B-Pref

B-Pref是一个专为基于偏好的强化学习设计的基准数据集，由加州大学伯克利分校的研究人员创建。该数据集包含多种移动和机器人操作任务，来自DeepMind控制套件和Meta-world。B-Pref通过模拟具有广泛非理性行为的教师，提出不仅评估性能，还评估对这些潜在非理性行为的鲁棒性的度量。数据集旨在解决在复杂任务中指定奖励函数的困难，通过使用教师的偏好来学习策略，无需预定义奖励。

B-Pref is a benchmark dataset tailored for preference-based reinforcement learning, created by researchers at the University of California, Berkeley. This dataset includes a variety of locomotion and robotic manipulation tasks sourced from the DeepMind Control Suite and Meta-world. B-Pref proposes evaluation metrics that assess both model performance and robustness against these potentially irrational behaviors, by simulating teachers that exhibit a broad spectrum of irrational behaviors. The dataset aims to address the challenge of specifying reward functions for complex tasks, enabling policy learning via teacher preferences without predefined reward signals.