jaekyeom / ibol Goto Github PK

Did anyone debug this code and get the diversity path as shown in the paper? I don't seem to get that effect. I used a diayn code and I can barely run the skill diversity path as shown in the paper. I just want to know if you can get a straight path like that? It really helps me.!!!

Sorry for the many rudimentary questions.

My last question.
What is the difference between function get_mode_actions, get_sample_actions, and get_actions in policy_ex.py?
Also, I would appreciate it if you could tell me when to use each of them.

Thank you for your quick answers to my many questions.

Hi.
First of all,thank you for publishing the IBOL code.

I have a rudimentary question:
I am aware that the dimension of actions in the HalfCheetah of Gym is 6.
However, the dimension of the action stored in the 'data' of the IBOL appears to be 18.
I would appreciate it if you could tell me how the actions are stored in the 'data' in this code.

Hi,

Thanks for publishing the code. In the skill discovery training stage, there is a bug of missing 'num_repeats' in Line 88 of ''garagei/envs/child_policy_env.py", which should have been defined in Line 190 of "tests/main.py".

I search the repo, and find that 'num_repeats' has a default value of 1. Is that a suitable value for skill discovery? Which variable does it correspond to in the paper?

Looking forward to your reply! Thanks!

Sorry for the repeated questions.
I also apologize if this is mentioned in the paper.

How can I get a post-study video like the one below?
As for HalfCheetah, I have only visualized actions in 6 dimensions. Sorry about that.