RLU (RL Unplugged)

RL Unplugged 是一套用于离线强化学习的基准测试。 RL Unplugged 的设计围绕以下考虑：为了便于使用，我们为数据集提供了统一的 API，一旦建立了通用管道，从业者就可以轻松地处理套件中的所有数据。这是论文 RL Unplugged: Benchmarks for Offline Reinforcement Learning 随附的数据集。
在这套基准测试中，我们尝试关注以下问题：
高维动作空间，例如运动人形域，我们有 56 维动作。
高维观察。
部分可观察性，观察具有以自我为中心的视野。
探索困难，使用最先进的算法和模仿为困难环境生成数据。
现实世界的挑战。

RL Unplugged is a suite of benchmarks for offline reinforcement learning. RL Unplugged is designed with the following considerations in mind: For ease of use, we provide a unified API for the datasets, allowing practitioners to easily work with all data in the suite once a general pipeline is established. This is the dataset accompanying the paper *RL Unplugged: Benchmarks for Offline Reinforcement Learning*.
In this benchmark suite, we aim to focus on the following key challenges:
High-dimensional action spaces: for example, in the humanoid locomotion domain, we have 56-dimensional actions.
High-dimensional observations.
Partial observability, where observations have egocentric visual fields.
Hard exploration scenarios, where data is generated for challenging environments using state-of-the-art algorithms and imitation learning.
Real-world challenges.

搜集汇总

背景与挑战

背景概述

RL Unplugged是一套由DeepMind和Google AI Research于2020年发布的离线强化学习基准测试数据集，旨在解决高维动作空间、高维观察和部分可观察性等现实挑战。它提供了统一的API，便于从业者处理套件中的所有数据，并作为相关论文的配套资源。