Some questions about ego attention DQN about rl-agents HOT 2 CLOSED

Hi,
Using several heads can allow them to specialize and attend to different parts of the input data, and it has shown to be beneficial in some use cases. However, the number of heads can be considered as a hyperparameter, and I don't know of any good tuning rule except for manual/grid search... On the intersection task, I think that I tried 2, 4 and 8 heads without seeing any significant effect (but 2 was better than 1). Also, note that with the current implementation, the dimensions of the attention features are shared between the heads, so (all things being equal) having more heads also means having smaller heads.

tldr: my advice is: try to change it, see if it helps ;)

from rl-agents.

I will try it, thanks a lot!!

from rl-agents.