1.g4 opening is losing?

[Jouni]

Interesting. Lc0 will soon see the win!?

[Ovyron]

I wouldn't trust Lc0's evaluation for anything, remember that if you follow blindly what it recommends against 1.d4 it suggests a losing line.

[Dann Corbit]

There are 5 bad opening moves. Here is the aftermath:
[d]rnbqkbnr/pppppppp/8/8/6P1/8/PPPPPP1P/RNBQKBNR b KQkq - acd 44; acs 74127; bm d5; c3 "d5"; cce 109; ce 59; id "C.A.P. 445895"; pm d5 {1176} c5 {482} e5 {263} d6 {161} Nc6 {135} e6 {116} c6 {18} h6 {16} h5 {15} g5 {13} g6 {10} Nf6 {8} a5 {8} f6 {7} a6 {5} b5 {5} f5 {5} Nh6 {4} b6 {2} Na6 {1}; pv d5 c4 dxc4 Qa4+ c6 Qxc4 e5 Nf3 Bd6; white_wins 716; black_wins 1315; draws 413; Opening Grob Opening: General. 1.g4; CaxtonID: 733; ECO: A00;
[d]rnbqkbnr/pppppppp/8/8/8/5P2/PPPPP1PP/RNBQKBNR b KQkq - acd 46; acs 83502; bm e5; c3 "e5"; cce 195; ce 55; id "C.A.P. 445903"; pm e5 {234} d5 {35} Nf6 {18} c5 {10} f5 {10} g5 {8} a5 {7} h5 {7} b5 {5} d6 {5} f6 {5} b6 {4} Nc6 {3} a6 {3} g6 {3} h6 {2} Nh6 {1} c6 {1} e6 {1}; pv e5; white_wins 89; black_wins 247; draws 24; Opening ECO:A00; Opening: Gedult's opening; 1. f3 *;
[d]rnbqkbnr/pppppppp/8/8/8/7N/PPPPPPPP/RNBQKB1R b KQkq - acd 41; acs 66413; bm d5; c3 "d5"; cce 121; ce 46; id "C.A.P. 445905"; pm d5 {51} e5 {18} Nf6 {3} a5 {3} Na6 {1} c5 {1} c6 {1} e6 {1} f5 {1} g5 {1} h6 {1}; pv d5 d4 c5 e4 dxe4 d5 Nf6 Nc3 a6 Be3 Qd6 Bf4 Qb6 Qd2 Bf5 O-O-O Nbd7 f3 exf3 gxf3 g6 d6 e6 Kb1 Bg7 Bh6 O-O Bxg7 Kxg7 Nf2 Qa5 h4 b5 h5 Nxh5 Ng4 Kg8 Qh6 Rad8 Be2 c4 Ne4 Bxe4 fxe4 c3 Rxh5 gxh5; white_wins 19; black_wins 41; draws 21; Opening Amar Opening: General. 1.Nh3; CaxtonID: 45; ECO: E00;
[d]rnbqkbnr/pppppppp/8/8/7P/8/PPPPPPP1/RNBQKBNR b KQkq - acd 43; acs 71197; bm d5; c3 "e5"; cce 139; ce 34; id "C.A.P. 445896"; pm e5 {195} d5 {78} h5 {28} Nf6 {24} c5 {16} f5 {10} h6 {10} a5 {9} e6 {9} b5 {7} g5 {7} g6 {7} b6 {6} f6 {6} Nc6 {4} a6 {3} c6 {3} Na6 {1} Nh6 {1} d6 {1}; pv d5 d4 c5 e3 Nf6 a3 cxd4 exd4 Nc6 Nc3 Bg4 Nge2 Bd7 Bf4 Rc8 f3 e6 g4 Be7 Qd2 O-O h5 Ne8 h6 g6 b3 Bh4+ Kd1 Be7 Kc1 Nd6 Bg2 Qa5 a4 Rfd8 Kb1 Nb4 Be5 f6 Bf4 Qa6 Rd1 b6 Bg3; white_wins 119; black_wins 252; draws 54; Opening Kadas Opening: General. 1.h4; CaxtonID: 952; ECO: A00;
[d]rnbqkbnr/pppppppp/8/8/8/N7/PPPPPPPP/R1BQKBNR b KQkq - acd 43; acs 81199; bm e5; c3 "e5"; cce 285; ce 28; id "C.A.P. 445908"; pm e5 {9} Nf6 {5} c5 {5} Nc6 {2} b5 {1} d6 {1} e6 {1} f5 {1} g5 {1} g6 {1} h6 {1}; pv e5 e3 d5 d4 Nd7 Be2 e4 Nh3 Ndf6 Nf4 c6 c4 g5 Nh5 Nxh5 Bxh5 Bb4+ Bd2 Bxd2+ Kxd2 Be6 Be2 Qf6 Qc2 Ne7 Raf1 O-O-O Kc1 Kb8 Kb1 h5 h3 h4 cxd5 Bxd5 Nc4 Qe6 b3 f5 Ne5 Ng6 Nxg6 Qxg6 Rc1 Qd6 Qc5; white_wins 2; black_wins 20; draws 6; Opening Sodium Attack: General. 1.Na3; CaxtonID: 2348; ECO: A00;

