zhengyinan-air / omiga Goto Github PK

I noticed you are using non-default SMAC arguments when initialising the environment. Can you explain why? And maybe provide a list of arguments you changed? Did you also change the SMAC code since you included the SMAC environment in your repository?

Thanks for the assistance.

Hi!
In your paper, you compared OMIGA with BCQ-MA, CQL-MA, ICQ and OMAR. I have found the official implementations for ICQ-MA and OMAR, can you share your version for BCQ-MA and CQL-MA or just share the implementations link? This will help me a lot!
Thanks for your work!

Inspecting the datasets for 2x4 Ant I notice that the observation dimension for each agent is 113. But the observation dim when agent_obsk=1 is supposed to be 52. Does that mean in your experiments agents got to see the whole state?

Many thanks for the contribution. I look forward to being able to use your datasets.

In the get_episode() function, the rewards have been turned into reward-to-gos, which is not describe in the paper.

for agent_trajectory in episode:
    rtgs = 0
    for i in reversed(range(len(agent_trajectory))):
        rtgs += agent_trajectory[i][3][0]
        agent_trajectory[i][3][0] = rtgs

return episode

Can you explain why should turn rewards to reward-to-go? Is this transformation applied for other baselines too?
Sorry to bother you.

Hi there, thanks for sharing your code for MAMuJoCo experiments. I was wondering if you plan to release the code to reproduce your experiments on SMAC?

Hi, I am trying to download the HalfCheetah datasets from the link you provided. But a few seconds after starting, it always fails. Please advise how to get access to these datasets.

I did successfully manage to download the Ant and Hopper expert datasets.