Benchmarks for Mechanistic Offline Reinforcement Learning (B4MRL)
收藏资源简介:
Benchmarks for Mechanistic Offline Reinforcement Learning (B4MRL)是由以色列理工学院和Google Research共同创建的数据集,旨在评估强化学习中结合离线数据和模拟器的混合方法。该数据集包含16个基准,涵盖了模拟器建模误差、部分可观测性、状态和动作差异以及混杂偏差等挑战。数据集基于MuJoCo机器人环境和Highway环境创建,旨在帮助研究者理解和解决使用离线数据和模拟器时的关键问题,推动强化学习领域的发展。
Benchmarks for Mechanistic Offline Reinforcement Learning (B4MRL) is a dataset jointly developed by the Technion – Israel Institute of Technology and Google Research, which is designed to evaluate hybrid reinforcement learning approaches that combine offline data and simulators. This dataset comprises 16 benchmarks covering key challenges such as simulator modeling errors, partial observability, discrepancies between state and action spaces, and confounding bias. Constructed based on MuJoCo robotic environments and Highway environments, the dataset aims to assist researchers in understanding and addressing critical issues arising from the use of offline data and simulators, thereby advancing the progress of the reinforcement learning domain.

- 1Benchmarks for Reinforcement Learning with Biased Offline Data and Imperfect Simulators以色列理工学院 · 2024年



