环境:
参考github:https://github.com/openai/gym
v0和v4的env差别在于有无0.25的repeat_action_probability,前者有,后者没有。
有无ram的差别在于input,是2d image(pixel) 还是1d array。更详细的解释可参考:https://stackoverflow.com/questions/45207569/how-to-interpret-the-observations-of-ram-environments-in-openai-gym
算法:
ppo
论文:Schulman J, Wolski F, Dhariwal P, et al. Proximal policy optimization algorithms[J]. arXiv preprint arXiv:1707.06347, 2017.(https://arxiv.org/pdf/1707.06347.pdf)
参考github:https://github.com/hill-a/stable-baselines
已尝试ppo2
muzero
论文:Schrittwieser J, Antonoglou I, Hubert T, et al. Mastering atari, go, chess and shogi by planning with a learned model[J]. arXiv preprint arXiv:1911.08265, 2019.(https://arxiv.org/pdf/1911.08265.pdf)
参考github:https://github.com/werner-duvaud/muzero-general
该开源项目使用了Ray,可方便的实现多机集群实验。👍
go-explore
原论文:Ecoffet A, Huizinga J, Lehman J, et al. Go-explore: a new approach for hard-exploration problems[J]. arXiv preprint arXiv:1901.10995, 2019.(https://arxiv.org/pdf/1901.10995.pdf)
最新论文:Ecoffet A, Huizinga J, Lehman J, et al. First return then explore[J]. arXiv preprint arXiv:2004.12919, 2020.