Fritz 18: Evolving Genius

[BrendanJNorman]

1. There is a large difference between an engine that is *aimed* at being as strong as possible (Elo chasing) and an engine *aimed* at simulating a human player.

Here's an example from a long-delayed (but active) private (for now) project I have been working on with the legend Pawel Koziol.

This engine is not a derivative like OpenTal (which was a Rodent personality), but designed to play like Tal from the beginning.

This also includes a "weakening" so that he isn't a superhuman calculator.

In the following game, played against me at 25 + 5, I was winning for most of the game against Tal's "typical" unsound sacrifices.

But as a human, how do you see through the maze of complications? I didn't just as many of Tal's opponent's couldnt.

[pgn][Event "25 5 Training Game"]
[Site "?"]
[Date "2021.01.07"]
[Round "1.1"]
[White "Mikhail Tal Prototype"]
[Black "Norman, Brendan"]
[Result "1-0"]
[ECO "C12"]
[PlyCount "77"]
[EventDate "2021.??.??"]

{So this was round 3 of a double round-robin rapid tournament I was running -
I'd already beaten Randy Trainer 2-0 and was feeling solid. Let's see how I'll
go against the legend himself...Mikhail Tal!} 1. e4 e6 2. d4 d5 3. Nc3 Bb4 $5 {
I decided upon a Winawer French because I thought it was the perfect line both
to stay safe AND to provoke Tal into an unsound sacrifice. This turned out to
be correct.} 4. Bd3 (4. e5 {is more usual when I was planning to try
Petrosian's favorite} b6 5. Qg4 Bf8 $1 {which is a bit cramped, but also super
solid.}) 4... dxe4 5. Bxe4 Nf6 6. Bg5 $5 {Offering black the bishop pair,
which I guess is Tal's idea. He loves to hound your king with his knights!} h6
{Taking the offer.} 7. Bxf6 ({White loses a piece after} 7. Bh4 $4 g5 8. Bg3
Nxe4 $19) 7... Qxf6 8. Qd3 ({Relevant:} 8. Nf3 c5 9. O-O Bxc3 10. bxc3 Nd7 11.
Rb1 Rb8 12. dxc5 Qf4 13. Qd4 e5 14. Qd6 Qxe4 15. Rfe1 Qc6 16. Rxe5+ Kd8 17.
Qe7+ Kc7 18. Nd4 Qf6 19. Nb5+ Kc6 20. Nxa7+ Kc7 21. Nb5+ Kc6 22. Na7+ Kc7 23.
Nb5+ Kc6 24. Re6+ Qxe6 25. Nd4+ Kc7 26. Nxe6+ fxe6 27. Qd6+ Kd8 28. c6 {
1-0 (28) Debray,C (2334)-Renman,N (2349) Menton 2018}) 8... c6 $5 {Fixing the
d5 square. My development plan is to castle and then play ....Nd7 and
eventually ...e5 liberating my c8 bishop.} 9. Nf3 Nd7 10. O-O O-O 11. Ne2 {
[#] A nice idea. The knight hops out of theline of fire of the black bishop
and prepares to transfer to the kingside.} Bd6 $146 {This is a novelty over a
game between two club players from 2015.} ({In that game, black tried} 11...
Rd8 {and eventually lost after} 12. c3 Ba5 13. Rad1 Kf8 14. Ng3 Nc5 15. Qe3
Nxe4 16. Qxe4 Bd7 17. Ne5 g6 18. Qe3 Qg7 19. Ne4 Bb6 20. f4 Be8 21. Kh1 Rd5 22.
Rf3 Rad8 23. Re1 Rc8 24. Nd3 a5 25. Rg3 Ba7 26. Nec5 b6 27. Ne4 c5 28. dxc5
bxc5 29. c4 Rd4 30. b3 Kg8 31. Rh3 Bc6 32. Ndf2 Rcd8 33. Qe2 Bxe4 34. Nxe4 f5
35. Nc3 Re8 36. Re3 Qf7 37. Nb5 Re4 38. Rxe4 fxe4 39. Nd6 Qxf4 40. Nxe8 Bb8 41.
g3 {Lamottke,S (1349)-Helmer,A (1718) Verden 2015 1-0}) ({The best move is}
11... e5 $1 {after which black is doing okay.}) 12. c4 Rd8 ({Again} 12... e5 $1
{would be a lot better. I guess I was hesitating because I was intimidated by
Tal haha.}) 13. Rad1 Nf8 $6 (13... e5 $1 {is still fine. I'm playing way too
passively.}) 14. h4 Qe7 15. Qe3 Qc7 $2 {Still too passive.} ({Black MUST
challenge the white center and this was the last chance. Now white gets to
prevent it and swing to the kingside.} 15... c5) 16. c5 $1 $16 {[#] Forcing
the bishop away from controlling the e5 and f4 squares, so now white's knights
can come forward. Black's bishop pair has been shut down.} Be7 17. Nf4 b6 18.
Nh5 Ba6 {I thought I was "gaining a tempo" by attacking this bishop, but Tal
has other ideas.} 19. Ne5 Bxf1 20. Qg3 g6 21. Qf4 $3 {[#] Objectively this is
an unsound sacrifice, but it is also typical Tal move leaving several pieces
hanging in a chaotic position. Black grabs material at the cost of seriously
putting his king in danger.} f5 $1 22. cxb6 axb6 23. Qxh6 gxh5 24. Bxc6 Ba6 25.
Bxa8 Rxa8 26. Rc1 {[#] Until this point I was sure that I was winning,
especially since from d1 there is no rook lift to g3 (via d3). But this move
was annoying to see.} Qd6 27. Qxh5 ({I planned to meet} 27. Rc3 {with} Bxh4 {
although white still has a draw after} 28. Rg3+ $1 Bxg3 29. Qg5+ Kh8 30. Qh6+
Kg8 (30... Nh7 $2 31. Nf7+ $18) 31. Qg5+ Kh7 32. Qxh5+ $11 {etc.}) 27... Rc8 {
This whole game I've been passive and too eager to trade pieces.} (27... Qxd4
$1 {is possible, but I was scared of that potentially dangerous rook on the
c-file. Turns out I was scared of ghosts. White has to take the draw, and
attempts to win can actually backfire and turn into a loss. For example,} 28.
Qf7+ Kh8 29. Qh5+ Kg7 30. Rc7 Qd2 $3 (30... Qxe5 {is also fine for black.}) 31.
Rxe7+ $4 Kf6 $1 $19 {and black is actually winning!}) 28. Qf7+ Kh8 29. Rxc8
Bxc8 30. Qh5+ Kg7 31. a4 Bb7 32. Qf7+ Kh6 $4 {[#] Only NOW I get ambitious?
Here I was trying to avoid the repetition. Afterall, I'm two pieces up. But I
completely missed the danger of Tal's response.} ({I should just play} 32...
Kh8 $1 {and Tal gets his draw.}) 33. Qg8 $1 {[#] This quiet move ends the game
on the spot. White threatens the knight fork on f7 as well as a sneaky mating
net. I couldn 't believe it.} Qd5 34. Nf7+ Kh5 35. Qh8+ Kg4 36. Qg7+ Kh5 (36...
Kxh4 $4 {gets mated after} 37. Qh6+ Kg4 38. Qh3+ Kf4 39. Qg3+ Ke4 40. Qe3#) 37.
Qh6+ Kg4 38. Ne5+ {and now I lose my queen.} Qxe5 39. dxe5 1-0[/pgn]

