<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Online-Planning on mlbot.blog</title><link>https://mlbot.blog/tags/online-planning/</link><description>Recent content in Online-Planning on mlbot.blog</description><generator>Hugo</generator><language>en-US</language><lastBuildDate>Tue, 04 Aug 2026 16:02:29 +0530</lastBuildDate><atom:link href="https://mlbot.blog/tags/online-planning/index.xml" rel="self" type="application/rss+xml"/><item><title>The Oracle Could Choose, but the Online Planner Could Not</title><link>https://mlbot.blog/posts/oracle-vs-online-planner/</link><pubDate>Tue, 04 Aug 2026 12:05:00 +0530</pubDate><guid>https://mlbot.blog/posts/oracle-vs-online-planner/</guid><description>&lt;p&gt;The encouraging result from
&lt;a href="https://mlbot.blog/posts/when-projection-choice-matters/"&gt;controlled belief projection&lt;/a&gt; was that
an exact-grid oracle could choose projection actions much better than a locked
fixed policy. The uncomfortable result was that this advantage depended on
information the online filter did not have.&lt;/p&gt;
&lt;p&gt;We tested the obvious bridge: plan from the bounded approximate belief, generate
causal futures under the known model, and score the projection actions through
self-rollout. It is a deployable information pattern. It was also the wrong
planner for this benchmark.&lt;/p&gt;</description></item></channel></rss>