Would recent bad-bishop patches help SF in Game 8 against A0?

[Ma Laoshi]

I have tried to see if recent updates to Stockfish to make it more aware of bad bishops would have helped in (published) game 8 of its match with AlphaZero, http://www.chessgames.com/perl/chessgame?gid=1899421 . There, a pawn sac by White on move 11 led to SF's LSB getting boxed in long-term, in a way that SF has had trouble understanding. So I looked at it with SF_2018-05-14, with contempt=0 for neutral analysis.

By White's move 19 (r2q1rk1/1b2bppp/2p1p3/ppP1P3/1Pp3Q1/2B3P1/P4PBP/R4RK1 w - a6 0 19), at d=40 (< 5 min on my average PC) SF evaluates the position as 0.00 (best move 19. a3 as in the game), so apparently it sees that White has full compensation for the sacrificed pawn. (All the same, SF doesn't play A0's key preceding move 18. dxc5, opting for 18. bxc5?! with a -0.2 score even at d=50 ~ 1 hour.) By White's move 25 (r1q3k1/1b2bppp/2p1p3/1pP1P3/1Pp1B1QP/2B3P1/5P2/2R3K1 w - - 3 25), it starts to see a plus score at d=51 ~ 30min, though it takes until d=56 ~ 3h until it sees a clear +0.41 White advantage (it suggests 25. Rd1 however; A0's Kg2 comes 5 moves later in the PV). With the hashtables warmed up from the latter search, you can step back through the game and SF continues to see that White is better, until at move 22 (5rk1/1b2bppp/2p1p3/1pP1P3/1Pp3Q1/2Bq2P1/5PBP/2R3K1 b - - 2 22) it thinks that Black could have gotten away (0.00 score) with 22. ... Rd8 instead of 22. ... Ra8 played by SF8 in the match, at least up to d=68. Without much more analysis I cannot be sure whether the latter is correct, and whether the White advantage on later moves implies a forced win from those positions already; there is no suggestion that A0's play in the game, while clearly very strong, was perfect.

For comparison, I gave move 19 also to SF9. Even at d=55 ~ 1 hour, it thinks White is slightly behind (-0.07), suggesting 19. Rfd1 instead of A0's 19. a3. At move 25, it sees a white advantage only at d=60, which take > 12h to complete (suggested move Qh5).

Bottom line, the changes since SF9 are not dramatic, but at least they go in the right direction. It's probably purely hypothetical to ask whether current Stockfish could have drawn this game with A0, these things play out differently each time. These findings are broadly in line with what I've seen DragonMist report on SF's TCEC loss to Komodo (http://tcec.chessdom.com/archive.php?se=11&di=P&ga=28), where the eval gap between the wrong 11. ... Rg8 and the correct 11. ... Bb7 has narrowed, although SF still misplays that position.