<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>PPO on dextermayhewjd</title><link>https://dextermayhewjd.github.io/tags/ppo/</link><description>Recent content in PPO on dextermayhewjd</description><generator>Hugo</generator><language>zh-cn</language><lastBuildDate>Wed, 07 Oct 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://dextermayhewjd.github.io/tags/ppo/index.xml" rel="self" type="application/rss+xml"/><item><title>Lecture 9 · 从旧样本复用到 PPO</title><link>https://dextermayhewjd.github.io/courses/cs285/lecture-09/</link><pubDate>Wed, 07 Oct 2026 00:00:00 +0000</pubDate><guid>https://dextermayhewjd.github.io/courses/cs285/lecture-09/</guid><description>逐步拆开换分布、轨迹重要性权重、因果性、单步替代目标与 clip/min，保留公式逐行拆解和四篇卡点精读。</description></item><item><title>Lecture 10 · 从策略改进界到 PPO-KL 与 TRPO</title><link>https://dextermayhewjd.github.io/courses/cs285/lecture-10/</link><pubDate>Wed, 07 Oct 2026 00:00:00 +0000</pubDate><guid>https://dextermayhewjd.github.io/courses/cs285/lecture-10/</guid><description>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。</description></item></channel></rss>