positional understanding of new chess engines

[peter]

When I have Syzygys installed at new machine and have some more hardware- time free again then, I'll maybe repeat the one or the other one game between SF and PlentyChess

Already fixed tbs and wanted to see them working in game playing anyhow, so here two more games of the known pairing. After 1...(28...)Ka5(!) PlentyChess 7.0.12 won again playing Black against SF dev. (White) and drew again playing White against SF playing Black after 28...Kb5(?):
[pgn][Event "DESKTOP-A1EFFT7, Schnellschach 45.0min+15sec"]
[Site "DESKTOP-A1EFFT7"]
[Date "2025.11.16"]
[Round "1"]
[White "Stockfish"]
[Black "PlentyChess 7.0.12"]
[Result "0-1"]
[Annotator "-1.67;-1.47"]
[WhiteFideId "-1"]
[WhiteFideId "-1"]
[SetUp "1"]
[FEN "2b3r1/6pp/1kn2p2/7N/ppp1PN2/5P2/1PP2KPP/R7 b - - 0 28"]
[PlyCount "123"]
[GameId "2244901569228852"]
[TimeControl "2700+15"]

28... Ka5 (28... Kb5 29. Ke3 {-0.71/52 291} Bd7 {-0.53/49 495 (Ka5)} 30. h4 {-0.41/45 93} Ka5 {-0.51/48 96 (Le8)} 31. Ng3 {-0.51/52 216 (Se2)} Bc8 {-0.25/49 406 (Td8)} 32. Nf1 {-0.19/56 233 (h5)} Rd8 {-0.11/43 82} 33. c3 {-0.19/57 0 (Sd2)} b3 {-0.31/42 43} 34. Nd2 {-0.17/54 0} Ne5 {-0.27/47 48} 35. Nd5 {-0.16/56 0} Ba6 {-0.20/47 70} 36. Nb4 {-0.16/49 0 (Kd4)} Bb5 {-0.27/38 38} 37. h5 {-0.11/47 0} Kb6 {-0.18/51 157} 38. Ke2 {-0.12/48 0} Kc5 {-0.10/50 317} 39. g4 {-0.13/54 0} Rd7 {0.00/54 146 (Ld7)} 40. Rf1 {-0.18/52 80 (Td1)} a3 {0.00/39 42} 41. bxa3 {-0.12/51 0} Nd3 {0.00/59 47 (Ta7)} 42. Rb1 {-0.12/55 224 (Tg1)} g5 {0.00/49 103 (h6)} 43. hxg6 {-0.05/52 75} hxg6 {0.00/55 0} 44. Rh1 {-0.28/15 0;½-½ PlentyChess 7.0.12-Stockfish DESKTOP-A1EFFT7, Schnellschach 45.0min+15sec (1) (g5) Remis angenommen}) 29. Nd5 {-1.67/50 668} Be6 {-1.47/42 88} 30. Nhf4 {-1.80/42 245 (Se3)} Bf7 {-1.98/41 109} 31. Ne3 {-1.93/37 30} Kb5 {-2.05/37 0} 32. Ke1 {-2.01/45 455 (Se2)} Ra8 {-2.45/43 117} 33. Ne2 {-2.24/42 119} Kc5 {-2.60/42 0} 34. Kd2 {-2.33/48 533 (Sd1)} Ra7 {-2.73/38 84 (h5)} 35. Nd1 {-2.49/51 441 (h4)} g5 {-3.05/52 416} 36. Nf2 {-2.61/48 0} h5 {-3.04/15 0} 37. c3 {-2.62/51 20} b3 {-3.12/38 0} 38. Ke3 {-2.64/42 50} Kb5 {-3.26/39 0} 39. Nd1 {-2.81/44 251} Kb6 {-3.60/50 0} 40. Nd4 {-2.83/33 10 (Ta3)} Ne5 {-3.63/43 84} 41. Nf2 {-2.87/37 0} Nd7 {-3.63/19 0} 42. Ra3 {-2.96/40 73 (Ke2)} Nc5 {-3.66/49 149} 43. Kd2 {-3.10/41 0 (Se2)} Be8 {-3.71/47 95} 44. Ra1 {-3.15/37 0} Bd7 {-3.33/13 0} 45. Ne2 {-3.22/34 50 (Ta3)} f5 {-4.31/43 63 (h4)} 46. e5 {-2.97/25 16 (exf5)} Bc6 {-5.01/33 68} 47. Nd4 {-3.62/28 0} Rd7 {-4.41/16 0} 48. Ke2 {-3.69/25 33 (h4)} f4 {-5.46/31 55 (Td5)} 49. e6 {-3.73/24 18 (Kd2)} Rd5 {-5.82/30 54} 50. g3 {-4.24/29 0} Bb5 {-5.77/17 0} 51. e7 {-4.29/26 16} Re5+ {-5.90/31 0} 52. Kd2 {-4.23/22 7} fxg3 {-5.97/29 0} 53. hxg3 {-4.61/28 28} Rxe7 {-6.05/31 0} 54. Nf5 {-4.69/25 22 (Th1)} Rd7+ {-6.45/36 70} 55. Nd4 {-4.77/28 0} Bc6 {-6.45/15 0} 56. Ke2 {-4.95/27 29 (Te1)} Re7+ {-6.74/34 73} 57. Kd1 {-4.98/24 0 (Kd2)} g4 {-6.94/32 74 (Le8)} 58. fxg4 {-5.77/23 30} hxg4 {-6.98/31 0} 59. Nxg4 {-6.44/23 17 (Sxc6)} Nd3 {-6.90/31 53} 60. Kd2 {-7.00/25 0} Nxb2 {-6.84/12 0} 61. Ne3 {-7.17/21 28} Rh7 {-6.95/32 0} 62. Ne2 {-7.59/18 17 (Sef5)} Rh2 {-7.07/36 62 (Lf3)} 63. Nf1 {-7.58/18 16 (Sg4)} Rh1 {-7.30/33 64} 64. Nd4 {-9.46/19 0 (Ke3)} Nd3 {-8.00/33 55} 65. Nc2 {-11.77/24 0 (Ke3)} bxc2 {-#2536/34 329} 66. Kxc2 {-74.51/22 0} Rg1 {-109.77/21 0 (Kc5)} 67. Rd1 {-11.16/25 29 (Kd2)} Kc5 {-112.67/25 267 (a3)} 68. Ne3 {-5.87/22 40 (Kd2)} Rxg3 {-112.67/21 214} 69. Kd2 {-199.77/35 0 (Sf1)} Nf2 {-112.67/20 172} 70. Re1 {-199.85/32 0} a3 {-112.67/17 0} 71. Ke2 {-199.88/29 44 (Sf5)} Ne4 {-112.67/19 22} 72. Kf1 {-199.92/28 0 (Tc1)} Rxe3 {#3864/16 28} 73. Rxe3 {-199.94/26 0 (Ta1)} a2 {#4164/23 138} 74. Re1 {-199.96/26 0} a1=Q {#4264/24 0} 75. Rxa1 {-199.98/37 13} Nxc3 {#4364/37 0} 76. Ke1 {-2/0 0} Na4 {-1/1 0} 77. Rd1 {-1/0 0} c3 {-8/1 0} 78. Rd8 {-8/1 0} Be4 {-7/0 0} 79. Kd1 {-7/1 0} Kc4 {-6/1 0} 80. Rc8+ {-6/0 0} Nc5 {-5/1 0} 81. Kc1 {-5/1 0} Bf5 {-4/1 0} 82. Rc6 {-4/1 0} Kb5 {-3/1 0} 83. Rc7 {-3/1 0} Kb6 {-2/1 0} 84. Rf7 {-2/0 0} Nb3+ {-1/1 0} 85. Kd1 {-1/1 0} c2+ {-1/1 0} 86. Ke2 {-1/0 0} c1=Q {-4/1 0} 87. Rf6+ {-4/1 0} Kb7 {-3/0 0} 88. Rf7+ {-3/0 0} Qc7 {-2/1 0} 89. Rxf5 {-2/0 0} Qc2+ {-1/0 0} 0-1
[/pgn]
Both with avx512 compiles and 15 threads of 16x4.3GHz CPU.

