research
Forward KL Kept More High-Value Actions in Play
FERPO improved policies without differentiating the critic with respect to actions.
Published Updated Story ID: mp-2026-10-02-015
Summary
FERPO improved policies without differentiating the critic with respect to actions.
FERPO derives an entropy- and KL-regularized target action distribution from critic values, then fits the actor with forward KL and self-normalized importance sampling. MuJoCo Playground and ManiSkill tests showed competitive performance, sample-efficiency gains and faster actor updates than REPPO.
Why it matters
FERPO improved policies without differentiating the critic with respect to actions.
Limits and context
No additional limitation was separately recorded.
Key claims
FERPO improved policies without differentiating the critic with respect to actions.
Evidence: source-2026-10-02-017
Sources
- arXiv preprint 2610.02198arXiv · primary research
Corrections
No corrections have been recorded for this story.