Trust Region on dextermayhewjdhttps://dextermayhewjd.github.io/tags/trust-region/Recent content in Trust Region on dextermayhewjdHugozh-cnWed, 07 Oct 2026 00:00:00 +0000Lecture 10 · 从策略改进界到 PPO-KL 与 TRPOhttps://dextermayhewjd.github.io/courses/cs285/lecture-10/Wed, 07 Oct 2026 00:00:00 +0000https://dextermayhewjd.github.io/courses/cs285/lecture-10/解释旧优势与旧状态分布的依据，逐步走过望远镜求和、耦合、TV、Pinsker、KL 惩罚和自然梯度。完整循环集中成框，保留 20 条公式拆解。