[Question] Is there a description on the testing methodology and the meaning of the scores in the performance tables ? about corl HOT 4 CLOSED

Many thanks !

from corl.

Hi,

Did you check https://github.com/corl-team/CORL/tree/main/results ?

from corl.

Thanks @vkurenkov for the link, however I guess I will need a little to start,

The results are the average rewards obtained?

For instance, let's take the Table of Gym-Mujoco:

https://github.com/corl-team/CORL#gym-mujoco

Task-Name | BC | 10% BC | TD3+BC | AWAC | CQL | IQL | ReBRAC | SAC-N | EDAC | DT

hopper-medium-v2 | 53.51 ± 1.76 | 55.48 ± 7.30 | 60.37 ± 3.49 | 63.02 ± 4.56 | 59.08 ± 3.77 | 67.53 ± 3.78 | 102.29 ± 0.17 | 40.82 ± 9.91 | 101.70 ± 0.28 | 65.10 ± 1.61

This numbers seems to me very low for meaning Hopper rewards.

Could you please give me some aid ?

Lots of thanks!

from corl.

Yes, sure. The results are averaged across seeds (best = best evaluation point within runs; final = the last ones within runs).

Note that these are normalized scores not the actual reward provided by the environment.

normalized_score = 100* (score - REF_MIN_SCORE) / (REF_MAX_SCORE - REF_MIN_SCORE )

For more details check section on normalized scores here https://github.com/Farama-Foundation/D4RL

from corl.