*This* is how an engine with HCE is (and can be) much more human-like than an NN engine.

We also have more positional engines with tactical blindspots too, which also replicate a positional strong, but not the master-level player.

Honestly, I was seconds from buying Fritz 18, but am a bit disappointed with the "human-like" games it plays - especially since it (the Evolving genius feature) seems to be locked to the Gingko engine.

[Odd Gunnar Malin]

Honestly, I was seconds from buying Fritz 18, but am a bit disappointed with the "human-like" games it plays - especially since it (the Evolving genius feature) seems to be locked to the Gingko engine.

How do you know it is the Ginkgo engine?
I did rename this engine to something else and the 'Easy game' (Evolving genius) works just as before. It is as I said earlier in this thread an internal engine. And most probably an engine with normal evaluation, how else could you ask question like' is there undefined pieces, is the Knight on b1 in play etc.). Maybe it is a special version of Ginkgo. who knows. You can still use your Tal engine in friendly or rated mode.

[AdminX]

Nice Game Brendan,

Well annotated at that, and I enjoyed going over it.


[fen]r2r1nk1/p1q1bpp1/bpp1p2p/2P4N/3PB2P/4QN2/PP3PP1/3R1RK1 w - - 0 19[/fen]


For the record I think these are the types of positions that Fritz 18 can help one improve at, but I do understand that you do not care for it's style of play.


[fen]r4nk1/2q1b3/bp2p2Q/4Np1p/3P3P/8/PP3PP1/2R3K1 b - - 0 26[/fen]


Here Fritz 18 (Ginkgo 2021) thought your Qd6 was still even 0.00, and that you had better in the following.




A cool new feature of Fritz 18 is that if you were to mouse over any of the two variation shown it would preview the line in the main board window, which is a feature that I enjoy seeing. Anyway thanks for sharing this game, I look forward to seeing the end product.

[AdminX]

Honestly, I was seconds from buying Fritz 18, but am a bit disappointed with the "human-like" games it plays - especially since it (the Evolving genius feature) seems to be locked to the Gingko engine.

How do you know it is the Ginkgo engine?
I did rename this engine to something else and the 'Easy game' (Evolving genius) works just as before. It is as I said earlier in this thread an internal engine. And most probably an engine with normal evaluation, how else could you ask question like' is there undefined pieces, is the Knight on b1 in play etc.). Maybe it is a special version of Ginkgo. who knows. You can still use your Tal engine in friendly or rated mode.

