mmt Vs. Ovyron (G4 D5 BG2)

[mmt]

SF with multi-PV=2 had:
21. ... Bc8 @ -2.16 @ depth 60
21. ... Be6 @ -1.95 @ depth 59

Looks like a few inaccuracies have been made.

Regards,

Zenmastur

SF also gets down to <-2.15 after Be6 for me, so I don't think it was an inaccuracy. Could be that the partial 7-piece EGTBs I got recently help SF see it quicker. Less effect on LC0 because of lower depths. I wonder if it's possible to improve LC0 somewhat by doing a SF-like search on some positions that are close to being converted to 7 pieces. We could take a bunch of random positions with 8 and 9 pieces from real games and compare SF to LC0 to see if there is a room for improvement.

[jp]

Less effect on LC0 because of lower depths. I wonder if it's possible to improve LC0 somewhat by doing a SF-like search on some positions that are close to being converted to 7 pieces.

So you feel from what you see that Leela is still weak in endgames?

[Ovyron]

SF with multi-PV=2 had:
21. ... Bc8 @ -2.16 @ depth 60
21. ... Be6 @ -1.95 @ depth 59

I'd like to hear about the PVs of those later, but I guess if mmt plays a move that isn't 22.Ba3... I mean 22.Bh3, then I'll know what was the difference.

I'm using "relative depth" of 62, constructed by interacting with the positions, so some lines are more shallow and others are deeper (specially the mainlines), but I'm still at some -1.80s scores, being unable to break into the -1.90s.

If my eval never goes past -2.16 I'll concede I need better hardware. If it does, and mmt doesn't miss any defense, that'd point to my Be6 > Bc8 conclusion being more accurate than depth 60. But this hangs from 22.Ba3... I mean 22.Bh3, so let's see if something else is played...

[Ovyron]

Actually it's possible that the only difference is complete lack of 6men TBs (or higher), so the positions are better than what my Stockfish can see at the same relative depth. mmt's best drawing chances might be to lead the game into positions that Stockfish with only 5-men TBs scores highly and thinks they're winning, but they're drawn.

[mmt]

So you feel from what you see that Leela is still weak in endgames?

I think there is not enough evidence yet. I'll run some tests after I do a few other projects.

[mmt]

mmt's best drawing chances might be to lead the game into positions that Stockfish with only 5-men TBs scores highly and thinks they're winning, but they're drawn.

Could be but I think I'll just play the best possible line I find regardless.

[Zenmastur]

SF with multi-PV=2 had:
21. ... Bc8 @ -2.16 @ depth 60
21. ... Be6 @ -1.95 @ depth 59

Looks like a few inaccuracies have been made.

Regards,

Zenmastur

SF also gets down to <-2.15 after Be6 for me, so I don't think it was an inaccuracy. Could be that the partial 7-piece EGTBs I got recently help SF see it quicker. Less effect on LC0 because of lower depths. I wonder if it's possible to improve LC0 somewhat by doing a SF-like search on some positions that are close to being converted to 7 pieces. We could take a bunch of random positions with 8 and 9 pieces from real games and compare SF to LC0 to see if there is a room for improvement.

I only used 6-man TB on this one and I did no reverse analysis. Just started it at let it run. After I got up this morning and posted I did a very quick reverse analysis on it and ended up with -2.40. I had it at over -3 but didn't want to spend the time it takes to get this back up to the root.

This is one reason I don't like turning on the machine and letting it run if I can help it. It doesn't produce very good results.

In the mean time I installed a new NVMe drive and am doing some tests on it. It took less than 15 minutes to copy 1TB of 5, 6, and partial 7-man TB to it. The only thing holding it back was the drive I was copying it from.

I install another drive as well but it was interfering with my boot drive so I temporarily unplugged it until I figure out how to resolve the issue. In the mean time I'm running speed test with 7-man TB's and other such things.

Regards,

Zenmastur

[mmt]

1. g4 d5 2. Bg2 Bxg4 3. c4 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 O-O 16. e3 Re8 17. Kf1 Bg4 18. Ne1 Bb4 19. Na4 Rb8 20. Nc2 Be7 21. f3 Be6 22. Nc5

[d]1r1qr1k1/p3bpp1/2p1bnn1/2Np2Bp/3P3P/1P1QPP2/P1N3B1/R4K1R b - - 1 1

[Ovyron]

Aha, so I missed Nc5. But if the reason was that black had some move that was killing 22.Bh3 that I couldn't see (it falls below Nc5), this is actually good. In that case I'd have played 22.Bh3 if I was white and lost faster. Let's see if I can reach the -1.90s now...

[Zenmastur]

My evals are nowhere close to that, but I suspect it's mostly because of lack of 6men TBs.

If TBs make all the difference, then this really is a matter of endgame ability. Maybe someone can give us evals with and without TBs to see what difference it makes.

I'm going to do that very thing. I have a couple of M.2 drives and just received a NVMe PCIe 4.0 drive. I plan on testing how much of a difference the speed of the drive makes when using 5, 6, and 7-man TB. On my current drives when using 7-man TB's the max requests per second is somewhere around 7,000-8,000. This slows down processing to about 40-45 Mnps. The new drive has twice the r/w speed and 3 times the IOPS speed. We'll see if this makes any difference. If it does, I may buy a couple more and make them into a striped array. If not, I'll buy 3 more of the cheaper drives and put them on a raid card.

In the mean time I'll run test without TB's.

Initial tests with the new drive show that it can sustain more than 3 times what my other drive can. Testing them is tricky. It's hard to find positions that will drive the TB requests to high numbers and keep them their long enough so that you can go to the right screens and record the data. Finding a position that can drive the request high enough for long enough to cause the drive to heat up to the point of throttling seems problematic. But this is a good thing I think. I haven't been able to test its peak yet but it sustains about 16k+ hard page faults a second and over 1.2GB of random reads per second with ease. My other drive could sustain 4k-5k and about 200Mb of random reads. This allow SF to maintain about 80-90 Mnps while heavy requests to 7-man TB files. I would say it's worth the price premium over a "standard" NVMe drive. One other advantage is it has is 8 times the write life. Not sure I need this since TB request are read only operations but I do have apps that beat the hell out of my hard drives with constant read/write cycles. This would be a good drive to run those apps. Now I just need 4 x 4TB drives like this one in a striped array and I'll be good to go!