Uri latest wrote, that Ka5 was a surprising move to him, that's what I thought at first glance too, but indeed the position just isn't to be solved "positionally" at all, it's "simply" a highly tactical one

[Tibono]

the position just isn't to be solved "positionally" at all, it's "simply" a highly tactical one

Hi; yes indeed, solving requires deep tactical search; but solving the position was not the OP question.
Many strong & recent chess engines immediately spot Ka5 as their intended move, starting from very low depth, including depth 1 immediate best move for some (lc0_032+BT4 policyhead does).

The question was: as tactical can't head the choice at such a low depth, there must be some positional knowledge involved, which one?

[peter]

The question was: as tactical can't head the choice at such a low depth, there must be some positional knowledge involved, which one?

"Tactics" and "positional" eval in computerchess never was to be separated at all the way humans tend to do that, considering humans say "positional" mainly for "being not able or willing to calculate" the tactics fully the way the engines do so. Look at the games and output- lines of the thread and you see the "positional" "knowledge" necessary to solve the position(s, the first one given and the follow- up ones), whether you're able to understand the moves more or less, depends on your capability of calculating the tactics on your own, even if shown by the engines already and evaluating them "positionally" as well.

"Tactics" "deep" for humans don't have to be deep for engines at all, never had to and don't have to in times of neuronal nets anymore at all nevertheless, so yes, engines can calculate tactics like those in this one position very quickly nowadays and that gives the appearance of them solving them "positionally", but that's not new, already "strategic test suites" like the one from Swaminathan and Corbit way back then were to be solved with (for the time and hardware being actual then or now much differently of course) out of "static eval", so "strategically" instead of out of hardware- time to calculate them deeply. What you want and demand was in positions like that the insight into the black box NN or NNUE and the tactical calculation based on this "knowledge", so what?

That different engines have different "positional" insight and different time to depth here (as well as in other tactically for humans "harder" or "easier" seeming positions, very much depending and differing as for the single one human judging and solving them) you see in parameters like output- eval, changing over time to ponder, even if best move is found quickly, the winning eval for 1...Ka5(!) and the drawing one for 1...Kb5(?) and the ponder- time to separate the both of them better and better, says much more then choice of move only does, moves in longer output- lines together with evals changing over ponder- time show best (and with or without Forward- Backward differently well), what engines "understand" "positionally" and what they don't and just choose a certain move by chance, sometimes out of "incorrect reasons" at even for them tactically too deep positions as we know and see again and again at also for engines really tactically deep puzzles, at such ones mostly used for "single best move" tests, it's easier to see the correct reasons in follow- up- moves and evals, in positions like the one here problem is the lesser easyly to be seen difference in follow- up- moves, because of the broadness of the search tree and the number of transpositions between the lines "only", regards