Before I reply to SP Lee's question, I just wanted to say how pleased I was to read about the progress being made with the Cosmic patch. It seems clear to me that GNUGo's play could be frighteningly good with just a modest amount of fullboard strategy. That Cosmic adds a couple stones of strength to GNUGo does not surprise me in the least.
SP Lee asked me to run the influence mapping algorithm I presented earlier on the following position:
19: . . . . . . . . . . . . . . . . . . .
18: . . . . O X X . . . . . . . . . O . .
17: . . . . O O X . . . X . O . O . O X .
16: . . O * . . X . . * . . . . . O X . .
15: . . . . . . . . . . . . . . . . X . .
14: . . . . . . . . . O . . . . . . . . .
13: . . . . O . O . . . . . . . . . . . .
12: . . O . . . . . . O . . . . . X . . .
11: . X O O X . . . . . . . . . . . . . .
10: . O X O . X . . X * . . . . . * . . .
9: . X X . . . . . . . . . . . . . X . .
8: . . . . . . . . . . . . . . . . . . .
7: . . X . . . X . . . . . . . . . . . .
6: . . . . . O . . . . . . . . O . X . .
5: . . . . . . . . . . . . . . . . . . .
4: . . . * . X . X . X . . . . . O . O .
3: . . . X . . . . . . . . . O . . . . .
2: . . . . . . . . . O . . . . . . . . .
1: . . . . . . . . . . . . . . . . . . .
a b c d e f g h j k l m n o p q r s t
For reference purposes, here is the influence map created by GNUGo:
The 8-step linear decrementing influence algorithm, with edge-reflection creates the following influence map:
19: -6 -7 -8 -9 -1 9 13 13 11 8 2 -5 -11 -15 -20 -21 -16 -9 -3
18: -11 -13 -16 -18 W B B 12 10 4 0 -8 -14 -17 -23 -26 W -3 3
17: -16 -18 -21 -24 W W B 10 9 6 B -9 W -20 W -30 W B 11
16: -18 -22 W -26 -28 -22 B 1 -2 -2 -5 -9 -13 -16 -12 W B 19 15
15: -17 -22 -27 -26 -26 -19 -13 -6 -6 -9 -8 -9 -13 -12 -6 3 B 19 15
14: -16 -21 -28 -30 -29 -20 -15 -8 -9 W -8 -10 -11 -9 -2 7 13 15 13
13: -10 -16 -24 -27 W -21 W -7 -6 -8 -7 -10 -10 -7 -2 7 16 20 15
12: -5 -10 W -23 -18 -16 -14 -6 -6 W -2 -4 -6 -4 1 B 20 22 18
11: 0 B W W B -7 -4 2 3 4 4 2 0 0 3 13 16 20 17
10: 3 W B W 25 B 6 11 B 9 8 5 2 3 6 12 15 16 14
9: 11 B B 28 27 21 14 17 16 12 9 8 5 4 4 10 B 12 9
8: 16 23 29 31 32 28 23 23 19 13 6 5 4 1 0 3 5 5 4
7: 15 22 B 32 34 29 B 24 17 12 5 2 0 -2 -4 -3 0 -1 0
6: 14 22 25 29 33 W 23 21 14 7 0 -4 -8 -11 W -7 B -4 -3
5: 16 21 25 30 31 29 27 24 16 9 1 -4 -9 -13 -13 -11 -11 -11 -9
4: 15 20 22 25 26 B 23 B 16 B -1 -8 -13 -16 -17 W -17 W -11
3: 13 18 18 B 23 22 19 17 13 6 -1 -6 -12 W -21 -21 -17 -17 -13
2: 10 14 14 15 15 17 14 12 9 W -3 -10 -15 -19 -21 -20 -17 -15 -11
1: 6 10 11 11 11 13 12 10 6 0 -3 -8 -13 -17 -18 -17 -13 -11 -8
a b c d e f g h j k l m n o p q r s t
The following will help in the interpretation of the map:
Positive = Black influence, negative = White influence.
+/- 8 = (or is roughly equivalent to) 1 side of a point being covered,
+/- 16 = (roughly) 2 sides of a point being covered,
+/- 24 = (roughly) 3 sides of a point being covered,
+/- 32 = (roughly) 4 sides of a point being covered.
(Perhaps integer division of the above numbers by 8, using a Min/Max of +/-3 would create a map more interesting to compare to GNUGo's.)
The 8-step algorithm claims that Black is ahead by an aggregate influence value of 305 (adding all the above values together), or 271 (adding up all the values on the above grid, subject to the maximum and minimum values of 24 & -24 applied compensate for overconcentration--I use this modified aggregate when getting the 1-ply juncture values, below), and points out Black has greater influence on 5 more points than White.
SP Lee asks why GNUGo does not credit White with more influence in the vicinity of D9. The GNUGo mapping, however does seem to be credible to my eye. While White can ruin Black territory here, it is only Black that has the potential to make territory. Thus a value biased towards White would be inaccurate. In fact, a single play at D9 or E9 effectively shuts White out of this little alcove.
The 8-step influence map reflects this assessment, showing very high values for Black adjacent to the White stone at D10.
My crude strategy recommendation includes not playing where the opponent has an influence value much greater than ABS 16 to 20 unless it is meant as a sacrifice, or if the connection to a live group is tactically rock solid. Since D10 is part of a group that is quite healthy, given the values such as -20 at its "north" end, plays at D9, E10 and E9 are obviously tactically sound, and might be worth checking out given the high Black values. However, anything further away would be considered an overplay. (And in fact, the effectiveness of a blocking move at D9 or E9 by Black verifies this assessment.)
As for these tactically safe moves, the one-ply test (used to find the "juncture" value of a play) shows that plays in this area should be considered as being amongst the lower priorities on the board at this point of the game. I'm sorry I haven't taken the time to create an output file of the "juncture" findings to share. However, from just eyeballing the numeric results, the top candidates for juncture moves are as follows:
For comparison, the lowest play currently on the board is valued at 201 points. The "juncture" value is the difference in the aggregate influence values of a White play vs. a Black play on a given point (but where I've also included the min/max of +/- 24 to compensate for overconcentration).
In other words, the 8-step algorithm believes that playing in this area of the board has little influence over the board as a whole at this time, relative to most plays elsewhere. And, in view of the Black stones already bounding this area (B9, C9, C7, G7, F11 and the stones on the 4th row), an incursion by White here would at the most only affect the status of the handful of points within that bounds. The open lower-center-right is a much higher priority.
I think it is cool that GNUGo's influence map also identifies this central neutral area, and perhaps neutral points should be used as candidate moves. But note that no distinction is made between neutral points that are pretty much irrelevant and those that cause potentially high shifts in score. Perhaps this shows that the 8-step influence algorithm, run at one ply's depth (since it can be run at a relatively low cost) would be a cost-effective tool? I offer this without knowing how the Cosmic algorithms work and if they also deal with the strategic concept of "juncture" moves. (I also think the 8-step influence mapping values could be useful as part of an early warning system to help hone a solution to the caching problem, but that is another email.)
I would be happy to share code or more analysis. I have less time to keep up with postings these days, so emailing me directly is the best way to make sure I see any requests.
Thanks, SP, for remembering my earlier post and sending me this request, giving me another opportunity to pitch this approach. I'm sorry if my analysis contradicts the direction you were wanting to take things, but that's my best understanding of the position.