神经网络 2023 年 Neural MMO 竞赛多任务强化学习结果
机器学习
2025-08-19 v1
摘要
我们介绍了神经网络 2023 年 Neural MMO 竞赛的结果，该竞赛吸引了超过 200 名参赛者。参赛者训练的目标条件策略能够泛化到训练期间从未看到的任务、地图和对手。最佳解决方案在单个 4090 GPU 上训练 8 小时内的得分比基线高出 4 倍。我们在 MIT 许可证下开源了与 Neural MMO 及竞赛相关的所有内容，包括我们的基线以及最佳提交方案的策略权重和训练代码。
关键词
引用
@article{arxiv.2508.12524,
title = {Results of the NeurIPS 2023 Neural MMO Competition on Multi-task Reinforcement Learning},
author = {Joseph Suárez and Kyoung Whan Choe and David Bloomin and Jianming Gao and Yunkun Li and Yao Feng and Saidinesh Pola and Kun Zhang and Yonghui Zhu and Nikhil Pinnaparaju and Hao Xiang Li and Nishaanth Kanna and Daniel Scott and Ryan Sullivan and Rose S. Shuman and Lucas de Alcântara and Herbie Bradley and Kirsty You and Bo Wu and Yuhao Jiang and Qimai Li and Jiaxin Chen and Louis Castricato and Xiaolong Zhu and Phillip Isola},
journal= {arXiv preprint arXiv:2508.12524},
year = {2025}
}