Yes,
And to push somebody I
propose how to quickly evaluate a move.
From some position play 1000 semi-random games
with this move first. Record all counts.
Set window [A,B] big enough to encompass all counts.
Go throught all counts and
Each time pick a random number R from [A,B]
If Count is less than R assign win, else assign loss
Calculate winrate WR = allwins/1000
From winrate WR and window [A,B] calculate value of a move V
V = A + WR * (B-A)
Now you go again and collect statistical information from Counts and V
like standard deviation and such to get confidence and dispersion.
If there is only one outcome then V = C, Standard deviation is 0,
and confidence high. Else apply statistics ...
Better play-out engine more confidence in V.
That is close to what MCTS does but fails to normalize.
Now you collect moves to game tree in do alpha-beta and pruning.
Leon.
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
