search

takuseno / d3rlpy

An offline deep reinforcement learning library
https://takuseno.github.io/d3rlpy
MIT License
1.3k stars 232 forks source link

[REQUEST] MuZero Unplugged #165

Closed tbskrpmnns closed 2 years ago

tbskrpmnns commented 2 years ago

Hey,

I was really impressed by DeepMind's latest progress in their offline RL version of the MuZero algorithm arXiv:2104.06294. Since it provides sota results for several offline RL benchmarks and is in addition data efficient because of the model-based RL approach, I was wondering whether you have any plans in implementing MuZero Unplugged?

0xJchen commented 2 years ago

Saw your comments on EfficientZero. I suppose you can just fill the replay buffer manually and trained the agent based on that.

tbskrpmnns commented 2 years ago

Thanks again for your help! If I understand it correctly, I can use then MuZero / EfficientZero or any other algorithm suitable for offline RL using a replay buffer that I can simply fill manually for the offline RL use-case?

takuseno commented 2 years ago

Not limited to MuZero, you can manually fill up the replay buffer. For the advanced support of model-based RL algorithms, it's still on the TODO list in this project.

tbskrpmnns commented 2 years ago

awesome, thank you very much!