On 16-11-17 18:24, Stephan K wrote:
> 2017-11-16 17:37 UTC+01:00, Gian-Carlo Pascutto <g...@sjeng.org>:
>> Third, evaluating with a different rotation effectively forms an
>> ensemble that improves the estimate.
> 
> Could you expand on that? I understand rotating the board has an
> impact for a neural network, but how does that change anything for a
> tree search? Or is it because the monte carlo tree search relies on
> the policy network?

It was a response to the statement "There are no reevaluations that
would improve your estimate."

Consider a quiet position where the tree search wouldn't reveal any
tactics. Normally, searching deeper won't give an immediate benefit. But
because of the rotations, the value network's score is improved from a
single estimate to an ensemble.

In chess/alpha-beta terms, the quiescence search resolves the tactics
(if any), so running it again with part of the tactics resolved would
produce the same score. But with value nets, this is not entirely true.

> Could it be possible to train a value net using only the results of
> already finished games, rather than monte carlo rollouts?

Isn't this how it works already?

> My (extremely vague and possibly fallacious) understanding of the
> situation was that monte carlo tree search was less effective for
> chess because of the more sudden changes there might be when
> evaluating chess positions. For instance, a player with an apparently
> lesser position might actually be a few moves away from a checkmate
> (or just from a big gain), which might be missed by the monte carlo
> tree search because it depends on one particular branch of the tree.

Life and death and capture races behave the same. The inability of MCTS
to switch to a new PV instantly isn't necessarily very different from
the requirement in chess that all moves are searched to an equal
(nominal!) depth. In practical alpha-beta implementations, failing high
on a new best move requires a re-search as well.

-- 
GCP
_______________________________________________
Computer-go mailing list
Computer-go@computer-go.org
http://computer-go.org/mailman/listinfo/computer-go

Reply via email to