A simple question about MAT

[KID0031]

Hello, I'm a beginner with MAT. In the paper, it is mentioned that the order of agents can be permuted during the training process. Additionally, the observations are concatenated with agents' IDs to preserve 'positional' information. However, the action embedding of different agents utilizes the same action encoder without incorporating positional information. How does the model handle heterogeneity in actions without distinguishing between them? Thanks :)

[morning9393]

hiya, thx for your attention~ the permutation mainly happens for the position of the input sequence, but for each obs, the agent id would not be changed, thus the policy could also distinguish the heterogeneity between agents~~