I am analyzing an interesting position, shown below. It is rich in lessons, at least for me, so I figured I would share it.
By the way, I have a simple way to find interesting situations. When Pebbles loses, it saves the *last* position that it thought it was winning (i.e., the rating of the selected play was > 0.5). Many such positions occur late in the game, which always means that critical tactics are mishandled by the playouts. Here is the position: 1 2 3 4 5 6 7 8 9 A - - - - - - - - - B - O O O O X X X O C - X X X O O X O - D - O - X O X O O - E O - O O X X O X X F - O X X X X X O - G - X X - - O X O - H X O O O - - X X O J - - - - - - - O - White (O) to play. There are several battles going on here: 1) The X string on C2 is locked in a semeai with the O group below. X loses that battle regardless of the side to move. 2) The O group on the bottom middle does not have enough room to make life. It is dead in alternating play, except when it connects to the O group on middle left. Connecting depends on winning a Ko on G1, and O should always lose that Ko in alternating play, but in random play O often survives. 3) The bottom right group of loose O stones is pretty much always killed off. But while alive those stones have aji because they take way liberties of the X string on E8-E9. 4) Finally, we have the semeai on top right. The X string on B6 has 3 liberties, and the O string to its left has 2 liberties, so it seems that X should win regardless of who moves first. Taking this altogether, X wins well over 45 points, so X should win this game. But Pebbles (playing O) showed O as winning this situation something like 75% of the time, even though Pebbles had to resign in just a few more turns. The variations showed that Pebbles easily solves any one of these battles, but the position as a whole is hard because at least two of those battles are always pushed into the playouts. The playouts do respect alternating play in local battles, so the wrong winner often emerges. As payoff for reading this far, here is the first lesson from this situation. I have been adding rules that inhibit self-ataris. In this position, X has a vital tactic that *requires* a self-atari: X plays on C9 to win O's C8 string in a snapback. The lesson is that if you inhibit self-atari then you should consider testing for snapback. Pebbles tests for snapback in the MCTS, where I am sure that it pays off. I have not yet tested what happens in the playouts. Further analysis convinced me that O is actually winning this game. My current engine likes A8 for O until iteration 7000, and then F9 for O, and switches to the winning move only on iteration 143,000. But it doesn't really "see" the win, because the evaluation remains around 50.3% no matter how long Pebbles searches. Best, Brian _______________________________________________ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
