知识卡片
离线策略价值不必需要跨世界耦合
评估目标策略的期望回报通常是部署该策略的干预目标,在顺序可交换、一致性和支持等条件下可由日志识别。若追问“同一已知轨迹在另一策略下会怎样”,才另外要求事实证据与跨世界耦合;不应只凭名称counterfactual认定全部离线评估都需层级3。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- · 查看原书依据
12.6.1 Counterfactual policy evaluation Counterfactual policy evaluation involves taking logged data from a policy in production and asking, “given we used this policy and got this cumulative reward, how much cumulative reward would we have gotten had we used a different policy?” See the chapter notes at https://www.altdeep.ai/p/causalaibook for references to techniques such as counterfactually guided policy search and counterfactual risk minimization. 12.6.2 Counterfactual regret
参考来源
- 位置:《Causal AI》第12章《因果决策与强化学习》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Causal AI/_epub-src/OEBPS/Text/chapter-12.xhtml)
- 结论依据:摘录中关于离线策略价值不必需要跨世界耦合的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:12.6.1 Counterfactual policy evaluation Counterfactual policy evaluation involves taking logged data from a policy in production and asking, “given we used this policy and got this cumulative reward, how much cumulative reward would we have gotten had we used a different policy?” See the chapter notes at https://www.altdeep.ai/p/causalaibook for references to techniques such as counterfactually guided policy search and counterfactual risk minimization. 12.6.2 Counterfactual regret