Closed famura closed 3 months ago
@NaturalGradient what do you think?
We forgot to close this, but having discussed it, @engintoklu and I had decided that sigma should remain a float. As noted by @NaturalGradient above, in the context of CMA-ES, sigma is not a variable that can be specified for each dimension (which can be true in the context of other distribution-based algorithms). Unfortunately the similar naming across algorithms may cause confusion.
Nevertheless, we appreciate you bringing this to our attention @famura!
See issue #101
Do not cast sigma to a builtin float but keep it as a torch.Tensor. Now the return type also matches with the superclass.