Clipping

القصّ

آلية تحدّ من حجم تحديث السياسة بقصّ نسبة الاحتمال ضمن نطاق [1−ε, 1+ε]، مما يمنع التغيرات الكبيرة المُزعزعة للتدريب.

A mechanism that limits policy update size by clipping the probability ratio within [1−ε, 1+ε], preventing large destabilizing changes during training.

Also translated asالتحديد، القطع

First appears in this corpus in: Proximal Policy Optimization Algorithms (2017)

Appears in these papers

Direct Preference Optimization: Your Language Model Is Secretly a Reward Model2023 in the sky ✦
Learning Dexterous In-Hand Manipulation2019 in the sky ✦
Proximal Policy Optimization Algorithms2017 in the sky ✦