Question about the dynamics network about efficientzero HOT 2 OPEN

Thank you for your comments.

Here we scatter the actions to planes to ensure the shape is the same as the shape of the feature plane (eg: feature is B x 64 x 6 x 6 and action is B x 1 x 6 x 6). And we choose a / action_space to scale the actions.

I agree that it is not a VERY good or natural way to shape the actions because the distance between (a=1, a=2) is different from that between (a=1, a=3). A Good way is to broadcast the action into shape B x Action_space x 6 x 6 with one-hot labels. But it's a large tensor because of the spatial feature, especially when action space is large. Moreover, when action space=18 or even 81, the current implementation can still work well. So we just keep the current implementation.

By the way, it is interesting to find a better method for the action representation under spatial features (not flatten features). Hope this can help you:)

from efficientzero.

Thanks for the reply.

I guess my thought was it'd be better to broadcast onto the height dimension - so you could one-hot encode the actions and end up with a B x 64 x 6 x 10 tensor (or 64 x 6 x 24 if you have 18 actions). But I can see for a large number of actions this would become a large tensor.

from efficientzero.