Igo Hatsuyōron 120 and
Go-Playing Programs (2009 - 2011)
Introduction
For a long time we have faced massive challenges during the analysis of tenuki paths after the guzumi ( 432), due to the complex situation in the upper left corner. We were very aware that our amateurish endgame was not very reliable. On the other hand, extensive support from stronger players was not in sight. |
|
In addition, there was also a large uncertainty with us, concerning the effectiveness of the White atari on the left side. If this atari were actually possible and forced Black to capture with , ... |
|
...before Black could play atari on his own with (partially threatening his two marked stones), White would gain two points ( in the previous diagram, here) and the game would end either with a one point victory for Black, or with a one point victory for White. Even with a Black win, the possible existence of a 71st Black stone would be excluded (the game would now end with a jigo only), and we would have obtained a very unsatisfactory result overall. The strength of Go-playing programs approached slowly the boundary between Kyu- and Dan-players, and I thought that I might once even try using one of them for further analysis. Although I did not have the very powerful computer that would have been required fully to exploit the capabilities of these programs, the program's support should prove itself in the long run. The benefits of a program should be, that, even after playing through millions and millions of variations, it will not become "exhausted", and will also not "overlook" decisive points, only because these were not successful in previous studies. I chose "Many Faces of Go" for this purpose. Many Faces was not the most powerful program, but in was the only one I knew of, which had a western interface, and also offered the opportunity to display consistently, in a variation tree, the candidate moves that appeared as the most promising after evaluation by the program. However, the first trials in late 2009 did not proceed very efficiently. Usually, there were too many candidate moves, many of which could be easily identified as being not correct, or as simple forcing moves, which had very little to do with the relevant move sequences. That changed abruptly in the first half of 2010, when a noticeable jump in quality could be observed. Now, not only was the number of candidate moves reduced significantly, but also I got the impression that the program now "understood", and "internalized", the vital points of the problem. The diagrams, shown in this chapter, are based on the use of the currently latest version (12.22). The numbers shown in the tables represent the percentage of games, which start with a move on a particular point, and which when played out by the program ended in success. Due to their architecture, the programs play to their utmost, when a close outcome of the game is to be expected. It is the basic purpose of their programming to win by (at least) half a point. To simulate this condition, it is necessary to set Komi for White to an amount that corresponds to the expected Black lead at the end of the game (without Komi). It has also been found favourable to do the analysis repeatedly, with varied Komi, in order to consider the possibility, especially in the beginning of the move sequences, that the program may calculate a final result, which is numerically different from ours. I used the program mainly for the analysis of the designated "critical" paths - i.e. the variations, which ended with the smallest lead for Black. If the calculation of the program resulted in a candidate move which did not exist in the variation tree to date, a new branch was opened. | |