Narxnet with Many timesteps to one time step

Hi All,

I have sensor readings with high frequency for an hour for example and I would like to predict one value out of these input measurements. How can I use narxnet to do that?

I do not want to concatenate all the inputs in one array and feed it to the network as one time step. In short how can I implement Many time steps to one value using narxnet.

consider Input is 100,000 * 3 where 100,000 is time steps and 3 is the number of sensors.

output is 10

i.e each 10,000 reading correspond to one output.

Thanks in advance

Answer 1

[Greg Heath]

1. Plot subsampled series so that you can see what range of subsampling rates makes sense.

2. Plot the autocorrelation function of the chosen subsampled series to obtain a subset of significant lags.

3. Given the chosen subsampled series and significant lags, the formidable problem is now reduced to a standard one.

Hope this helps.

Thank you for formally accepting my answer

Greg

[amr alanwar]

Thank you for your answer. But I do not want to formulate it as one to one problem. The whole time steps are important. Other tools like keras support time steps as a third dimension (many to one problem). Is there another way to solve the problem with Narxnet ?

Thanks again!

[Greg Heath]

It is very hard to believe that subsampling 100,000 points will lead to irrelevance.

Greg

[amr alanwar]

Thanks for follow up. I can do subsampling with a relevant subsampling. But I do not want to formulate one to one problem as it does not make sense. Until now I see that Matlab does not support timesteps in the input. right?

For example, The LSTM network in Keras expects the input data (X) to be provided with a specific array structure in the form of: [samples, time steps, features]. Thank you

[Greg Heath]

Until now I see that Matlab does not support timesteps in the input. right?

> Yes.