Funny game

[Joost Buijs]

Today I came across the following funny game between Ares and Kingsrow:

[Event "Game 16, opening 31-27 20-24"]
[Date "2025.04.21"]
[White "Ares-NNUE v1.82 exp"]
[Black "Kingsrow 1.63"]
[Result "0-2"]
[FEN "W:W27,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50:B1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,24"]
1. 37-31 14-20 2. 41-37 10-14 3. 34-30 17-21 4. 30-25 21-26 5. 33-28 18-23 6. 39-34 12-17 7. 34-30 17-21 8. 44-39 7-12 9. 40-34 24-29 10. 45-40 20-24 11. 39-33 4-10 12. 49-44 12-17 13. 27-22 8-12 14. 31-27 12-18 15. 37-31 26x37 16. 42x31 14-20 17. 25x14 9x20 18. 31-26 20-25 19. 36-31 15-20 20. 46-41 2-7 21. 44-39 10-14 22. 50-45 7-12 23. 47-42 1-7 24. 41-36 3-9 25. 42-37 5-10 26. 48-42 10-15 *

A complete blockade after 27 moves with both players having 18 pieces on the board.

[Ed Gilbert]

At first glance I thought the black/white positions might be symmetrical mirrors, but they aren't quite, the man on 43 would have to be moved to 44.

Kingsrow thinks that white's first really weak move was 11. 39-33. The search score jumps from 12 to 56 almost instantly. 50-45 instead maintains a more even eval. What does ares think about the turning points in the game?

--Ed

[Joost Buijs]

Ares played this game on a single thread without incremental NNUE update (about 1 Mn/s), so it didn’t see the problem after 11. 39-33 coming in time. If I give Ares a bit more time, it agrees with Kingsrow that 50-45 is the best move, albeit with a slightly negative score.

The network sometimes struggles with mobility; the only way to solve this is to search deeper in branches with low mobility.
Compared to a pattern evaluator the slowness of network inference remains a drawback, the search depth is on average 2 to 3 plies less.
Maybe this will get better on a modern processor with a built-in NPU and BF16 support, time will tell.

Joost