Closed carlosgmartin closed 11 months ago
Thanks for the suggestions. It sounds good. We do not need to preserve the backward compatibility with the None defaults. The stochastic_muzero_policy is still experimental.
Will you send a pull request or should I do that?
Thanks for the improvements.
It is possible to automatically determine the
num_actions
andnum_chance_outcomes
parameters tostochastic_muzero_policy
from theroot
,decision_recurrent_fn
, andchance_recurrent_fn
parameters, viaChanceRecurrentFnOutput.action_logits.shape
andDecisionRecurrentFnOutput.chance_logits.shape
. That would reduce the number of parameters that users need to pass in by two.I suggest allowing
num_actions
andnum_chance_outcomes
to beNone
, making themNone
by default, and determining their values automatically if they'reNone
. This preserves backward compatibility. Thoughts?