outputs from mgpr.predict_on_noisy_inputs and pilco.propagate confusion about pilco

[mcyip]

Thanks! This is a silly mistake on my part. Thanks for the clarification. Will do a PR.

[nrontsis]

Hi,

Thanks for the interest. If you spend time on writing comments, we would be very happy if you were to make a PR so as to incorporate (some of) them in the main repo. We plan on adding more documentation & examples soon, and while writing them it would be helpful to get feedback from other people.

The delta comes because the datasets X, Y you provide to PILCO should be of the form:

X[i] = [state at time i;
       control action at time i]
Y[i] = (state at time i + 1) - (state at time i).

So the underlying GPs are trained to provide delta as their output.

See here how we create X and Y on the example.