[lkaufman]

I wouldn't trust Lc0's evaluation for anything, remember that if you follow blindly what it recommends against 1.d4 it suggests a losing line.

No engine currently in existence is close enough to perfection to trust it blindly, but unless we are talking about a clearly tactical position such as the Vienna variation double pawn gambit you are probably talking about, I would rather trust its evaluation than that of Stockfish, and I have found that opening positions with expected scores above 0.7 by Lc0 usually lead to wins when played out, and those below more often lead to draws. With Stockfish the evals jump around too much from ply to ply to set any reliable thresholds. Playing out the position either with or without human help is not any better, as there is no way to know if your very first move for either side was a mistake. Determining whether a bad opening position that is around this 70% threshold is a win or not is simply impossible today in most cases.
I note that the 5 bad opening moves analyzed in another reply all have overall losing scores (meaning more than 50% losses) in the database, but that probably just reflects that the players who play these bad openings tend to be weaker players than their opponents in general. Probably none of White's opening moves is actually losing, whereas it is likely that several Black replies to 1.e4, 1.d4, and 1.Nf3 are losing, not only the obvious blunders.

[Zenmastur]

I wouldn't trust Lc0's evaluation for anything, remember that if you follow blindly what it recommends against 1.d4 it suggests a losing line.

No engine currently in existence is close enough to perfection to trust it blindly, but unless we are talking about a clearly tactical position such as the Vienna variation double pawn gambit you are probably talking about, I would rather trust its evaluation than that of Stockfish, and I have found that opening positions with expected scores above 0.7 by Lc0 usually lead to wins when played out, and those below more often lead to draws. With Stockfish the evals jump around too much from ply to ply to set any reliable thresholds. Playing out the position either with or without human help is not any better, as there is no way to know if your very first move for either side was a mistake. Determining whether a bad opening position that is around this 70% threshold is a win or not is simply impossible today in most cases.
I note that the 5 bad opening moves analyzed in another reply all have overall losing scores (meaning more than 50% losses) in the database, but that probably just reflects that the players who play these bad openings tend to be weaker players than their opponents in general. Probably none of White's opening moves is actually losing, whereas it is likely that several Black replies to 1.e4, 1.d4, and 1.Nf3 are losing, not only the obvious blunders.

I'm not sure I agree with the first statement. There are MANY analyzing techniques that significantly improve the viability of A/B type engines even with modest hardware. As a CC player I spend most of my time analyzing endgames, EVEN, when in the opening phase of the game. In fact, almost all my search time for any particular games is in the endgame. This has a profound affect on the ability of A/B engines to produce accurate lines of play. To be sure, engines like SF have trouble with certain endgames, and this can be a MAJOR pain in the ass if you find yourself in one of those it doesn't “understand”. I learned very early on to avoid those endgame like the plague! They eat compute time for breakfast and ask for seconds and thirds. You can, in fact, resolve most of them using advanced analysis techniques but the time involved isn't worth the benefit so, it's much better to avoid them.

