search

FangchenLiu / MaskDP_public

Code for "Masked Autoencoding for Scalable and Generalizable Decision Making". NeurIPS 2022
MIT License
43 stars 4 forks source link

Question about generating expert dataset #2

Open tianyunzhe opened 4 months ago

tianyunzhe commented 4 months ago

Thank you for sharing your code.

Could you please explain how to generate the dataset for _maskdp_evalexpert?

Additionally, could you provide the code and commands for this part? Thank you.

FangchenLiu commented 3 months ago

We train a TD3 policy for 0.5M - 1M interaction steps (depends on task), store the checkpoint, and unroll the stored checkpoint with a small gaussian noise to collect such data, so we call it (near-)expert data. Train and eval expert data are generated in the same way, and we split them for evaluation. The conclusion and relative performance still hold for our method if you use other data (e.g. sup, semi).

It's should be easy to implement, but I'll also add the code support in the data collection branch.