Damir Desevac
Posts : Join date : Age : Location : Here is the position:
Ghppn likes this post
Like 1
Eelco
Posts : Join date : Damir Desevac wrote: Here is the position: Yes I think you are right that the move proves much better for White but I do not understand where the difference (in evaluation) comes from, unless it is simply a matter of training with more games or simply with the relevant opening position in the case of Stockfish, helped that it will try to maximize better depth and better King safety already since the times of Glaurung. The difference is so large... If stockfish is using its HCE after sacrificing a Bishop, because of the material imbalance does that not defeat the purpose of NNUE, when it is supposed to have knowledge especially of gambits? That makes not an awful lot of sense but probably Stockfish is using the hybrid eval here, a Bishop down is after all a large imbalance. White has the King safety, Black king trapped in the center, Rook on h8 immobilized, Black lost its castling rights, White two far advanced pawns against the king and against an isolated pawn Rook on a8 also not in the game. More than enough compensation and the PV proves it. But if Stockfish's NNUE is not playing it must rely on the very good King safety in the handcrafted evaluation here... And the NNUE trainer would need more gambit games that is obvious, if you don't want the HCE to take over again...
[pgn]
[Event "?"]
[Site "?"]
[Date "2022.08.11"]
[Round "?"]
[White "Blue Marlin 14.9a"]
[Black "Rebel 15.1 MOD"]
[Result "1-0"]
[SetUp "1"]
[FEN "r1bqk1nr/pp1pnppp/1bp5/1B6/3PP3/5N2/PP3PPP/RNBQ1RK1 w kq -"]
1. d5 cxb5 2. d6 Nc6 3. e5 h6 4. a3 Kf8 5. Nc3 Qe8 6. b4 a6
7. a4 bxa4 8. Rxa4 Ra7 9. Bb2 Qe6 10. Qd2 Qb3 11. Rfa1 f6
12. Nd5 1-0
[/pgn]
Playing out the opening position, shown in the pgn above with both engines running in parallel, me arbitrarily deciding when one of the engines may make its next move.
Rebel 15.1 (with verification search but that is not important, evaluation is everything here, until Rebel sees that Black is lost) on the final position after 12. Nd5:
Code:
Blue Marlin standing for the Stockfish family, same position:
Code:
Blue Marlin on the five best moves in the starting position, where also for Blue Marlin it must not yet be obvious how good White is here if playing the sac, just the compensation is a bit larger than for Rebel who also shows compensation but only to equality, well that difference can be 'explained' by style. But in the final position of the 'game' especially after playing Qb3 Rebel is much too optimistic again for Black.
Code: r1bqk1nr/pp1pnppp/1bp5/1B6/3PP3/5N2/PP3PPP/RNBQ1RK1 w kq -
Eelco
Posts : Join date : I think that I found a part of the reason why the evaluations were so different. A version was playing where in eval.cpp the NNUE evaluation was divided by 200 instead of 100. In this case (this opening position) where eval is so important that may easily have been fatal. For instance the engine would appear to have trouble to see a win if the evaluation would be halved, just by appearances alone (for the user) but also because the values of positional compensation are halved.
I made a version of Rebel MOD were the NNUE is divided by 95, so a little bit in the opposite direction. The PV changes a bit but White still has the same compensation to start with of course, it is at least worth the Bishop.
The new PV could be like this with 3...Ba5:
[pgn]
[Event "?"]
[Site "?"]
[Date "2022.08.11"]
[Round "?"]
[White "Blue Marlin 14.9a"]
[Black "Rebel 15.1 MOD"]
[Result ""]
[SetUp "1"]
[FEN "r1bqk1nr/pp1pnppp/1bp5/1B6/3PP3/5N2/PP3PPP/RNBQ1RK1 w kq -"]
1. d5 cxb5 2. d6 Nc6 3. e5 Ba5 (3... h6 4. a3 Kf8 5. Nc3
Qe8 6. b4 a6 7. a4 bxa4 8. Rxa4 Ra7 9. Bb2 Qe6 10. Qd2 Qb3
11. Rfa1 f6 12. Nd5) 4. a3 b4 5. Bg5 Qb6 6. h3 Kf8 7. Bd2
Qd8 8. Re1 b6 9. axb4 Bxb4 10. Bxb4 Nxb4 11. Qd2 Na6
12. Qe2 Nb4 13. Qc4 Nc6 14. h4 Bb7 15. Nc3 h6 16. h5 Na5
17. Qg4 $16 {possibly this is still won for White}
[/pgn]
Eval of Rebel after 17. Qg4:
Code:
Eval of Blue Marlin:
Code:
The structural problems that Black has in the position still seem very large. But the White squared Bishop is not totally immobilized anymore on c8 I forgot to mention that as one of the weaknesses in the first PV, that alone is worth a Bishop.
Ghppn likes this post
Like 1
texium
Posts : Join date : my cfish 12 can see it by depth 15 on nnue=pure, although its not using a stockfish net but a net trained on the caissa database. interestingly, the same net does not output d5 when using with the rodent nnue engine which tends to play like a GM, probably because a human would never play d5 and has yet.
Eelco and Ghppn like this post
Like 2
Sponsored content
Home 
