It seems that every Go study notes the importance of "distance to the last
move" as a predictor. Several heuristic measures of "distance" have been
proposed in the literature. I tested three:
1) The Manhattan distance dx+dy.
2) dx + dy + max(dx,dy), which is Manhattan plus a term that measures
the difference in row/column.
3) Mogo's function: "a topological measure that is the same for all
points in a string." Fuego also implements this function.
It seems that the threshold for nearness is "distance <= 3", which includes
link points and excludes most non-link points. Note that dx+dy+max(dx,dy)
increases faster than the other two, so you have to increase the threshold
for that function.
Pebbles uses distance as a progressive widening bias (only). In this
implementation, you give a bonus for nearness as a linear function of the
distance, with zero assigned to all points beyond the threshold distance.
I tested the three functions above, hoping to definitively identify which
Function produced the best performance. Unfortunately, the answer isn't
clear.
All functions have strengths and weaknesses.
But I had to choose, so I picked Mogo's function. It has some strengths:
1) With its best parameters, it did produce the best results. By a
little bit.
2) Second and third best parameters were still among the best
performers.
3) The function is useful in both open boards (making space) and closed
(semeai).
The disadvantage of this function is that it can bonus a lot of points at
distance 3,
which blurs the focus a bit. I decided that I can refine the bonus function
to
improve behavior in that case. So with more work, I may be able to produce
clearly better results.
What do other programs do?
_______________________________________________
computer-go mailing list
[email protected]
http://www.computer-go.org/mailman/listinfo/computer-go/
