search

takuseno / d3rlpy-benchmarks

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

I am wondering if there will release the result on the d4rl-v2 dataset in the future? And how to compute the normalized return? #2

Closed houmaolin closed 2 years ago

takuseno commented 2 years ago

Currently, I'm focusing on producing benchmark results for -v0 with 10 random seeds each to make them available for journal submission. Once we get it accepted, we would work on -v2 benchmarking.

houmaolin commented 2 years ago

thanks for applying. image In d4rl, I found the way for computing the normalized score. but how could I get the expert score and random score used in the picture? Is there a certain number for every task, or I should run SAC and random policy in every task to get the certain score?

takuseno commented 2 years ago

Here it is. https://github.com/rail-berkeley/d4rl/blob/master/d4rl/gym_mujoco/__init__.py#L23

houmaolin commented 2 years ago

Here it is. https://github.com/rail-berkeley/d4rl/blob/master/d4rl/gym_mujoco/__init__.py#L23

Thanks a lot!

houmaolin commented 2 years ago

Here it is. https://github.com/rail-berkeley/d4rl/blob/master/d4rl/gym_mujoco/__init__.py#L23

Are those scores same between the v2 or v0?