With proper hardware (proper meaning very powerful) and sufficient time I believe it's possible to “prove beyond a reasonable doubt” that moves like 1.g4 are in fact losses. I'm not saying they can be calculated all the way to mate, but that they can be calculated to +7/-7 or better/worse. I tried this with 1.g4 right after I started playing CC and couldn't quite get there. IIRC the best I could get after about a week of analysis was -3.xx or so. This was on 5M-6M nps machine with 16 GB of TT and 6-man EGTBes. I plan on trying this again when I build my new system and, before I increase my ICCF game load significantly. I think this is just a matter of how powerful a machine you use. If I had a dual EPYC 7742 machine with 2 or more TB or RAM, 7-man EGTBs and a week of spare time (2 at the most) I think 1.g4 would be history!

I'm not sure you would actually need this powerful of a machine for 1.g4. I'm more concerned about having enough TT that if can hold a sufficient number of moves from all the sub-trees that the sub-trees can still be traversed in a reasonable period of time near the end of the analysis. My experience from using machines with small TT's is that the size of the TT limits the maximum depth that can be attained in a reasonable period of time. It becomes virtually impossible for the program to drag the needed evaluations deep in the sub-trees to the root when the nodes needed are being overwritten long before they can be used. This I think is the limiting factor and NOT the raw speed of the machine ( assuming a reasonably fast machine, say 80+M to 100+M nps.)

Anyway, I plan on giving it a go when I get my new machine. It should make a nice “break-in” problem for it.

Regards,

Zenmastur

[MikeGL]

I saw at CCC (chess.com), they were using big hardware so maybe result was relevant.
Lc0 as black only drew against SF after SF as white played 1.g4 (0.5 - 0.5)
But SF (as black) convincingly won after Lc0 (as white) played 1.g4. (0-1)

edit: name of engines

[Zenmastur]

I saw at CCC (chess.com), they were using big hardware so maybe result was relevant.
Lc0 as black only drew against SF after SF as white played 1.g4 (0.5 - 0.5)
But SF (as black) convincingly won after Lc0 (as white) played 1.g4. (0-1)

edit: name of engines

While interesting, we're not talking about the same level of analysis. I'm talking about pulling the evaluations of leaf nodes that might be 150 to 200 plies from the root up into the sub-tree where the will have a "real" impact on move selection that is close enough to the root to cause the root move or moves close to the root to change. This can and does radically alter the evaluation of the root and the moves deemed best at or near the root. Someone looking at this type of analysis and comparing it with SF or other programs who have searched it to say depth of 80 ply will think you are nuts as the analysis will be like looking at ancient greek. Few if any moves will match. The only way to "prove" the analysis is valid is to do a complete reverse analysis on it with a machine that has sufficient TT to be able to keep the sub-trees in TT on VERY deep searches. This by itself can take several days. It's not like you can take a big machine start SF and let it run for a few days and expect to get similar results. You won't get anything even close. The main difference between this type of search and a "normal" search is that as you walk down the line of play in a "normal" search each move is, on average, searched to one less ply than the move preceding it. This type of compound search each move is searched to approximately the same depth until the end of the line of play. The lines are normally truncated at a point where a win is highly probable. As a practical matter you are no more likely to find a "serious" error on the last move of the line of play than you are on the first move of the line of play. This makes the line of play MUCH MUCH more reliable that the "crap" lines most engines spit out ad nauseam. That's not to say these lines are perfect, they aren't (you still have a limited time to do the analysis), they are just orders of magnitude better that what an engine will give you even if you let the engine run for a long time (i.e. days).

Regards,

Zenmastur

[zullil]

I tried this with 1.g4 right after I started playing CC and couldn't quite get there. IIRC the best I could get after about a week of analysis was -3.xx or so. This was on 5M-6M nps machine with 16 GB of TT and 6-man EGTBes.

Regards,

Zenmastur

Wow, that's quite surprising to me. My gut would suggest that 1. g4 draws.

I don't suppose you've retained the PV that yields such a decisive advantage for Black?

[Zenmastur]

I tried this with 1.g4 right after I started playing CC and couldn't quite get there. IIRC the best I could get after about a week of analysis was -3.xx or so. This was on 5M-6M nps machine with 16 GB of TT and 6-man EGTBes.

Regards,

Zenmastur

Wow, that's quite surprising to me. My gut would suggest that 1. g4 draws.

