How to effectively utlize Stockfish's strength for one's chess development

[KLc]

I try to avoid being led around by the nose with a multipv engine running in a window. My approach is using a shallow opening rep (5 moves) with good model games at all of the leaf positions. In some of those leaves, there aren’t many good human games to use as a model. I use GM Sadler’s method of generating model games with engines as I described above.

At the end of the day, you need to study those model games with your own brain and some strategy and endgame books, reaching for an engine to spellcheck your work only at the end.

It depends on what you want to do but when the goal is playing humans then I feel a problem with the shallow opening repertoire combined with model games by strong engines is that you won't consider some inferior but natural looking (and maybe even more frequently played) human moves (this is only relevant if you're not facing super GMs of course). I like creating model games at the leaves as you said, but this can be far down a line that computers never play—but humans do.

[towforce]

Let me preface this by saying I am open to being persuaded otherwise. But, honestly, I think you are mostly chasing an inefficient chimera here.

I agree. It sounds like taking a weak man to a muscle gym and telling him that the world record for a deadlift is 501 kg - oh, and look - I've loaded 501 kg onto this barbell for you!

Not only would the weak man fail to build muscle, but if he did somehow get the barbell off the ground, he would likely injure his back and then be out of the gym for many weeks.

These engines are stronger than top humans: it's likely that you'll be unable to discern the reason for a move choice just by playing lines.

Looking at a top chess engine's games is good fun, but it's not likely to be the best way to get better at chess IMO. Think of top engines as a magic show in which you see impossible things being done before your eyes!

Right now, the fastest way to improve your chess would probably be to pay a human trainer to work with you on your game. If you're rich, that's definitely the way you should go.

[bmp1974]

The problem with SF is that it is sometimes very difficult to make a connection between his move choices and the static features of the position. I prefer using one of the GM personalities in Lucas chess and an engine with a good evaluation, but a slower search. The engines I recommend are Komodo 8, Dark Toga (or Fire/Ethereal NNUE + the Dark Horse net [dh02]), and Imho, the best for this purpose, Marvin with the Rebel/Benjamin NNUE. One has also to know the characteristics of the engine it uses -- Komodo is extremely positional, but can misses tactical shots, Marvin_Rebel is a bit overoptimistic, but plays very active, searches for initiative, Dietrich's dh-02 net is also very positional (I think that, for this purpose, it is a bit better than The White Rose).

Playing "assisted" is also good for improving one's understanding of positions, but also for improving focus and calculation abilities. It has some serious drawbacks, though.

In general SF was not optimal for human play and improvements. It could be used to analyze a critical variation though, where the depths it achieves in a short time are useful. But there, I would rather recommend one of the derivatives/clones: Sugar, Shashchess etc. where the authors tried to adapt/improve SF usefulness for analysis. I think the situation has improved with this new SF/Leela hybrid in this regard, and an advanced player (not me, obviously!) could make really good use of this engine.

Thanks 'matejst' for your valuable suggestions. I will use your recommended engines to seek more human moves.

[bmp1974]

Let me preface this by saying I am open to being persuaded otherwise. But, honestly, I think you are mostly chasing an inefficient chimera here. How can an engine - which does not convert moves into words, do much to help (some GUI’s do so…to various degrees of success) in “One’s chess development”?

Openings…well, I can see how a player with an existing good understanding of how to play decent chess can get help there; but you used the blanket “one’s chess development” and reference a book which as I remember is about ones...general ‘chess development’.

Perhaps you just mean using an engine to weed out bad opening moves and get a better idea of what you should/should not do in a given opening?

Even there…I think you could do better to just use your brain directly in conjunction with an engine which has good multi-pv – Dragon MCTS, for example but even Stockfish would do - and investigate specific openings by tabbing thru...opening lines/book lines/ pgn file of an opening which has words from a competent annotator, supplemented by your own investigations… to check for interesting ideas or problematic moves in the lines and for ways to exploit those?

