-
Notifications
You must be signed in to change notification settings - Fork 6k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[RLlib] APPO enhancements (new API stack) vol 02: Cleanup loss function, add GAE-lambda to vtrace, make rho-clip configurable. #48800
[RLlib] APPO enhancements (new API stack) vol 02: Cleanup loss function, add GAE-lambda to vtrace, make rho-clip configurable. #48800
Conversation
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM. Everything gets cleaned up now. This looks really good. Do we have already first training results?
|
||
# Target policy. | ||
target_action_dist = action_dist_cls_train.from_logits( | ||
module.forward_target(batch)[TARGET_ACTION_DIST_LOGITS_KEY] |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Can we do this here, if we want ot use multiple learners? I remember for SAC and CQL we needed to run a single forward pass before the loss calculation.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
You are right, but it does seem to run fine in the CI (multi-GPU). 🤷♂️
# The discount factor that is used should be gamma except for timesteps where | ||
# the episode is terminated. In that case, the discount factor should be 0. | ||
# The discount factor that is used should be `gamma * lambda_`, except for | ||
# termination timesteps, in which case the discount factor should be 0. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Why not gamma^(t-1)?
…on, add GAE-lambda to vtrace, make rho-clip configurable. (ray-project#48800) Signed-off-by: hjiang <dentinyhao@gmail.com>
APPO enhancements (new API stack) vol 02:
Why are these changes needed?
Related issue number
Checks
git commit -s
) in this PR.scripts/format.sh
to lint the changes in this PR.method in Tune, I've added it in
doc/source/tune/api/
under thecorresponding
.rst
file.