Alchemy

Alchemy是一个由DeepMind创建的3D视频游戏环境，用于元强化学习研究的基准测试。该数据集包含167424种可能的化学组合，每种组合都涉及一个潜在的因果结构，该结构在每次游戏时都会程序化地重新采样，从而允许进行结构学习、在线推理、假设测试和基于抽象领域知识的动作排序。Alchemy旨在通过其透明和可深入分析的结构丰富性，为元RL研究提供一个具有挑战性的基准。此外，Alchemy还提供了一套分析工具和样本代理轨迹，以帮助研究人员评估和改进元学习算法。

Alchemy is a 3D video game environment created by DeepMind, serving as a benchmark for meta-reinforcement learning (meta-RL) research. This dataset encompasses 167,424 possible chemical combinations, each paired with a latent causal structure that is programmatically resampled at each gameplay iteration. This design enables structural learning, online inference, hypothesis testing, and action ordering based on abstract domain knowledge. Alchemy aims to provide a challenging benchmark for meta-RL research via its transparent, deeply analyzable structural richness. Additionally, Alchemy offers a suite of analytical tools and sample agent trajectories to help researchers evaluate and refine meta-learning algorithms.