Bound std of Gaussian policy via beta-sigmoid ? about softlearning HOT 6 CLOSED

The reason for clipping the std is mainly to prevent nans during initial transients, and the bounds are not typically active afterwards. You could try out different ways to implement clipping, but my intuition is that there isn't going to be much difference in performance and stability.

from softlearning.

Thanks @haarnoja for your reply. I have a tiny further question, it seems the 'pre-tanh' Gaussian sampled actions is implemented by Affine transformation of a unit Gaussian, is it because of better numerical stability or it can be identical implementation to directly using mean and diagonal std for MultivariateNormalDiag ?

from softlearning.

it seems the 'pre-tanh' Gaussian sampled actions is implemented by Affine transformation of a unit Gaussian, is it because of better numerical stability or it can be identical implementation to directly using mean and diagonal std for MultivariateNormalDiag ?

Good question! You're right, this would be identical with directly using mean and std for the MultivariateNormalDiag. The reason I implemented this way was that at the time of writing these I wasn't sure what would be the best way to fit different distributions into the keras models. Also, I wanted to fit the RealNVPPolicy (which unfortunately doesn't yet exist in this code base) and the GaussianPolicy into a similar pattern to make it more readable and usable.

I don't actually think that the policies we have right now are implemented in the best possible way. Better way would be to utilize the new tensorflow probability layers to do these transformations. I'm planning to refactor them once we upgrade to tf2.

from softlearning.

@hartikainen Thank you so much for your detailed explanation. I have a quick further question, in PG methods, we often stop gradients of a normally sampled action (say with reparametrization trick) when computing log-probability, otherwise there'd be zero gradient for the mean head. However, in SAC (tanh-transformed Gaussian), it seems to me that we need the gradient flow during log-abs-det part, but should we also allow gradient flow to the base distribution (i.e. original Gaussian) ?

I've tried to stop gradient of tanh-transformed sampled action, it breaks the algorithm where the gradient explodes (grad norm can grow to e+20).

from softlearning.

Thanks @hartikainen @haarnoja for your kind replies. It is much clearer to me now.

By the way, the numerically stable formula for log-abs-det proposed in this repo is really beautiful, and compared with clamping of original formula, it also leads to a bit better final performance.

from softlearning.

Thanks! This was actually originally proposed by @gjtucker and we have indeed found it to be really great.

from softlearning.