The learning rate of the inverse KKT method in the code is inconsistent with that in the paper.

[beijiguang94]

hi @wanxinjin. I noticed that in the paper the learning rate for PDP, inverse KKT, and neural policy cloning methods in imitation learning was set to $\eta=10^{-4}$. But in scripts like "cartpole_inverseKKT.py", the parameter lr equals 1e-7. Why so?

[wanxinjin]

Hi @beijiguang94, thanks for your interest.
I highly likely have changed the codes afterward. Please use a learning rate that is stable.

[beijiguang94]

thanks. When useing different learning rates, I suppose it would be better to compare the imitation loss of these methods by changing the X label from 'iterations' to 'consumed time'.