Man verses machine

[Lyudmil Tsvetkov]

Eval at the root says black has some 50cps advantage.

Can anyone disprove the axiom that eval at the root, with perfect play, equals eval at all nodes propagating along the tree? I.e., a draw at the root is always a draw anywhere at the tree, a white win is a white win anywhere at the tree, and a black win is a black win anywhere in the tree?

[hgm]

Only the score along the PV is equal to the root score (in absenseof a delayed-loss bonus). Other nodes can have arbitrarily different scores. If the position in the root is won, there can be plenty of moves that lose. Even moves that get you mated in one. E.g. in KQKR.

[Eelco de Groot]

I hope that is clear what Harm Geert said, Lyudmil? Or is it not what you meant to ask? It is not clear to me what picture you have of basic alpha beta searching. Nothing to worry about, you just need to see it in your mind. I would recommend the pages of Bruce Moreland but it is only available still on the Wayback machine. I don't know for how long. Because of their succinctness I found them helpful to understand some basic things, back in the previous century it was or maybe around 2000 So much wasted time.. Wiki has the link to slightly updated pages, but the original is long gone.

[Lyudmil Tsvetkov]

I hope that is clear what Harm Geert said, Lyudmil? Or is it not what you meant to ask? It is not clear to me what picture you have of basic alpha beta searching. Nothing to worry about, you just need to see it in your mind. I would recommend the pages of Bruce Moreland but it is only available still on the Wayback machine. I don't know for how long. Because of their succinctness I found them helpful to understand some basic things, back in the previous century it was or maybe around 2000 So much wasted time.. Wiki has the link to slightly updated pages, but the original is long gone.

Is not perfect play, as I mentioned, synonymous with pv?
Though, you should make here a sligth distinction: while a pv migth contain false best moves, perfect play does not.

What I wanted to say is that if the above position favours black in terms of eval at the root, it should favour black at presumably sound pv line also at ply 5,10,15,etc. And if engines see white in advantage above, there is obviously something very wrong with eval.

What I have observed lately is that, in less than 90% of cases, but in definitely more than 80%, in complex mg positions, when I say white leads by half a pawn, and SF and Komodo say, wait a minute, black leads by half a pawn, when I continue feeding in suggested alternatives to SF and Komodo main lines, SF and Komodo scores agree with me, but not immediately, just after 5 or 6 full moves.

Btw., above position is very strange, at the least - I do not know from where it is coming. Black seems to just have played Ng4 (why on Earth, if f4 instead sealed black advantage), the knigth hailing either from f6, or h6. Very strange move.

I tend to agree with Komodo that ef5 above is best option, but not to remain leading in score, but to just attempt to equalise with difficulty at some point. Objectively speaking, above position is a draw, but definitely black has the advantage. Is not it about time, that instead of heuristic score, engines start showing W/D/L? Easy to do: any score above a certain threshold at a specific phase is either win or draw.

Do not mention alpha-beta to me: I have been doing a lot of reading on it lately, but it is a bit tough to grasp comprehensively. Not plain alpha-beta, of course, but all the refinements.

The most interesting part of computer chess is still in stall. There are stupendous positions from what I see, involving deeper search, but they are accessible to engines currently only with much more refined eval. Current top engines are skipping a huge amount of promising lines, simply because their eval fails them. In terms of eval, not much to get from tuning terms, maybe 25% of available eval strength should come from tuning, the rest has to do with king safety(25%), new pawn features(25%), and other new eval features(25%). Overall, 50% of what could be done has to do with king safety, where all the deep lines are.

Started talking too much, sorry, just got carried away...Back to alpha-beta:)

[Lyudmil Tsvetkov]

My pv runs ef5 gf5 Nde4 Nf6 (certainly not Nh2, as some engines prefer) Nf6 Rf6 f4 (Komodo chooses this easily as best move, the only engine to do so rightly) e4 Be3 (so far all my best moves are also Komodo 8 best moves) c5 (here my Komodo tends to prefer something else, like Bd7):

[d]r1bq2k1/pp2n1b1/3p1r1p/2pP1p2/2P1pP2/2N1B3/PP2B1PP/R2Q1RK1 w - c6 0 6

and, although the position is already relatively simple, both Komodo and SF still give some 40cps white advantage, when actually black has the edge, and I believe this is already veryfiable with longer engine search. Black plays Kh7,Rf8,Ne7-g8-f6 next, with noticeably better position.

[tpoppins]

Btw., above position is very strange, at the least - I do not know from where it is coming. Black seems to just have played Ng4 (why on Earth, if f4 instead sealed black advantage), the knigth hailing either from f6, or h6. Very strange move.

<snip>

I tend to agree with Komodo that ef5 above is best option, but not to remain leading in score, but to just attempt to equalise with difficulty at some point. Objectively speaking, above position is a draw, but definitely black has the advantage. Is not it about time, that instead of heuristic score, engines start showing W/D/L? Easy to do: any score above a certain threshold at a specific phase is either win or draw.