I glanced at my notes to an odd line for a tournament this weekend…which sadly I did not go to because of forecasted bad weather. This isn’t actually my ‘black opening tournament book’ which I created a while back with help of the method I mention above, just a fragment from a game referenced in it:

[pgn]
[Event "?"]
[Site "?"]
[Date "2019.01.20"]
[Round "?"]
[White "IM; 2436"]
[Black "Me; 2040"]
[Result "1-0"]
[ECO "B01"]
[WhiteElo "2436"]
[BlackElo "2040"]
[Annotator "Me"]
[PlyCount "24"]
[EventDate "2019.??.??"]

1. e4 d5 2. exd5 Nf6 3. d4 Bg4 4. Bb5+ c6 {One plays these gambit lines simply
to get an unbalanced position where one hopes to outplay their opponent. I've
had this position in blitz but never put much study into the line and probably should never have ventured it against this particular opponent.} 5. dxc6
Nxc6 6. Nf3 Qb6 7. Nc3 O-O-O 8. Bxc6 Qxc6 9. h3 Bxf3 10. Qxf3 Rxd4 ({In the
actual game I played} 10... Qxf3 {thinking I could play} 11. gxf3 Rxd4 12. Be3
{when I thought I could play 12...Rb4...but quickly realized after 13. b3 my
rook would be terribly misplaced and otherwise I lose my a7 pawn when the
long-term 3 vs 1 on the Queenside is just not going to be holdable.}) {By
tabbing thru this with multi-pv it became obvious that if I am to ever play
this line again, the position is more holdable with 10...Rd4} 11. Qxc6+ (11.
Be3 $5 Rc4 12. Qxc6+ Rxc6 13. Nb5 Rxc2 14. Nxa7+ {Is still 'a game'.}) 11...
bxc6 12. Be3 Rd7 $14 {and while White is still better, we still have a game to
play.} 1-0
[/pgn]

The human understanding and improvement from my quoted game is seen after move black's 8th move. 8...b5 & 9...Bb6 are slightly inaccurate and is not seen so far in the database. SF exploits with 10. Bg5 and the 11.Bh4 inviting g5. For me, this game plan deserves attention and can be adopted in the tournament game if Black delays h6. This is what I am trying to compare with "Master vs Amateur" model games.

[pgn][Event "Tour1, Rapid 15.0min+15.0sec-1"]
[Site "?"]
[Date "2022.01.15"]
[Round "1.1"]
[White "Stockfish 130122"]
[Black "Critter 1.6a 64-bit"]
[Result "1-0"]
[ECO "C54"]
[Annotator "0.22;0.08"]
[SetUp "1"]
[FEN "r1bq1rk1/2p2ppp/p1np1n2/1pb1p3/4P3/1BPP1NN1/PP3PPP/R1BQK2R b KQ - 0 9"]
[PlyCount "80"]
[EventDate "2022.01.15"]
[EventType "tourn"]
[SourceTitle "Fritz Engine Tournament"]
[Source "BM"]

