ruizhaogit / energybasedprioritization Goto Github PK

Hello,

Thank you for your work!
But I didn't get the same result for FetchPickAndPlace as in your paper.
I installed all necessary packages and just run this command on my laptop:
python baselines/her/experiment/train.py --env_name FetchPickAndPlace-v1 --prioritization energy --clip_energy 0.5 --n_epochs 50 --num_cpu 1
The model was trained, but the max success rate I could get is 0.65 for FetchPickAndPlace-v1 with 100 epochs. What batch size did you use?

Another question is: have you tested your algorithm in the FetchSlide environment?

Thank you

Hi, atfer read paper 'Energy-Based Hindsight Experience Prioritization' , i run the algorithm in FetchPickAndPlace-v0 with energy prioritization on 5 cpu and 11 cpus, and found the test success rate is quite diffierent. Compare with 11 cpus, the performentce of 5 cpus is worse, and the curve is unstable.
I wonder why the number of cpu could affect the performence greatly, and could i get the similar performence with the paper without so many cpus.