Zenmastur vs Ovyron No it's not another GROB!

[Zenmastur]

1. e4 e5 2. f4 exf4 3. Nf3 d6 4. d4 g5 5. h4 g4 6. Ng1 Qf6 7. Nc3 Ne7 8. Nge2 Bh6 9. Qd2 Nbc6 10. Nb5 Kd8 11.d5

[d]r1bk3r/ppp1np1p/2np1q1b/1N1P4/4PppP/8/PPPQN1P1/R1B1KB1R b KQ - 0 11

Unassisted Stockfish now at 0.00. Depth 70, but no 7-man endgame tables. Looks like Ovyron is being led like a lamb to the slaughter, unaware of who's really in charge here.

lc0-384x30-t60-3070, also unassisted, at -0.32, depth 45. So far - since my last comment - lc0 and SF-dev have been following the same line as the one played here.

Once I got a graphics card that could run NN engines at a reasonable speed I thought this would "fix" many of Lc0 short comings. It hasn't. I ran Lc0 side-by-side with Stockfish for several hours last night while I was analyzing. When ever Lc0 would select a different line than SF I would save both and then analyze them to see which was "right" or "better" or which had errors in it. In every case I investigated Lc0 lines had errors in it. By "errors" I mean moves in the line that evaluated significantly different than the evaluation given to the line as a whole by Lc0. This wasn't a one-off per line of analysis, there were several moves like this in each of Lc0 prefered lines. I tried several different nets, including 62078 320x24 and segio's 512x40 test net. I figured the larger nets would do better since the analysis times, while they varied, were around 300 seconds per move. This didn't seem to help.

So, anyone have any suggestions on a net that under longtime control doesn't have this issue?

Not exactly sure, but the line given by SV-384x30-t60-3070 was in agreement with SF for more than 30 plies - when I checked them both a few moves ago -.

I just looked at all the nets I have and that one wasn't among them, so I'm downloading it now. I'll give it a spin and see how it does.

How deep were the searches?

Regards,

Zenmastur

[drewdrew]

1. e4 e5 2. f4 exf4 3. Nf3 d6 4. d4 g5 5. h4 g4 6. Ng1 Qf6 7. Nc3 Ne7 8. Nge2 Bh6 9. Qd2 Nbc6 10. Nb5 Kd8 11.d5

[d]r1bk3r/ppp1np1p/2np1q1b/1N1P4/4PppP/8/PPPQN1P1/R1B1KB1R b KQ - 0 11

Unassisted Stockfish now at 0.00. Depth 70, but no 7-man endgame tables. Looks like Ovyron is being led like a lamb to the slaughter, unaware of who's really in charge here.

lc0-384x30-t60-3070, also unassisted, at -0.32, depth 45. So far - since my last comment - lc0 and SF-dev have been following the same line as the one played here.

Once I got a graphics card that could run NN engines at a reasonable speed I thought this would "fix" many of Lc0 short comings. It hasn't. I ran Lc0 side-by-side with Stockfish for several hours last night while I was analyzing. When ever Lc0 would select a different line than SF I would save both and then analyze them to see which was "right" or "better" or which had errors in it. In every case I investigated Lc0 lines had errors in it. By "errors" I mean moves in the line that evaluated significantly different than the evaluation given to the line as a whole by Lc0. This wasn't a one-off per line of analysis, there were several moves like this in each of Lc0 prefered lines. I tried several different nets, including 62078 320x24 and segio's 512x40 test net. I figured the larger nets would do better since the analysis times, while they varied, were around 300 seconds per move. This didn't seem to help.

So, anyone have any suggestions on a net that under longtime control doesn't have this issue?

Not exactly sure, but the line given by SV-384x30-t60-3070 was in agreement with SF for more than 30 plies - when I checked them both a few moves ago -.

I just looked at all the nets I have and that one wasn't among them, so I'm downloading it now. I'll give it a spin and see how it does.

How deep were the searches?

Regards,

Zenmastur

I do not remember the actual depth, but the node count would have been about 1 billion.

