Mapo Mixed Advantage Policy Optimization

Dynamically reweight advantage functions based on trajectory certainty to improve policy optimization in foundation models. Addresses advantage reversion and mirror problems by mixing standardized and mean-normalized advantage formulations. Enables more stable gradient signals across high- and low-certainty samples.

adu2021 Updated 7 repo stars

File contents

adu2021/skillxiv commit 744b40f43d

Frequently asked questions

npx get-skill add adu2021/mapo-mixed-advantage-policy-optimization