search

takuseno / d3rlpy-benchmarks

Benchmark data for d3rlpy
MIT License
20 stars 5 forks source link

Add learning d4rl mujoco dataset for *-random-v0/v2 #3

Closed im-Kitsch closed 11 months ago

im-Kitsch commented 2 years ago

Hi,

thanks for the excellent repo, I would like to ask could you add the test of *-expert-v2 data?

I just tried to learning halfcheetah-expert-v0 dataset on AWAC, but unfortunately it learns nearly nothing, if you could add some tests on expert dataset, it would be quite helpful, thanks!

takuseno commented 2 years ago

Thank you for your request. I'm focusing on filling -v0 datasets for journal submission. Once it's published, more datasets including -expert-v0 will be added.

Regarding AWAC training, I've observed that AWAC is not really robust for offline training. I'd recommend IQL unless you really need AWAC.