It is instructive to analyze one of the games in detail to get a feeling for Gobble's non-human, probalistic way of ``thinking''. Computer go is usually not pretty, even to the eyes of the author who is a 13 kyu, but there are some interesting observations to be made. Remember that Gobble lacks even the simplest go heuristics, and any signs of go understanding produced by our algorithm are therefore of interest.
Figures 2 and 3 show the game record of one of the games played with strategy B and a two stone handicap. Black wins by 9 points.
Move 2. Figure 4 shows the board after White moved onto 2. To give an impression of how Gobble evaluates the position, the six highest and lowest values according to strategy A are shown. Border moves are considered bad by Gobble while moves somewhere in the center or near White 2 are preferred. Incidentally, a move onto the field now occupied by White,i.e. after White's stone has been captured, is worth about 40 points.
Let us again emphasize that there are noticable random fluctuations in the result, although 2000 random games were played. Figure 5 shows exactly the same position for a second evaluation. Notice that the average value has shifted by 2 points and only one move made it into the top six twice, and for some reason the idea to attack White's stone does not play a role. The lesson is that there may be many moves which cannot be differentiated by Gobble. However, among the top six there always seems to be one or two good moves on an objective scale, which is remarkable in itself. The challenge is to find them, which is one motivation for strategy B. As with human go players, many will recognize that a move should be played ``somewhere in the center'', fewer are able to pinpoint the best move.
Move 4. White has knowledge about fuseki.
Move 5. Definitely much too tight. We tried to reproduce this move with strategy A, but failed. 400 games per move in strategy B may allow to much randomness. Also see moves 9 and 11.
Move 6, 7, 8. Nice shape.
Move 9, 11. These moves look good to Gobble because on average a connected group with many liberties is more likely to survive than separate stones.
Move 12. Defends territory.
Move 13. If connected pieces are strong, a cut produces a weakness.
Move 14–24. White defends its territory while Black is cluttering the center.
Move 25. Having one eye under all circumstances is another feature which is valuable on average, and this move is sente.
Move 26. By now White has a big advantage and it is still open whether Black is capable of making a second eye.
Move 27, 28. Black tries to make territory for the first time, White takes it away at the other end of the board.
Move 29, 30. A futile invasion.
Move 31. Black probably considers White's stones in the lower right corner as half dead, simply because with a probabilistic algorithm there is always a chance that a player does not notice an atari. This chance may be slim, but if the gain involved is large enough, such a move will be played.
Move 32. Uncharacteristically poor move for White, the loosing move.
Move 33. Black is able to take advantage of White's mistake.
Let us pause here for a moment and consider the values strategy A would assign to this position. Figure 6 shows that Black 33 clearly stands out as the best move for Black with 20.2 points (gained by starting from this position). The worst move is worth 12.2 points. There is a large number of moves worth about 15 points. The difference of 5 points between these moves and the best move can only mean that Black wrongly considers part of White's stones in the lower right corner as dead.
Figure 7 shows the situation before White 32. That the best move is the one worth 5.7 points (for White, the lower the value the better) indicates that Gobble is consistent in that he makes the same mistake as in the evaluation for Black 33. Still, this move is much better than White 32 which scores a 10.5. The move values in the upper left corner suggest that Gobble does not understand the position that Many Faces of Go has build. We return to the game.
Move 47 (figure 3). This move should of course be played at 51. The bonus from an oversight by White must have outweight the territorial loss in the upper right corner. The single most important reason that Gobble looses in the endgame are such nonsense threats, while defending the borders is consistently underrated.
Move 49, 51. More meaningless ataris.
Move 57, 59, 61, 65. Meaningless.
Move 63, 69, 71, 73. At times when no competing flawed goals exist, Gobble's statistic is good enough to detect even small gains.
The remaining moves are not shown since they do not affect the outcome of the game.
Let us summarize what we have learned from this example. Gobble displays its own brand of typical computer go: surprising (relative) strengths in some areas and glaring weaknesses in others. Examples are life-death and territorial evaluations, respectively. Based on these observations there certainly are many natural suggestions for improvement of Gobble's strategy, some of which are mentioned in the following discussion.