what's the output dimension of actor network?

[zkengz]

hi. in the paper you've run an experiment to compare DMC with actor-critic. in the case of actor-critic method, I am curious about the output dimension of actor network.

since the action space is so large, it would be impossible to output the probability of all 27472 actions. but if you set the number of legal actions as output dimension, it is still large and varies from turn to turn. and the meaning of each action can change (e.g. the first action stands for solo 2 this turn, and solo 3 the next turn).

so... how do you deal with it?

your prompt reply would be appreciated.