Killer moves works well in alpha-beta search I guess because it found
by searching the PV then when alternative moves are played to the best
they are cut of one at a time with the killer move. Thus the killer
move is applied to almost the same position in every attempt.
In a playout the boardstate is arbitraty every time a "killer move" is
attempted. So you may try a 5*5 pattern and collect statistics for
them. The problem is that either this pattern is either to large or
too small to capture the necessary information you need. The solution
may be a smarter way of defining patterns and learn them dynamically.
-Magnus
A response heuristic could also be used in a subtree fashion as in RAVE.
That should increase applicability.
A 5*5 vicinity response table was indeed a crude suggestion, my apology.
But the fundamental need remains. Most of the time when bots
misinterpret a position, it's not that they didn't find the solution to
tactical problems.
It's that they didn't find them often enough, both in the tree and in
the playouts.
So there needs to be some memory of success and failure - not just in
the leaf of the tree, but of critical tactical sequences everywhere.
Stefan
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