[Ovyron]

Zullil could never guess who's in charge because the only thing he has learned about computer chess is how to put the engine to analyze a position

Right predictions: 4

Prediction: 14.Qe3

1. e4 e5 2. f4 exf4 3. Nf3 d6 4. d4 g5 5. h4 g4 6. Ng1 Qf6 7. Nc3 Ne7 8. Nge2 Bh6 9. Qd2 Nbc6 10. Nb5 Kd8 11. d5 a6

[d]r1bk3r/1pp1np1p/p1np1q1b/1N1P4/4PppP/8/PPPQN1P1/R1B1KB1R w KQ -

[Zenmastur]

I do not remember the actual depth, but the node count would have been about 1 billion.

Hmmm....How much memory is Lc0 using to get to a billion nodes?

Has the memory requirements per node searched by Lc0 changed? I thought it was 350-ish bytes per node. I noticed last night that it didn't seem to be using as much memory as I figured it should. I let a small net that was running over 100K nps run for over an hour and the memory usage didn't seem to change much. Not sure what's up with that. I figured it would crash when it exceeded the 100,000,000 node cache I had set. It didn't seem to care as it past 300M nodes.

[Zenmastur]

1. e4 e5 2. f4 exf4 3. Nf3 d6 4. d4 g5 5. h4 g4 6. Ng1 Qf6 7. Nc3 Ne7 8. Nge2 Bh6 9. Qd2 Nbc6 10. Nb5 Kd8 11. d5 a6 12.Nbc3

[d]r1bk3r/1pp1np1p/p1np1q1b/3P4/4PppP/2N5/PPPQN1P1/R1B1KB1R b KQ - 1 12

When you can predict what my 50th move is let me know!

[drewdrew]

I do not remember the actual depth, but the node count would have been about 1 billion.

Hmmm....How much memory is Lc0 using to get to a billion nodes?

Has the memory requirements per node searched by Lc0 changed? I thought it was 350-ish bytes per node. I noticed last night that it didn't seem to be using as much memory as I figured it should. I let a small net that was running over 100K nps run for over an hour and the memory usage didn't seem to change much. Not sure what's up with that. I figured it would crash when it exceeded the 100,000,000 node cache I had set. It didn't seem to care as it past 300M nodes.

I am not exactly sure what it was at 1 billion nodes, but right now at 220 million nodes, it is using 190GB.

[zullil]

Zullil could never guess who's in charge because the only thing he has learned about computer chess is how to put the engine to analyze a position

With that false and mean-spirited comment, now I'm really cheering for Zenmastur. But unless you blunder, this one's a draw.

[Alayan]

Except Zenmastur is playing with white and a draw is a poor result when the goal was to win.

[zullil]

Except Zenmastur is playing with white and a draw is a poor result when the goal was to win.

He very well might win. But only if Black blunders. The opening is obviously a draw with correct play.

Isn't that the point of the challenge? That Black, due to insufficient hardware and/or flawed methodology, will make a move that he considers correct but that is actually losing?

I mean, drewdrew playing Black using Stockfish with 400 threads isn't going to lose from here, no matter what Zenmastur plays.

[Ovyron]

Zullil could never guess who's in charge because the only thing he has learned about computer chess is how to put the engine to analyze a position

With that false and mean-spirited comment, now I'm really cheering for Zenmastur.

And what would be the difference? Think about it, imagine that this game was Ovyron Vs. Zenmastur, so I was white. Would you also be saying that Zenmastur was a lamb about to be slaughtered? Suppose I was white and we played some other line, what would you have said about white having a score of 0.00 at Depth 70 at move 12? Perhaps that my slow hardware made a difference and that's why I couldn't maintain my white advantage?

You're blinded by your biases.

The truth of the matter is that I've been playing at a level higher than Zenmastur's, as any sane person would pick the black's pieces here. We made decisions that had nothing to do with the hardware we had, with reversed colors people like Uri would be talking about how black could still win the game, but it's the same position!

Right predictions: 5

Prediction: 15.Nce2

1. e4 e5 2. f4 exf4 3. Nf3 d6 4. d4 g5 5. h4 g4 6. Ng1 Qf6 7. Nc3 Ne7 8. Nge2 Bh6 9. Qd2 Nbc6 10. Nb5 Kd8 11. d5 a6 12. Nbc3 Ne5

[d]r1bk3r/1pp1np1p/p2p1q1b/3Pn3/4PppP/2N5/PPPQN1P1/R1B1KB1R w KQ -