Off-the-Grid Multi-Agent Reinforcement Learning (OG-MARL)

OG-MARL 是一个为离线多代理强化学习研究设计的增长数据集库，由InstaDeep和开普敦大学合作创建。该数据集提供了多种环境动态、异构代理、非平稳性、多代理、部分可观测性、次优性、稀疏奖励和示范协调等特性的设置，以模拟真实世界系统。数据集包括不同类型的数据集（如Good, Medium, Poor, Replay），并详细描述了每个数据集的经验组成。OG-MARL旨在作为社区可靠的数据集来源，推动该领域的进步和为新研究者提供入门点。数据集适用于开发有效的分散控制器，解决如交通管理、能源管理等实际问题。

OG-MARL is a growing dataset library designed for offline multi-agent reinforcement learning (MARL) research, co-developed by InstaDeep and the University of Cape Town. This dataset offers configurations featuring a wide range of characteristics, including diverse environmental dynamics, heterogeneous agents, non-stationarity, multi-agent scenarios, partial observability, suboptimality, sparse rewards, and demonstrated coordination, to emulate real-world systems. The library encompasses multiple dataset variants such as Good, Medium, Poor, and Replay, with detailed documentation of the experience composition for each dataset. OG-MARL aims to act as a reliable dataset resource for the global research community, promoting advancements in the field and providing an accessible entry point for novice researchers. This dataset is suitable for developing effective decentralized controllers to address real-world practical problems including traffic management and energy management.