The position arises in a sideline of the 7.Be3 system in the Classical KID, typically via the following move order:

[pgn][White ""]
[Black ""]
[ECO "E92"]

1. d4 Nf6 2. c4 g6 3. Nc3 Bg7 4. e4 d6 5. Nf3 O-O 6. Be2 e5 7. Be3 h6 8. O-O Ng4 9. Bc1 Nc6 10. d5 Ne7 11. Nd2 f5
[/pgn]

11...f5 was apparently introduced into GM practice by Nunn in his game with 1988 World Cup game with Kasparov. The game continued 12.Bxg4 fxg4 13.b4 a5 14.bxa5 Rxa5 15.Nb3 and the world champion managed to draw in 41 moves.

Three years later Nunn essayed the same line against Gelfand at Belgrade Investbank 1991. Black pressed for twenty moves longer than in the previous game, but despite the rating gap between the two being 1/3 of the gap in the previous case this game, too, ended in a draw.

12.Bxg4 fxg4 13.b4 a5 14.bxa5 Rxa5 15.Nb3 became the standard continuation in GM practice ever since, one notable exception being Wessman-Shirov, Gausdal 1990 where White, in a display of amazing foresight (the first Komodo release being more than 20 years away), chose to exchange on f5 before taking the knight and obtained the coveted draw in just 17 moves.

The line seems to have gone out of fashion by mid-90s, perhaps because White had been scoring rather heavily ever since for some reason. The only time a 2500+ player lost with White in this line was Iskusnyh (2533)-Yandarbiev (2360) at the 2001 St. Petersburg Open, where White went astray around moves 23-25, allowed his opponent to build an imposing pawn phalanx opposite his king, threw his knight in front of it (hoping to exploit the "hole" on f5) and ran into a powerful exchange sac.

The line is still encountered in master practice occasionally, and it seems like every time Black manages a win he also happens to outgun White by a couple hundred Elo point. The score is something like +11-4=3 for the 2001-2014 period, the latest game being Husman (2573)-Smirin (2660), Israel ch 2014 (32, 1-0).

[Lyudmil Tsvetkov]

Btw., above position is very strange, at the least - I do not know from where it is coming. Black seems to just have played Ng4 (why on Earth, if f4 instead sealed black advantage), the knigth hailing either from f6, or h6. Very strange move.

<snip>

I tend to agree with Komodo that ef5 above is best option, but not to remain leading in score, but to just attempt to equalise with difficulty at some point. Objectively speaking, above position is a draw, but definitely black has the advantage. Is not it about time, that instead of heuristic score, engines start showing W/D/L? Easy to do: any score above a certain threshold at a specific phase is either win or draw.

The position arises in a sideline of the 7.Be3 system in the Classical KID, typically via the following move order:

[pgn][White ""]
[Black ""]
[ECO "E92"]

1. d4 Nf6 2. c4 g6 3. Nc3 Bg7 4. e4 d6 5. Nf3 O-O 6. Be2 e5 7. Be3 h6 8. O-O Ng4 9. Bc1 Nc6 10. d5 Ne7 11. Nd2 f5
[/pgn]

11...f5 was apparently introduced into GM practice by Nunn in his game with 1988 World Cup game with Kasparov. The game continued 12.Bxg4 fxg4 13.b4 a5 14.bxa5 Rxa5 15.Nb3 and the world champion managed to draw in 41 moves.

Three years later Nunn essayed the same line against Gelfand at Belgrade Investbank 1991. Black pressed for twenty moves longer than in the previous game, but despite the rating gap between the two being 1/3 of the gap in the previous case this game, too, ended in a draw.

12.Bxg4 fxg4 13.b4 a5 14.bxa5 Rxa5 15.Nb3 became the standard continuation in GM practice ever since, one notable exception being Wessman-Shirov, Gausdal 1990 where White, in a display of amazing foresight (the first Komodo release being more than 20 years away), chose to exchange on f5 before taking the knight and obtained the coveted draw in just 17 moves.

The line seems to have gone out of fashion by mid-90s, perhaps because White had been scoring rather heavily ever since for some reason. The only time a 2500+ player lost with White in this line was Iskusnyh (2533)-Yandarbiev (2360) at the 2001 St. Petersburg Open, where White went astray around moves 23-25, allowed his opponent to build an imposing pawn phalanx opposite his king, threw his knight in front of it (hoping to exploit the "hole" on f5) and ran into a powerful exchange sac.

The line is still encountered in master practice occasionally, and it seems like every time Black manages a win he also happens to outgun White by a couple hundred Elo point. The score is something like +11-4=3 for the 2001-2014 period, the latest game being Husman (2573)-Smirin (2660), Israel ch 2014 (32, 1-0).

Thanks for the informative reply.

I assume this all comes from some kind of database, but I do not read databases.

Seemingly, 7.Be3, and 11.Nd2 are inaccuracies, so black is fully in the rigth to enjoy an advantage.

So, the belief is Bg4 is the best move, because Kasparov played it, and ... hardly managed to draw...
Even worse with Shirov. Obviously, Bg4 is no good.

I do not know about subsequent games, if the stats is true, relevant, non-biassed in some statistical way, etc. What I know is that static eval favours black by a non-winning margin.

Of all engines, Komodo performs best here.

Thanks for the informative post anyway.