https://mlbot.blog/tags/distribution-shift/Recent content in Distribution-Shift on mlbot.blogHugoen-USTue, 04 Aug 2026 16:02:29 +0530https://mlbot.blog/posts/useful-offline-harmful-online/Tue, 04 Aug 2026 12:10:00 +0530https://mlbot.blog/posts/useful-offline-harmful-online/<p>After <a href="https://mlbot.blog/posts/oracle-vs-online-planner/">causal self-rollout failed</a>, the next attempt removed online tree search. A permutation-invariant neural controller would read the bounded Gaussian-mixture belief and directly predict the full-information oracle’s six action values.</p> <p>Offline, the idea worked well enough to be interesting. On fresh repeated control, the frozen policy was harmful. One capped dataset-aggregation round improved the states the policy actually visited and produced a promising known-schedule specialist—but the same corrected controller failed under random hazards.</p>