{Intel(R) Core(TM) i9-9900K CPU @ 3.60GHz 3600 MHz W=38.8 plies; 6,135kN/s; 1,
869,277 TBAs B=22.2 plies; 14,659kN/s} 9... Bb6 {0.02/22 49 (h6)} 10. Bg5 {
0.94/32 18 (0-0)} h6 {0.00/24 36} 11. Bh4 {1.09/32 18} Na5 {0.00/23 85 (Be6)}
12. Bc2 {1.12/32 22 (0-0)} g5 {0.20/20 24 (Kh8)} 13. Nxg5 {1.65/35 20} Ng4 {
0.46/21 25} 14. Nf5 {1.49/37 22 (Nf3)} Bxf5 {0.31/23 23} 15. Nf3 {1.46/38 29}
Bxf2+ {0.31/21 0} 16. Bxf2 {1.45/40 26} Nxf2 {0.37/24 41} 17. Kxf2 {1.33/40 55}
Be6 {0.38/24 32} 18. Rf1 {1.25/38 25} c5 {0.37/24 33} 19. Kg1 {1.30/35 21} Kh7
{0.37/22 124 (Nc6)} 20. Qe2 {1.47/38 44 (Qd2)} Rg8 {0.35/21 29 (Nc6)} 21. Rf2 {
1.65/33 27 (Kh1)} Qe7 {0.41/21 100} 22. Kh1 {1.42/35 155} Rac8 {0.41/20 36
(Raf8)} 23. Raf1 {1.67/33 27 (g3)} Rg7 {0.41/21 26} 24. h3 {1.65/35 28 (Qe3)}
Kg8 {0.46/20 59 (Nc6)} 25. Bd1 {2.03/30 29 (Qd2)} Rc7 {0.48/20 24 (Nc6)} 26.
Qe1 {1.97/38 121 (Qe3)} Rg6 {0.50/21 23 (Nc6)} 27. d4 {2.68/36 36 (Be2)} Bc8 {
0.67/20 22 (Bc4)} 28. Nh2 {4.28/30 18 (b3)} Qh4 {0.61/21 53 (Qg5)} 29. Rxf7 {
5.74/31 30 (b3)} Qxe1 {0.00/21 8} 30. Rxc7 {6.17/30 23} Qd2 {0.12/22 28 (Qxf1+)
} 31. Rxc8+ {6.74/28 48} Kg7 {0.12/21 0} 32. Bg4 {6.90/28 14 (Ng4)} h5 {
0.00/22 12} 33. Rcf8 {7.38/26 17} Kh6 {0.00/23 11} 34. R1f7 {7.95/27 15 (Rh8+)}
Rg5 {0.40/22 97 (Qe1+)} 35. Nf1 {9.10/34 37 (Rf6+)} Qc1 {1.01/18 17 (Qf4)} 36.
Rh8+ {10.71/26 12} Kg6 {1.01/17 0} 37. Be6 {10.79/25 13} exd4 {3.10/19 74 (h4)}
38. h4 {11.94/60 28 (cxd4)} Re5 {8.68/21 20} 39. Rg8+ {11.94/59 24} Kh6 {
8.68/20 0} 40. Bf5 {10.74/55 86} Qxf1+ {9.71/23 17} 41. Kh2 {11.94/18 1} Qf4+ {
10.26/24 24} 42. Kh3 {11.94/54 5} Rxf5 {10.89/24 35 (Qg4+)} 43. exf5 {11.94/79
6} Qg4+ {11.56/25 27} 44. Rxg4 {12.53/62 8} hxg4+ {11.56/24 0} 45. Kxg4 {
13.23/49 8} Nc6 {11.80/24 22} 46. Rf6+ {13.62/55 19 (cxd4)} Kg7 {10.70/22 16}
47. Rxd6 {14.15/52 8} Ne5+ {10.82/23 15} 48. Kf4 {15.15/39 9} Nf7 {11.50/22 18
(Nd3+)} 49. Rxa6 {13.08/50 74 (Rg6+) adjud.} 1-0

[/pgn]

[Cornfed]

The human understanding and improvement from my quoted game is seen after move black's 8th move. 8...b5 & 9...Bb6 are slightly inaccurate and is not seen so far in the database. SF exploits with 10. Bg5 and the 11.Bh4 inviting g5. For me, this game plan deserves attention and can be adopted in the tournament game if Black delays h6. This is what I am trying to compare with "Master vs Amateur" model games.

My point is two-fold:

1). Overall “chess development” isn’t helped much – particularly compared to tried and true methods of increasing ones overall “chess development”…which largely require ‘words’, not variations and statistics.

2). Specifically, your approach which I see as largely for better opening play. In your example, Toga vs Stockfish, in Stockfish Multi-PV mode slots 8….b5 down to no better than the 6th ‘best’ move and that was going fairly deep (while preferring 8…h6) and it’s the same for 9…Bb6, just further down. Just glancing at that…might it not tell you exactly what you contend...that the move pair is...'lacking'?

