stockfish 10 vs. Mephisto III S Glasgow

[mclane]

We have different point of views uri. But I do respect your opinion.

[PK]

many programmers often speak out : why should I put knowledge into my programs when throwing knowledge out gives more ELO

Encoding human knowledge in the engine is the polar opposite of imbuing engine with chess knowledge. Throughout the years, I developed a way of parametrising engine to play like a human grandmaster. The funny thing is that terms I use became less and similar to human knowledge, material and positional weights more alien and pieces of information more and more generalised. I'm not speaking about removing knowledge - due to some mathematical trick with piece/square tables the last Rodent uses more than 2000 unique eval parameters, and I have a private engine that is almost as strong, but 30% slower and more evaluation-heavy. I am speaking about moving away from human abstractions of chess knowledge. I know for sure that to emulate human game one must move away from human thinking.

[Ovyron]

What Stockfish does within 1000 nodes is completely irrelevant because no sane people would use it like that. Basically, if Stockfish at 1000 nodes could defeat Mephisto, then it'd be used at 100 nodes and the question would be why it can't defeat Mephisto, with the same arguments.

[Raphexon]

What Stockfish does within 1000 nodes is completely irrelevant because no sane people would use it like that. Basically, if Stockfish at 1000 nodes could defeat Mephisto, then it'd be used at 100 nodes and the question would be why it can't defeat Mephisto, with the same arguments.

If the Mephisto emulator was "properly" working I'd have tested SF vs M3 at different nodes.
Right now if I wanted to do a tournament between SF and Mephisto at 5000 nodes (or whatever) it doesn't listen to the chessGUI and always take the same amount of time to make a move regardless of what the ChessGUI asks from it.

[mclane]

It’s much easier to reduce Stockfishs NPS to the level of mephisto III


There is a race between Komodo and Stockfish. Or other top chess programs (Houdini).
Shredder is Sometimes in this combat too.
Others don’t have the time or energy anymore because of family issues.

But why these fights ?
It would be much more senseful to concentrate forces and develop new ideas how the programs could Plan instead of only rely on search tree.

[mclane]

Experiment:

Mephisto III 8 bit 3.5 MHz on 4 mignon (AA) batteries. When on batteries the clock speed is reduced from 6.1 to 3.5 MHz.
On that level it makes 1 (!!!). NPS. Level 6 = 40/120 or 3 minutes a move.


I put stockfish10 on 180 Nodes per move.


[pgn][Event "Computer Schach Partie"] [Site "ORION8"] [Date "2019.12.06"] [Round "?"] [White "mephisto 3 8bit, 3.5 mhz"] [Black "Stockfish_10_x64"] [Result "*"] [BlackElo "2200"] [ECO "C15"] [Opening "Französische Verteidigung"] [Time "15:00:47"] [Variation "Winawer-Variante, 4.Bd3 dxe4 5.Bxe4"] [WhiteElo "2400"] [Termination "normal"] [PlyCount "36"] [WhiteType "human"] [BlackType "program"] 1. c4 e6 2. d4 c6 3. e4 d5 4. cxd5 exd5 {-0.13/2 0} 5. e5 Bf5 {+1.18/3 0} 6. Bd3 Be4 {+1.72/2 0} 7. Bxe4 dxe4 {+0.52/4 0} 8. Nc3 Bb4 {-0.77/2 0} 9. Ne2 Ne7 {-0.86/3 0} 10. Bg5 h6 {-0.06/3 0} 11. Bxe7 Qxe7 {-0.78/4 0} 12. Qc2 f5 {-0.75/3 0} 13. O-O O-O {-1.14/3 0} 14. Qb3+ Kh7 {-1.29/5 0} 15. Nf4 Nd7 {-1.17/3 0} 16. Ncd5 Nc5 {-4.61/3 0} 17. Qxb4 cxd5 {-4.61/3 0} 18. Nxd5 Qf7 {-5.44/3 0 Schwarz gibt auf} * [/pgn]

[mclane]

now with 6.1 mhz and stockfish with 360 Nodes per move, doubling on both machines:


[pgn][Event "Computer Schach Partie"] [Site "ORION8"] [Date "2019.12.06"] [Round "?"] [White "Stockfish_10_x64"] [Black "mephisto III 8 bit 6.1 mhz level 6"] [Result "*"] [BlackElo "2400"] [ECO "B45"] [Opening "Sizilianisch"] [Time "19:27:38"] [Variation "Klassisches System, Vierspringer, 6.Ndb5 d6 7.Bf4"] [WhiteElo "2200"] [Termination "normal"] [PlyCount "45"] [WhiteType "program"] [BlackType "human"] 1. e4 c5 2. Nf3 e6 3. d4 cxd4 4. Nxd4 Nf6 5. Nc3 Nc6 6. Ndb5 d6 7. Bf4 e5 8. Bg5 {+2.25/5 0} a6 9. Na3 {+1.92/5 0} Be6 10. Nc4 {+0.21/5 0} Rc8 11. Nd5 {+0.37/4 0} Bxd5 12. exd5 {+1.62/5 0} Ne7 13. c3 {+0.96/4 0} Nexd5 14. Qb3 {+3.84/5 0} b5 15. Nd2 {-0.70/4 0} h6 16. Bxf6 {-1.43/5 0} Nxf6 17. c4 {+0.23/5 0} Qa5 18. cxb5 {+0.77/6 0} Ne4 19. Qe3 {+0.31/5 0} Rc2 20. Bd3 {-2.54/4 0} Rxd2 21. Qxe4 {-4.65/4 0} Rxd3+ 22. Kf1 {-6.15/5 0} Qxb5 23. Kg1 {-6.58/4 0 Weiss gibt auf} * [/pgn]

