How Large Is the Action Space for the Policy?

If you ignore the language, the action space is still huge. The branching factor is something like 10 to 20. And that's also ignoring the language. So we sample from the policy net a set of actions for each player that are highly likely. We iterate 256 times and we end up with hopefully and in practice a better prediction of what everybody's going to do than just with the invitation learned policy.