Frankly, the multi-pv approach is so much more efficient (but maybe some have all the time in the world?) for opening 'understanding' than having engines play matches against each other in a host of different openings and hoping to use ones brain with it’s current level of development (I like thinking of towforce's weight lifter analogy )to interpret the results.
FYI – if you listen to Peter Svidler and Jan Gustafson’s commentary toward the end of Carlsen vs Giri in today’s Tata Steel game, they talk a little bit about engine use by GM’s. They specifically mention what I have recommended. Also, they also say that GM’s do not practice/train against engines…just use them alongside them as they are preparing their openings.

Actually I have Sadlers book but have only glanced at it. I have a nagging suspicion that he is perhaps trying to cash in a bit (not so much money wise, but attention from the chess public…it is his former profession after all…can't blame him) with a follow up to his Alpha Zero book ‘fame’.

[matejst]

Thanks 'matejst' for your valuable suggestions. I will use your recommended engines to seek more human moves.

You are welcome. You can find the Rebel neural net here: http://rebel13.nl/epoch-78x.nnue You should use it with Marvin 5.2, instead of Martin's original NN. It is much faster than Rebel 14, and has multiPV. I strongly recommend using multiPV for analysis -- even only two lines. The engine very often finds better moves. Be careful: this net over-evaluates initiative and attack! But for human play, and in dynamic openings, it's great.

I will share some of my insights about training [without help, by yourself] in the hope of someone finding it useful. In my experience, the most important thing is to work on tactical awareness. The things I do every day is to use Lucas Chess and the module "Learn tactics by repetition", or any other module (mate in 2, 3, mates in GM games, etc.). It helps tremendously. In the Serbian third league I faced players rated from 2000-2300 (25 years ago!) and at this level tactical awareness is still one of the two most important things. Playing blindfold is also a key exercise to improve calculation and awareness. There is also a module for this in Lucas Chess.

The second thing is to work on your technique. Two aspects are important: the endings and the middle games arising from your openings. In both case, using an engine is beneficial: analyzing, playing sequences of moves, training against an engine. It is very important to try all the ideas one has in a certain position to see why it is not good. _Learning what's wrong is as important as learning what's right._

The third thing are the openings. I recommend using systems that one can remember. If you play the English attack in the Najdorf, play it against the Classical variation, the Taimanov, etc. Be2 in the Sicilian is a great move, and similar moves can be used in the Pirc, against the Modern etc. The fianchetto is much more dangerous than it is usually thought, especially at lower levels. So many players lose their cool when you start playing f4, g4, g5.

Just looking at the lines given by an engine, trying to visualize them and understand them has a lot to offer. First, one learns how to chose candidate moves: I improved my understanding when I noticed that the first moves engines were considering were moves that attacked something. In my youth I learnt to defend, but I had problems not losing the initiative. It helps also not making blunders.

Don't lose your time playing blitz and rapid! In my experience, without one hour per game, you just wasted your time. The key is to develop the habits of methodically calculating variations until you simply "see" them.

[bmp1974]

The human understanding and improvement from my quoted game is seen after move black's 8th move. 8...b5 & 9...Bb6 are slightly inaccurate and is not seen so far in the database. SF exploits with 10. Bg5 and the 11.Bh4 inviting g5. For me, this game plan deserves attention and can be adopted in the tournament game if Black delays h6. This is what I am trying to compare with "Master vs Amateur" model games.

My point is two-fold:

1). Overall “chess development” isn’t helped much – particularly compared to tried and true methods of increasing ones overall “chess development”…which largely require ‘words’, not variations and statistics.

2). Specifically, your approach which I see as largely for better opening play. In your example, Toga vs Stockfish, in Stockfish Multi-PV mode slots 8….b5 down to no better than the 6th ‘best’ move and that was going fairly deep (while preferring 8…h6) and it’s the same for 9…Bb6, just further down. Just glancing at that…might it not tell you exactly what you contend...that the move pair is...'lacking'?

