Question about alphago zero evaluator about leela-zero

[gcp]

There are several sources of randomness, they are all mentioned in the paper.

They add noise to the policy network priors:
https://github.com/gcp/leela-zero/blob/f321b568ec23d10d024661c6227ddc393a5d8ac8/src/UCTNode.cpp#L181

All network evaluations are done with a random symmetry. This is RANDOM_ROTATION in Leela's source.

In practice one expects some randomness from multiprocessing with virtual loss as well.

[Narsil]

Isn't the noise only used in self play and not in evaluation ?

I will add the symmetries and rotations, I forgot about those.

[gcp]

They only talk setting the softmax temperature (really: disabling the move selection randomization), not about disabling the Dirichlet noise. I suspect you can get away with not doing that anyway, the symmetries add a lot of variation to what it will play. On the other hand leaving it enabled probably doesn't really make the program weaker, at worst it will spend just a few playouts on bad moves.