how to predict multiple steps ahead

[wilbertmatthew]

I am using PyRCN version PyRCN-0.0.17 and the mackeyglass example. I would like to know how to modify the example to predict multiple steps ahead instead of 1 step ahead as it currently does.

Thanks

Wilbert Jackson

[renierts]

Hi Wilbert,

Thank you for this question. Currently, this is not supported natively. However, I am working on a better implementation based on PyTorch, which makes it easier to do tasks like this.

For now, the easiest way would be to predict one step, append the output to the input and then predict the next step with this modified input. This can be done in a simple loop. Does this help you for now?

[wilbertmatthew]

Thanks Peter for your quick response.
Is there any code available for your new design? I tried your suggestion for multi step prediction as follows:

stepsahead = 10
for i in range(stepsahead):
y_test_pred = esn.predict(X_test)

append the last prediction value to the X_test array

X_test = np.concatenate((X_test., y_test_pred[-1:])
y_test_pred = esn.predict(X_test)

Is this the right idea? The multiple calls to the incremental_regression function makes for an very slow prediction process.

[renierts]

Hello,
If you are still interested in using PyRCN, I recommend you to check the latest update (0.1). I added a PyTorch backend now, and iterative training. I think that this helps to speed up things, because you don't have to concatenate numpy arrays anymore.
Your idea was correct, but I for myself didn't find time to make the implementation more efficient.
Since it is a long time ago (my fault, I was too busy with new projects), I understand if you don't need this anymore.