The idea this time is to make use of your opponent's moves (as the proverb says). I tried a few different weights for the opponent move and 0.1 was the best of
the ones I tried. The win rates for the version with opponent moves versus the version without opponent moves are as follows (both versions use the weighted
AMAF recently discussed):
At 8 playouts, modified version wins 66.8% after 500 games.
At 64 playouts, modified version wins 56.6% after 500 games.
The advantage continues to decrease with more playouts. So it seems to perform better for a small number of playouts but doesn't scale well. I think this is
ok for a tree search program because you only run a few simulations before the node gets expanded. But I have not tried it in a tree search program yet.
Here is the code, so there is no ambiguity.
for (int mv = 0; mv < NNN; mv++)
{
credit[mv] = true; // initial assumption is that credit is awarded
for any move
}
double weight = 1.0;
double weightDelta = 2.0 / (ctm - savctm + 1);
for (int i = savctm; i < ctm; i += 2)
{
int mv = mvs[i] & MASK;
if (credit[mv])
{
wins[mv] += weight * sc;
hits[mv] += weight;
credit[mv] = false; // do not award credit for this move again
}
int opponentMove = mvs[i + 1] & MASK;
if (credit[opponentMove])
{
double opponentWeight = 0.1 * weight;
wins[opponentMove] -= opponentWeight * sc;
hits[opponentMove] += opponentWeight;
credit[opponentMove] = false; // do not award credit for this
move again
}
weight -= weightDelta;
}
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