Yes, whatever it is, it is clearly embedded and since this is a proprietary product we may never know. However I think it does a good job for what it was designed to do.

[Odd Gunnar Malin]

I tried Dragon 2.5 in Friendly mode (Dragon don't have UCI_Elo) and it work at level 700 so it does use some internal logic for other engines too. I think all what the engine need is multipv.
Then it probably comes to that it isn't importent which engine you use except when playing in full strength.

[BrendanJNorman]

Honestly, I was seconds from buying Fritz 18, but am a bit disappointed with the "human-like" games it plays - especially since it (the Evolving genius feature) seems to be locked to the Gingko engine.

How do you know it is the Ginkgo engine?
I did rename this engine to something else and the 'Easy game' (Evolving genius) works just as before. It is as I said earlier in this thread an internal engine. And most probably an engine with normal evaluation, how else could you ask question like' is there undefined pieces, is the Knight on b1 in play etc.). Maybe it is a special version of Ginkgo. who knows. You can still use your Tal engine in friendly or rated mode.

Well, pretty much everyone referred to "Fritz 18" as Gingko, so that's how I *assume* - only ChessBase really knows I guess.

If the Evolving Genius feature isn't locked to a single engine (whatever Fritz 18 is, not relevant to my statement), then I'll buy it for sure.

[BrendanJNorman]

Hi Ted, nice to hear from you.

For the record I think these are the types of positions that Fritz 18 can help one improve at, but I do understand that you do not care for it's style of play.

Hmm probably true, but I'm a bit stubborn in my training approaches.

Like you mentioned in this thread, I'd feel like I was cheating if the engine had to keep telling me there was a good move on the board. This is a little gimmicky to me.

When I play training games, the PV/eval is switched off, time on clock is equal, no takebacks, I relay moves to a real physical board and try to simulate a real tournament game as much as possible.

Hints kill the realism to me.

In the game I played, I KNEW that Tal's sacrifices were unsound, but very dangerous, yet was unable to find the right defenses.

Playing against a mainstream engine, you can pretty much resign if they sacrifice something because your position *must* be lost for them to do it.

This is the reason these "human" ones are good for strict tournament-like settings.

[fen]r4nk1/2q1b3/bp2p2Q/4Np1p/3P3P/8/PP3PP1/2R3K1 b - - 0 26[/fen]


Here Fritz 18 (Ginkgo 2021) thought your Qd6 was still even 0.00, and that you had better in the following.

Yeah, have analyzed the game deeply with engines too (obviously )

The reason for ...Qd6 instead of ...Qa7/b7 is because this move covers the e6 weakness, the c7 rook entry and gives the potential if possible of ...Qd5/Bb7 counter-attack. ...Qd6 is the "natural" move, but unfortunately doesn't work tactically.

Most of the game I was doing quite well, only missing 33.Qg8! was my final mistake. Good training, no doubt.

The game tells me I need deep calculation training.

A cool new feature of Fritz 18 is that if you were to mouse over any of the two variation shown it would preview the line in the main board window, which is a feature that I enjoy seeing. Anyway thanks for sharing this game, I look forward to seeing the end product.

This is a VERY cool feature and another reason I like CB products is the analysis features and cool bells and whistles like this. I think I'll definitely grab it! Thanks for "selling" it to me haha

[Odd Gunnar Malin]

I tried Dragon 2.5 in Friendly mode (Dragon don't have UCI_Elo) and it work at level 700 so it does use some internal logic for other engines too. I think all what the engine need is multipv.
Then it probably comes to that it isn't importent which engine you use except when playing in full strength.

Some more try, I don't think it ask the engine for any moves if my setup is correct, at least for this level (700). I only saw a startup of the engine with setting the hash and ponder to false. Then nothing before the games was ended and it closed the engine with quit. If I get time I will try a little more with rated game and higher level.

[Cornfed]

The issue is that engines with handcrafted evaluations simply do not mimic a strong human, regardless of what elo it is at. They play like superhuman calculators who are masters at tactics, but absolutely suck at positional play and are crap in the opening.

I disagree with this, but feel that you are getting your tools mixed up. Let me explain...

Well said...and better than I could so I just deleted my intended response.

Perhaps all that is left to be said is...in the words of that great philosopher Taylor Swift: "Haters will hate".

I say that mainly to all who are posting in this thread who too often seem to have an automated/internal bias against all things made by Chessbase and those who want to compare their pet beta GUI with the Fritz GUI, which is marketed to a much broader audience, doing things the other GUI (which is promising for what IT is designed for) probably ever will...despite the almost daily request for 'this or that' to be added or changed or fixed.

[Madeleine Birchfield]

Did Fat Fritz (Lc0 net) come with Fritz 18 as well, or was that only a Fritz 17 exclusive?