Frankly, the multi-pv approach is so much more efficient (but maybe some have all the time in the world?) for opening 'understanding' than having engines play matches against each other in a host of different openings and hoping to use ones brain with it’s current level of development (I like thinking of towforce's weight lifter analogy )to interpret the results.
FYI – if you listen to Peter Svidler and Jan Gustafson’s commentary toward the end of Carlsen vs Giri in today’s Tata Steel game, they talk a little bit about engine use by GM’s. They specifically mention what I have recommended. Also, they also say that GM’s do not practice/train against engines…just use them alongside them as they are preparing their openings.

Actually I have Sadlers book but have only glanced at it. I have a nagging suspicion that he is perhaps trying to cash in a bit (not so much money wise, but attention from the chess public…it is his former profession after all…can't blame him) with a follow up to his Alpha Zero book ‘fame’.

I agree with you but in parts. When I say 'one's chess development', its basically enhancing one's understanding of certain opening structure, how to lead them into a favorable middle game and so on. I am not talking about just opening improvement move for move. As SF & other engines have moved into the NNUE era, one can trust the engines for providing better insight into positional understanding of certain opening structures, themes, middle game nuances etc. For example, in the Sicilian Najdorf from white's side, one is always pondering about the correct time to play Nd5 or Kb1 or g4/h4 break etc.. So these types of Engine 'Master vs Amateur' games will certainly help. Also, there is always a chance to find novelty along the way or some crazy idea that has not been found in human DB games. In the multi -pv analysis one might miss on these subtle advantages as one is constantly looking at many lines/moves like a maze. Another disadvantage imho is that while doing a multi-pv analysis, one is not considering the positions in a frame work of a time controlled game as such.

[carldaman]

Thanks 'matejst' for your valuable suggestions. I will use your recommended engines to seek more human moves.

You are welcome. You can find the Rebel neural net here: http://rebel13.nl/epoch-78x.nnue You should use it with Marvin 5.2, instead of Martin's original NN. It is much faster than Rebel 14, and has multiPV. I strongly recommend using multiPV for analysis -- even only two lines. The engine very often finds better moves. Be careful: this net over-evaluates initiative and attack! But for human play, and in dynamic openings, it's great.

I will share some of my insights about training [without help, by yourself] in the hope of someone finding it useful. In my experience, the most important thing is to work on tactical awareness. The things I do every day is to use Lucas Chess and the module "Learn tactics by repetition", or any other module (mate in 2, 3, mates in GM games, etc.). It helps tremendously. In the Serbian third league I faced players rated from 2000-2300 (25 years ago!) and at this level tactical awareness is still one of the two most important things. Playing blindfold is also a key exercise to improve calculation and awareness. There is also a module for this in Lucas Chess.

The second thing is to work on your technique. Two aspects are important: the endings and the middle games arising from your openings. In both case, using an engine is beneficial: analyzing, playing sequences of moves, training against an engine. It is very important to try all the ideas one has in a certain position to see why it is not good. _Learning what's wrong is as important as learning what's right._

The third thing are the openings. I recommend using systems that one can remember. If you play the English attack in the Najdorf, play it against the Classical variation, the Taimanov, etc. Be2 in the Sicilian is a great move, and similar moves can be used in the Pirc, against the Modern etc. The fianchetto is much more dangerous than it is usually thought, especially at lower levels. So many players lose their cool when you start playing f4, g4, g5.

Just looking at the lines given by an engine, trying to visualize them and understand them has a lot to offer. First, one learns how to chose candidate moves: I improved my understanding when I noticed that the first moves engines were considering were moves that attacked something. In my youth I learnt to defend, but I had problems not losing the initiative. It helps also not making blunders.

Don't lose your time playing blitz and rapid! In my experience, without one hour per game, you just wasted your time. The key is to develop the habits of methodically calculating variations until you simply "see" them.

The highlighted part is music to my ears - over-optimistic NNUE play, not exactly what we've been used to!
Thanks for sharing this info!

