arXiv · Machine Learning Theory· Du Cheng, Junyu Lu, Jingfan Zhang, Yuzhuo Zhang, Haoxuan Jia, Shichao Weng·· 3 小时前AI 评分37
Cross-Behavior Policy Optimization：用共享回归目标统一离线演示与在策/离策 rollout 学习
Cross-Behavior Policy Optimization: Unified Learning from Offline Demonstrations and On- and Off-Policy Rollouts via a Shared Regression Target
AI 导读
论文提出 Cross-Behavior Policy Optimization（CBPO），用固定参考目标的单一回归损失统一离线演示与在策/离策 rollout，无需存储行为策略概率、重要性加权或学习 critic。
来源：arXiv · Machine Learning Theory · arxiv.org