PKU-MARL / Multi-Agent-Transformer

339 stars 72 forks source link

State (cent obs) not used in evaluate actions #30

Open brithbutter opened 8 months ago

brithbutter commented 8 months ago

To whom it may concern, I noticed that in forward() of ma_transformer.py, the state (cent obs) is converted into zeros. However, it seems to be an essential element for critic function of the RL. And I roughly can not identify the reason or explanation about this in the paper. Could you provide an elaboration in code? Sincerely,

morning9393 commented 2 months ago

hiya, thx for your attention, the cent obs is a redundant parameter in MAT (just used for debugging, I forgot to delete it). In MAT, the encoder will combine obs from all agents, performing the role of cent obs (but including less information than cent obs, if you are chasing the ultimate performance, you could also replace the obs with cent obs for encoder I think~).