Unverified50% confidenceFactExact time
离线RL的核心难题是如何避免对数据集未覆盖的状态-动作对产生过于乐观的价值估计（分布外外推问题）
1
Sources
50%
Confidence
Long-term
Relevance
7/14/2026
First Seen
Sources
Vizuara机器人学习课程笔记获取指南与免费资源推荐
redditr/deeplearning7/12/2026
Related Claims
Unverified离线RL的核心挑战是分布偏移，CQL通过惩罚项压低未见过动作的Q值，IQL则完全避免对数据集外动作的直接查询75% similarUnverifiedA model that performs well on offline metrics may perform poorly online due to differences in data distribution, a phenomenon known as the offline-online gap.74% similarUnverified错误的归纳偏置会让模型花费容量拟合数据中不存在的虚假模式，导致分布外数据上严重失效70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/506763API
curl https://kongchang.com/api/v1/knowledge/claims/506763MCP
get_claim(id=506763)