLayerSize);

[inputs,inputStates,layerStates,targets] = preparets(net1,inputSeries(1:20000),{},targetSeries(1:20000));

net1.divideParam.trainRatio = 70/100; net1.divideParam.valRatio = 15/100; net1.divideParam.testRatio = 15/100;

[net1,tr] = train(net1,inputs,targets,inputStates,layerStates); outputs = net1(inputs,inputStates,layerStates);

Then I want to use the NARX network as closed. I use a delay of 15 to preparing the inputStates. Then I want to use the NARX network and the X input variable to create the Y variable.

netc = closeloop(net1);

index = 1600;

[inputs,inputStates,layerStates,targets] = preparets(netc,inputSeries(index-15-4+1:index),{},targetSeries(index-15-4+1:index));

outputs = netc(inputs,inputStates,layerStates);

When I use this code I get a good starting condition/layerStates conform the data. Then I use the next four X inputs inputSeries(index-3:index) (inputs) and the NARX network to calculate the next four Y outputs targetSeries(index-3:index) (targets).

I would expect that the calculation of the output is independed on targetSeries(index-3:index) because it should be depended on the starting condition layerStates and inputSeries(index-3:index).

However when the targetSeries(index-3:index) has different value also the output values of the NARX closed network is different. It looks like that targetSeries(index-3:index) is somehow used in the calculation and this is not desired. Does anyone have any idea what goes wrong?

Many thanks in advance and best wishes