I was literally just struggling to download it and I found your video! Amazing man! You have a new subscriber! Keep up the great videos!

Would love to see more on custom PettingZoo envs and how to train multiple agents with e.g SB3

Great tutorial. I know that it isn't relevant, but if I were to create my own environment for marl, where do you think it is best to code? I have done in python, but my environment seems to be limited by my processing power so I can't simulate environment fast enough to speed up training (as you do with your gym environments without rendering). Do you think that I need to rebuild it in c-like language?

@johnnycode

Ай бұрын

I'm guessing your code is not utilizing 100% of your computer's CPU and GPU. SB3 has something called "vectorized environment", where you can train several instance of the env in parallel, so you might want to check out their code. You can also pay for Google Colab or high-end virtual cloud servers, if you need more processing power.

@FaryhSan

Ай бұрын

@@johnnycode Yes, I also think that my code isn't optimized to its max as I am using structures that are known as why python is so slow. I was just having a dilemma of whether to improve on that (with numba/cython) or recreate full environment on faster language, like c++ (then using smth like pybind11 as what box2d-py does). Also thanks for pointing out on vectorized env, will look into it, since ppo is one of available algos in there.

Can you do a tutorial for Nvidia Gym or Nvidia Sim? The videos on KZread are not that good, and I think you would be good at it.

@johnnycode

2 ай бұрын

I’ll check it out.

@Build_the_Future

2 ай бұрын

@@johnnycodeThanks, I look forward to it.