Time Series Multiple Inputs

Hello,

I am confused about the multiple values of inputs and targets in a neural network. More so about whether Matlab can do it and how would I correctly format the data.

Let's go to the pollution dataset:

LOAD pollution_dataset.MAT loads these two variables:

pollutionInputs - a 1x508 cell array of 8x1 vectors representing

eight measurements over 508 timesteps.

1. Temperature

2. Relative humidity

3. Carbon monoxide

4. Sulfer dioxide

5. Nitrogen dioxide

6. Hydrocarbons

7. Ozone

PollutionTargets - a 2x508 cell array of 3x1 vectors representing

a 219 timesteps of three kinds of mortality to be predicted.

1. Total mortality

2. Respiratory mortality

3. Cardiovascular mortality

How would I enter or refine the data for multiple trials at exactly same time? e.g. I measured the eight measurements above from several different people at the same time and then got multiple outputs as well? Can somebody explain how the data would be formatted correctly? Could you show me a quick example using the pollution dataset.

Thank you.

[Srivardhan Gadila]

Can you provide more information on whether you are trying to predict future/current data based on past inputs or is it fitting some neural network to some inputs and outputs(i.e., output and input are of same time)?

Answer 1

[christttttttophe]

Yes, I am trying to predict future outputs based on the inputs and past outputs (this is Narxnet). So let's say I have 10 timesteps of data and therefore a 1x10 cell array for both the inputs and targets that I want to train.

For the Inputs (X) each individual cell (single timestep) has 20 colunms of sample data (each at 7x1) consisting of:

1. Temperature

2. Relative humidity

3. Carbon monoxide

4. Sulfer dioxide

5. Nitrogen dioxide

6. Hydrocarbons

7. Ozone

For the Targets (T) we have again a 1x10 cell array for the 10 timesteps. Except this time each individual cell array has 20 columns of (3x1) output data.

1. Total mortality

2. Respiratory mortality

3. Cardiovascular mortality

So basically we 20 sets of the same type of data at a single time. This could be Bob, Joe, and up to 20 friends measuring that data at a single time for example.

If we have multiple samples at a single time, is the neural network being trained using the all of the multiple sample data at a single time to train the network or is something else happening? Is the training metric of RMS error all of the sample data or is it simply picking the 1st column of data in each cell and ignoring the rest and not using it in the training.

Hope this makes sense.