ColorGrid

该数据集名为ColorGrid，是一种新颖的多代理强化学习环境，旨在评估代理在人与代理协作背景下的学习能力。它具有可定制的非定常性、不对称性和奖励结构特点。该环境允许实验者探索不同密度的区块和目标切换概率，并包含模型检查点和轨迹可视化功能。它支持任意数量的代理，但主要关注1对1的场景。其任务是评估在目标推断和辅助场景中的多代理强化学习算法。

This dataset, named ColorGrid, is a novel multi-agent reinforcement learning environment designed to assess agents' learning capabilities within the context of human-agent collaboration. It features customizable non-stationarity, asymmetry, and reward structures. The environment enables experimenters to investigate varying block densities and target switching probabilities, and incorporates model checkpointing and trajectory visualization functionalities. It supports an arbitrary number of agents, with the primary focus on 1:1 collaborative scenarios. Its core task is to evaluate multi-agent reinforcement learning algorithms in target inference and assistance scenarios.