[Uri Blass]

Experiment:

Mephisto III 8 bit 3.5 MHz on 4 mignon (AA) batteries. When on batteries the clock speed is reduced from 6.1 to 3.5 MHz.
On that level it makes 1 (!!!). NPS. Level 6 = 40/120 or 3 minutes a move.


I put stockfish10 on 180 Nodes per move.


[pgn][Event "Computer Schach Partie"] [Site "ORION8"] [Date "2019.12.06"] [Round "?"] [White "mephisto 3 8bit, 3.5 mhz"] [Black "Stockfish_10_x64"] [Result "*"] [BlackElo "2200"] [ECO "C15"] [Opening "Französische Verteidigung"] [Time "15:00:47"] [Variation "Winawer-Variante, 4.Bd3 dxe4 5.Bxe4"] [WhiteElo "2400"] [Termination "normal"] [PlyCount "36"] [WhiteType "human"] [BlackType "program"] 1. c4 e6 2. d4 c6 3. e4 d5 4. cxd5 exd5 {-0.13/2 0} 5. e5 Bf5 {+1.18/3 0} 6. Bd3 Be4 {+1.72/2 0} 7. Bxe4 dxe4 {+0.52/4 0} 8. Nc3 Bb4 {-0.77/2 0} 9. Ne2 Ne7 {-0.86/3 0} 10. Bg5 h6 {-0.06/3 0} 11. Bxe7 Qxe7 {-0.78/4 0} 12. Qc2 f5 {-0.75/3 0} 13. O-O O-O {-1.14/3 0} 14. Qb3+ Kh7 {-1.29/5 0} 15. Nf4 Nd7 {-1.17/3 0} 16. Ncd5 Nc5 {-4.61/3 0} 17. Qxb4 cxd5 {-4.61/3 0} 18. Nxd5 Qf7 {-5.44/3 0 Schwarz gibt auf} * [/pgn]

It is clear that stockfish made a losing move because of some search pruning trick that is bad at conditions of searching small number of nodes per move.


Black is worse but not clearly losing before 16...Nc5 so I analyzed the reason that stockfish did not play 16...cxd5 and the problem is that after 17.Nxd5 stockfish prune all the queen moves so it does not see that the queen can escape and believe that 17.Nxd5 wins the queen(only at depth 6 it can see that the queen can escape).

The analysis also suggest that stockfish developement version prune 17.Nxd5 for some reason and need depth 3 to see it(also seems strange).



[d]r4r2/pp1nq1pk/7p/3pPp2/1b1PpN2/1Q6/PP3PPP/R4RK1 w - - 0 17

Stockfish_19112109_x64_modern:
1/1 00:00 118 118k -1.20 17.Qg3
2/2 00:00 281 281k -0.96 17.Qg3 a6 18.Qg6+ Kh8
3/4 00:00 556 556k +7.18 17.Nxd5 a6 18.Nxe7
4/4 00:00 851 851k +6.69 17.Nxd5 Nb6 18.Nxe7 Bxe7
5/5 00:00 1k 562k +6.55 17.Nxd5 Rab8 18.Nxe7 Bxe7 19.f4
6/6 00:00 3k 1,273k +2.03 17.Nxd5 Qd8 18.Qxb4 Nb6 19.Nxb6 Qxb6


Of course it is a stupid search trick by human standard but it gives elo at long time control and it seems that the trick is from the beginning of 2015 based on analysis by stockfish developement versions that I have.

Unfortunately I could not find the relevant patch because the last page of stockfish patches is page 18 and page 19 and later pages seems not to be available for some reason.


http://abrok.eu/stockfish/?page=18

[Ovyron]

Experiment:

People have told you so many times that those are not relevant because what is a node for Mesphito is nothing like what is a node for Stockfish?

I hold that any decent programmer could make an engine that completely destroys Mephisto while searching even less nodes than Mephisto, but it wouldn't be an engine that makes plans and it wouldn't resemble smart AI. And it would be much weaker than Stockfish so my question would be, what would be the point of minimizing nodes searched?

[Ovyron]

Unfortunately I could not find the relevant patch because the last page of stockfish patches is page 18 and page 19 and later pages seems not to be available for some reason.


http://abrok.eu/stockfish/?page=18

They have deleted all old versions and only have pages that lead to the ones they hadn't deleted.

tmokonen has provided an archive of old versions hosted at Mega:

https://mega.nz/#F!G9tCiIZL!dBfqyOLHshiukt48WHbGTQ