rail-berkeley/softlearning

target_entropy discrete_space

Maggern3 opened this issue · 2 comments

Maggern3 commented

softlearning/softlearning/algorithms/sac.py

Line 42 in 46f1443

"TODO(hartikainen): implement for discrete spaces.")

Are you planning to implement this?
What would be a good value for a MultiDiscrete([3 3 2 3]) action space?
Depending on how I calculate I get -4, -11 or -54 but I'm a bit clueless to what would be a good value.

hartikainen commented

Hey @Maggern3, thanks for the question! Unfortunately, I don't have plans to implement the discrete version at least in the near future, mainly because I don't really work on discrete tasks.

Hello-Ki commented

softlearning/softlearning/algorithms/sac.py

Line 42 in 46f1443

"TODO(hartikainen): implement for discrete spaces.")

Are you planning to implement this?
What would be a good value for a MultiDiscrete([3 3 2 3]) action space?
Depending on how I calculate I get -4, -11 or -54 but I'm a bit clueless to what would be a good value.

Hello! Have you found a solution to implement the discrete action?