Regarding time controls, I tend to prefer slower rapid for serious play, with games that end in about one hour, with some time for careful analysis afterwards.
Regular slow chess is too exhausting for me and takes way too much time.

[Cornfed]

The human understanding and improvement from my quoted game is seen after move black's 8th move. 8...b5 & 9...Bb6 are slightly inaccurate and is not seen so far in the database. SF exploits with 10. Bg5 and the 11.Bh4 inviting g5. For me, this game plan deserves attention and can be adopted in the tournament game if Black delays h6. This is what I am trying to compare with "Master vs Amateur" model games.

My point is two-fold:

1). Overall “chess development” isn’t helped much – particularly compared to tried and true methods of increasing ones overall “chess development”…which largely require ‘words’, not variations and statistics.

2). Specifically, your approach which I see as largely for better opening play. In your example, Toga vs Stockfish, in Stockfish Multi-PV mode slots 8….b5 down to no better than the 6th ‘best’ move and that was going fairly deep (while preferring 8…h6) and it’s the same for 9…Bb6, just further down. Just glancing at that…might it not tell you exactly what you contend...that the move pair is...'lacking'?

Frankly, the multi-pv approach is so much more efficient (but maybe some have all the time in the world?) for opening 'understanding' than having engines play matches against each other in a host of different openings and hoping to use ones brain with it’s current level of development (I like thinking of towforce's weight lifter analogy )to interpret the results.
FYI – if you listen to Peter Svidler and Jan Gustafson’s commentary toward the end of Carlsen vs Giri in today’s Tata Steel game, they talk a little bit about engine use by GM’s. They specifically mention what I have recommended. Also, they also say that GM’s do not practice/train against engines…just use them alongside them as they are preparing their openings.

Actually I have Sadlers book but have only glanced at it. I have a nagging suspicion that he is perhaps trying to cash in a bit (not so much money wise, but attention from the chess public…it is his former profession after all…can't blame him) with a follow up to his Alpha Zero book ‘fame’.

I agree with you but in parts. When I say 'one's chess development', its basically enhancing one's understanding of certain opening structure, how to lead them into a favorable middle game and so on. I am not talking about just opening improvement move for move. As SF & other engines have moved into the NNUE era, one can trust the engines for providing better insight into positional understanding of certain opening structures, themes, middle game nuances etc. For example, in the Sicilian Najdorf from white's side, one is always pondering about the correct time to play Nd5 or Kb1 or g4/h4 break etc.. So these types of Engine 'Master vs Amateur' games will certainly help. Also, there is always a chance to find novelty along the way or some crazy idea that has not been found in human DB games. In the multi -pv analysis one might miss on these subtle advantages as one is constantly looking at many lines/moves like a maze. Another disadvantage imho is that while doing a multi-pv analysis, one is not considering the positions in a frame work of a time controlled game as such.

Yes...it pays to be specific and not talk in generalities like "one's chess development", if one actually means something specific.

Multi-pv...I think we are on two different tracts here. "One might miss the subtle advantages"...I see it differently. Mainly: look at the difference in the eval for each line - tiny differences really mean nothing. Big ones do so (only speaking for myself here) I don't see a 'maze' to get lost in...just options to explore and outright worse moves. Engine matches do not seem worth my time in comparison.

I am not sure what you mean by "Another disadvantage imho is that while doing a multi-pv analysis, one is not considering the positions in a frame work of a time controlled game as such." I mean, what I highlighted in red...just doesn't make any sense to me with regards to using engine matches helping in "one's chess development". I am probably missing something...it's late.

[KLc]

You can find the Rebel neural net here: http://rebel13.nl/epoch-78x.nnue You should use it with Marvin 5.2, instead of Martin's original NN. It is much faster than Rebel 14, and has multiPV. I strongly recommend using multiPV for analysis -- even only two lines. The engine very often finds better moves. Be careful: this net over-evaluates initiative and attack! But for human play, and in dynamic openings, it's great.

Thanks for the link. It was impossible for me to find it on the website after you mentioned "Rebel/Benjamin net".