I don't suppose you've retained the PV that yields such a decisive advantage for Black?

I'm sure I have it. Unfortunately I have about 4 TB of of PGN's, Game DB's etc. They are not well organized. I did look and I found about 20 game databases devoted to the GROB. About 20 MB in total. So I guess the first order of business would be to go through them. I'm not hopeful it's in those, IIRC they are mostly MCTS games for the grob (i.e. short time control games to use in book making.) I have several other directories where it could be. I assume I gave it some name I would recognize. I hope so, I spent enough time analyzing it. We'll see, in the mean time, I have some time to kill so I started anther analysis. If I recall I used SF DD or a build close in time to SF DD to do the analysis. This may help find it as I can search the entire disk for this. The reason I started a new one is the 300 ELO difference in SF since I last did it, my hardware is much faster, (although not as fast as I would like) I have partial 7-man Syszgy now (vice 5 or 6 man) and the 7-man is loaded onto a NVMe drive ( vice a slow 5400 RPM HDD) and I have more than twice the ram as before.

This should make it more accurate and it should take less time than the original. It should also give me a better idea if I have a shot at scores like -7. A 300 ELO difference in engine strength is quite a bit. Who knows maybe it will tell me the position is dead even. I doubt it, but I've had stranger things happen even with scores over 4.xx. So far, I've done just about two passes (one down (d35) and almost 1 back to the root (d45)) and the score is -1.65 or so. I let the individual searches go deeper the more passes I do. So maybe 6 or 8 passes total, well see.

I'll do a HDD search tomorrow to see if that turns up anything. I could do a position search through the DB's, unfortunately I have literally hundreds of them (maybe a thousand) so that could take a while since I have to hand load each before I can search it.

I do have a medical procedure I have to have done as soon as it clears insurance (maybe the end of this week or early next week), assuming I survive it I should have a new analysis in the next week or so.

Regards,

Zenmastur

[zullil]

I tried this with 1.g4 right after I started playing CC and couldn't quite get there. IIRC the best I could get after about a week of analysis was -3.xx or so. This was on 5M-6M nps machine with 16 GB of TT and 6-man EGTBes.

Regards,

Zenmastur

Wow, that's quite surprising to me. My gut would suggest that 1. g4 draws.

I don't suppose you've retained the PV that yields such a decisive advantage for Black?

I'm sure I have it. Unfortunately I have about 4 TB of of PGN's, Game DB's etc. They are not well organized. I did look and I found about 20 game databases devoted to the GROB. About 20 MB in total. So I guess the first order of business would be to go through them. I'm not hopeful it's in those, IIRC they are mostly MCTS games for the grob (i.e. short time control games to use in book making.) I have several other directories where it could be. I assume I gave it some name I would recognize. I hope so, I spent enough time analyzing it. We'll see, in the mean time, I have some time to kill so I started anther analysis. If I recall I used SF DD or a build close in time to SF DD to do the analysis. This may help find it as I can search the entire disk for this. The reason I started a new one is the 300 ELO difference in SF since I last did it, my hardware is much faster, (although not as fast as I would like) I have partial 7-man Syszgy now (vice 5 or 6 man) and the 7-man is loaded onto a NVMe drive ( vice a slow 5400 RPM HDD) and I have more than twice the ram as before.

This should make it more accurate and it should take less time than the original. It should also give me a better idea if I have a shot at scores like -7. A 300 ELO difference in engine strength is quite a bit. Who knows maybe it will tell me the position is dead even. I doubt it, but I've had stranger things happen even with scores over 4.xx. So far, I've done just about two passes (one down (d35) and almost 1 back to the root (d45)) and the score is -1.65 or so. I let the individual searches go deeper the more passes I do. So maybe 6 or 8 passes total, well see.

I'll do a HDD search tomorrow to see if that turns up anything. I could do a position search through the DB's, unfortunately I have literally hundreds of them (maybe a thousand) so that could take a while since I have to hand load each before I can search it.

I do have a medical procedure I have to have done as soon as it clears insurance (maybe the end of this week or early next week), assuming I survive it I should have a new analysis in the next week or so.

Regards,

Zenmastur

Hope the medical procedure goes well. No rush for the analysis!