Recurrent Neural Network output not working on test data (it converges abruptly after a number of steps)

Question

Deskj 2020-10-22

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/623223-recurrent-neural-network-output-not-working-on-test-data-it-converges-abruptly-after-a-number-of-st

I have a dataset of 5 inputs and 2 outputs of 11520 samples and i'm using it to train a recurrent neural network and see how it works on some test data of 2881 samples.

From the learning data, I use my input 5x11520 matrix to generate X_learn (1x11520 cells of 5x1 double) and my output 2x11520 matrix to generate T_learn (1x11520 cells of 2x1 double).

I created and trained the network as shown in the official documentation. My code is the following:

net = layrecnet([1:5],[3]); %a network with one hidden layer of 3 nodes and 5 delay steps
[Xs,Xi,Ai,Ts] = preparets(net,X_learn,T_learn);
net = train(net,Xs,Ts,Xi,Ai);

This is what the network looks like with view(net):

To see how the trained network behaves on both the learning and test data, i do the following.

Fist, to evaluate it on the learning data:

Y_learn=net(X_learn); %feeds the inputs to the network
c_learn=cell2mat(Y_learn); %turns the result into a matrix of 2x11520

And if i plot, let's say the output number 1 (of 2), both from the target and the net's output, here the result:

To test the net on the test data, I used my input 5x2881 matrix of the test samples to generate X_test (1x2881 cells of 5x1 double). Then I feed it to the network:

Y_test=net(X_test); %feeds the test inputs to the network
c_test=cell2mat(Y_test); %turns the result into a matrix of 2x2881

But when i plot both the target and the net's output from the test data, here's the result:

Apparently it works on the first 381 samples (kind of) and then the net's output converges to a specific value. Same happens with the output 2 (only the value it converges to changes).

I'm obviously missing something important, but i can't understand what.

Thank you