Quoting Jason House <[EMAIL PROTECTED]>:
Does anyone have any data on just how optimistic or pessimistic the results
would be? I'd like to use some heuristics that inherit winning percentages
from a parent node to bias the expected winning percentage of the children
nodes... and maybe pruning away portions of a search in a fixed depth monte
carlo search with iterative deepening.
This is confusing to me. Are you asking how UCT behaves in order to implement
something which is not UCT?
Anyway, UCT scores has the property that the score of a node changes
very slowly
when it is searched deep. But does this mean that sibling scores can be
compared
to each other? I would hesitate here because the scores change as a
function of
the search. In difficult positions the scores for all good moves are often
very similar. The scores often move up or down slowly together as
function of if
the position is good or not.
I attach an sgf file where i added the principal variation of Valkyria
after the
black opening move of the center point using 5 minutes of thinking time. For
each move I give the winning percentage and the number of playouts that passed
this node. For the second move of black I also give the winning percentage of
all siblings. The imortant thing here is that many of those move are only
searched 1000 times whereas the few best move were searched 1000000 times, but
there is still not much of a difference in the actual scores.
Best
Magnus
--
Magnus Persson
Berlin, Germany
(;GM[1]FF[4]SZ[9]AP[SmartGo:1.4]
;B[ee];W[cc]
C[Main variation
54.3% 705920
Winrate is shown for all black moves searched more than 1000 times. Note how
close all values are. If all moves were searched as much as the best move these
numbers would certainly be much lower in many cases.
]
LB[bb:43][cb:42][db:43][eb:44][fc:46][gc:45][dd:44][ed:45][gd:45][ge:45][ff:43][gf:45][gg:45]
;B[ec]
C[46% 340813];W[be]
C[54.6% 138298];B[fg]
C[46.3% 75060];W[db]
C[54.7% 9330];B[eb]
C[47.6% 3923];W[eh]
C[54.9% 1236];B[dc]
C[47.0% 279])
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
