mmt Vs. Ovyron (G4 D5 BG2)

[Uri Blass]

I believe better hardware can help unless the position is so bad that black is winning even with a relatively weak hardware.

[zullil]

I have never needed more than 128 MB for the TT, and never have needed high depth to analyze.

Link to your previous game?

It's here.

And my move:

1. g4 d5 2. c4 Bxg4 3. Bg2 c6 4. Qb3 e6 5. Qxb7 Nd7 6. Nc3 Ne7 7. cxd5 exd5 8. d4 Rb8 9. Qa6 Rb6 10. Qd3 Ng6 11. h3 Be6 12. Nf3 Bd6 13. h4 h5 14. b3 Nf6

[d]3qk2r/p4pp1/1rpbbnn1/3p3p/3P3P/1PNQ1N2/P3PPB1/R1B1K2R w KQk -

Komodo's top three at depth 45 were:

14...O-O (-1.07)
14...Nf6 (-1.02)
14...Bg4 (-0.84)

Nf6 was the top choice through most of the search.

[Ovyron]

I've seen what faster processors can do at Playchess, and I'm not impressed (my loses on there have been book loses, from decent positions the up to 9 higher depths that they're reaching seem to be doing nothing.) Software has bridged the gap, and my experiments with giving the engines 128MB hash, 256MB hash, 512MB hash, or 1GB hash (which I can give, no problem) produced analysis that was indistinguishable (because that thing about showing the engine refutations and making it remember them can be done with Learning engines, and everything else can be done with an analysis graph and exclude moves that doesn't benefit from more hash.) 64MB hash was noticeably worse, so there's a bare minimum, somehow.

I'm highly interested in playing Zenmastur in a *normal game* in the future, where I get black, and if I lose I'll concede that I need a faster processor and more RAM.

Komodo's top three at depth 45 were:

14...O-O (-1.07)
14...Nf6 (-1.02)
14...Bg4 (-0.84)

Nf6 was the top choice through most of the search.

For me it was the opposite, 14...O-O was a top choice for most of my search. The problem was that (CPU) Leela has hundreds of variations that were scored sub-1.00, and though I could refute them to some -1.40 one by one, I just wondered if there was one I couldn't refute, or if refuting her lines to -1.30 was going to be the subject of the rest of my life.

14...Nf6 is the Leela's move and it doesn't have this problem because she already shows high scores for black, so it's just must faster to work with it, which probably makes it best because white's choices are much more restricted.

[jp]

where I get black, and if I lose I'll concede that I need a faster processor and more RAM.

If you only draw with black against 1.g4 (against anyone), it'll show the same: either you need better resources or you need to use them better.

[Uri Blass]

I've seen what faster processors can do at Playchess, and I'm not impressed (my loses on there have been book loses, from decent positions the up to 9 higher depths that they're reaching seem to be doing nothing.) Software has bridged the gap, and my experiments with giving the engines 128MB hash, 256MB hash, 512MB hash, or 1GB hash (which I can give, no problem) produced analysis that was indistinguishable (because that thing about showing the engine refutations and making it remember them can be done with Learning engines, and everything else can be done with an analysis graph and exclude moves that doesn't benefit from more hash.) 64MB hash was noticeably worse, so there's a bare minimum, somehow.

I'm highly interested in playing Zenmastur in a *normal game* in the future, where I get black, and if I lose I'll concede that I need a faster processor and more RAM.

Komodo's top three at depth 45 were:

14...O-O (-1.07)
14...Nf6 (-1.02)
14...Bg4 (-0.84)

Nf6 was the top choice through most of the search.

For me it was the opposite, 14...O-O was a top choice for most of my search. The problem was that (CPU) Leela has hundreds of variations that were scored sub-1.00, and though I could refute them to some -1.40 one by one, I just wondered if there was one I couldn't refute, or if refuting her lines to -1.30 was going to be the subject of the rest of my life.

14...Nf6 is the Leela's move and it doesn't have this problem because she already shows high scores for black, so it's just must faster to work with it, which probably makes it best because white's choices are much more restricted.

1)Book losses means that you have a mistake in the book and I guess it means that you could avoid them by using faster hardware to check the book.
2)getting optimal results with 1.g4 for white or black may be harder than getting a draw from the opening position(I do not know).
The point is that in the opening positions you can make some non accurate moves in the first 10 moves and still draw the game when I am not sure if you can do the same with 1.g4

[Ovyron]

1)Book losses means that you have a mistake in the book and I guess it means that you could avoid them by using faster hardware to check the book.

My hardware is fast enough to check mistakes in the book. What is happening is that I'm using lines that were developed before July 2019, and back then people hadn't found their refutations yet. And I only check for improvements after losing games, because book tweaking solo isn't my thing.

The point is that in the opening positions you can make some non accurate moves in the first 10 moves and still draw the game when I am not sure if you can do the same with 1.g4

Maybe 1.g4 loses, if mmt with all his resources can't save it against me, I'll jump over the fence and join the people that think it loses. Because so far he hasn't played a blunder that I can see, so losing without a blunder would be very abnormal, unless 2.Bg2 loses by force (and white's second move choices are falling like flies.)

[mmt]

Yeah, with long analysis I think white can't save it. I added a new weapon, some 7-piece TBs, and I'll pay it out though. I guess 14. Ng5 might be worth playing out also. I think LC0 has done a lot worse than SF 11 in this opening. I could push LC0 to come close to SF11's evals when given SF's responses but the opposite wasn't true.

[mmt]

1. g4 d5 2. c4 Bxg4 3. Bg2 c6 4. Qb3 e6 5. Qxb7 Nd7 6. Nc3 Ne7 7. cxd5 exd5 8. d4 Rb8 9. Qa6 Rb6 10. Qd3 Ng6 11. h3 Be6 12. Nf3 Bd6 13. h4 h5 14. b3 Nf6 15. Bg5

[d]3qk2r/p4pp1/1rpbbnn1/3p2Bp/3P3P/1PNQ1N2/P3PPB1/R3K2R b KQk - 2 15

[zullil]

1. g4 d5 2. c4 Bxg4 3. Bg2 c6 4. Qb3 e6 5. Qxb7 Nd7 6. Nc3 Ne7 7. cxd5 exd5 8. d4 Rb8 9. Qa6 Rb6 10. Qd3 Ng6 11. h3 Be6 12. Nf3 Bd6 13. h4 h5 14. b3 Nf6 15. Bg5

[d]3qk2r/p4pp1/1rpbbnn1/3p2Bp/3P3P/1PNQ1N2/P3PPB1/R3K2R b KQk - 2 15

Komodo at depth 46 with MultiPV = 3 had

15. Na4 (-0.89)
15. Kf1 (-1.12)
15. Bg5 (-1.13)

[jp]

I think LC0 has done a lot worse than SF 11 in this opening. I could push LC0 to come close to SF11's evals when given SF's responses but the opposite wasn't true.

This on its own is interesting. We wouldn't say this opening is very tactical, would we? That's the usual excuse for Lc0's failures. "Strategic" doesn't have meaning for computer play (it's all calculation), but if we pretend it does, I'm not convinced NN engines are so fantastic "strategically". The problem may be Lc0's weak endgame play. If lines aren't resolved by tactics, long searches will end up in endgames, which will harm Lc0's evaluations at the roots.