mmt Vs. Ovyron (G4 D5 BG2)

[mmt]

An update on my new drive. I was able to find a few positions that stressed it for extended periods of time. It was able to sustain 35+k page faults a second and 3.5 Gb a second. At this rate it does have a large negative affect on NPS. They dropped to 45M nps. So, for 7-man TB's it looks like you really do need a striped array of fast PCIe M2 cards to get best/better performance from 7-man TBs.

Yes, I've seen SSDs and NVMe (when I tested it for a bit) actually get stressed when SF was hitting tablebases.

One other note, I've come to the conclusion that 5-man with DTM for general analysis is better than 7-man syzygy and better yet 6-man with DTM would be far superior. I actually have 6-man Nalimov but few programs can use them.

Do you mean that without them it will be faster by enough that a completely accurate leaf score will matter less? Or another reason like converting a 7-piece position to a quick win?

[Zenmastur]

I collected 4,228 Grob games. I'm mass analyzing them to 15 ply. Then I'm going to cut off all the bad parts, hack them up into little bits and play engine games from the stubs. I think I'll start with 2, 3, 4, 5, and 6 ply start positions. I plan on playing enough games to build a reasonable book. Any suggestions on what the time controls work best for this. I plan on doing this in progressive steps. The first step will be a lot of short Tc games so it gets a wide variety of moves in the games. I'm thinking three engine round robin, with two games per engine-pair for each position multiplied by some small constant. Then collect the new games in a database chop them up and repeat with the new positions at a longer time control. And then repeat as needed. Etc. Unknow how long this will take as I haven't ever tried anything like this before.

Anyone ever do this before or have any suggestions?

Regards,

Zenmastur

[Zenmastur]

One other note, I've come to the conclusion that 5-man with DTM for general analysis is better than 7-man syzygy and better yet 6-man with DTM would be far superior. I actually have 6-man Nalimov but few programs can use them.

Do you mean that without them it will be faster by enough that a completely accurate leaf score will matter less? Or another reason like converting a 7-piece position to a quick win?

I don't have enough memory to hold all 7-man TBs (obviously.) When analyzing overnight it fills memory and then the page file with TB data> It then slows to a crawl. I've found it's better to use 5 or 6 man for overnight runs. I thought about buying 4 x 4 TB NVMe drives and a board to put them on but, from testing 5 and 6 man with DTM, it's a lot faster I don't have to spend a couple of grand on hardware to support it.

I wish Syszygy had DTM for 5 or 6 man. But they don't and I don't expect that they ever will.

[Ovyron]

Nd3

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 Bc8 23. Kf2 Nd7 24. Ne6 Qa5 25. Bxe7 Rxe7 26. b4 Qb6 27. Ng5 Ba6 28. Qa3 Rbe8 29. Bf1 Bxf1 30. Raxf1 Qc7 31. Qd3 a5 32. a3 axb4 33. axb4 Qd6 34. Rfg1 Nb6 35. Qf5 Nf8 36. Re1 Nc4 37. Nh3 Ra7 38. Qxh5 Ra2 39. Re2 Qe7 40. e4 dxe4 41. fxe4 Qd7 42. Ng5 f6 43. Nf3 Nd6 44. Nfe1 Rxe4 45. Qf3 Qe6 46. Rxe4 Nxe4+ 47. Kg1 Rb2 48. Rh2 Rb3 49. Nd3 f5

[d]5nk1/6p1/2p1q3/5p2/1P1Pn2P/1r1N1Q2/2N4R/6K1 w - -

[Ovyron]

Anyone ever do this before or have any suggestions?

I've done it. I have such a backsolved tree for all of chess (which was used to build my chessmap.)

My suggestion is to resist the temptation of quantity over quality, all those 15ply games will provide a lot of noise, and variety isn't a good thing (the only reason you want variety is that there's some mainline you don't want to miss, so playing more lines would catch it; if you'd somehow miss it there's something else to fix before starting this.)

What has worked best is doing it the other way around, you build trunks of analysis lines of the highest quality moves that you can find, then you search for alternatives to those moves at decreasing depth. Starting with low depth will only make you realize later how much time you wasted in some lines (say, having 2000 lines for a poor move and 100 for the best move when it should be the other way around.)

Another thing is game adjudication, you shouldn't bother with playing games to completion, depending on how badly scored a line is you'd only play games from it until they reach a certain score, and if they do you look for an alternative line.

Unfortunately the optimal way of carrying it out depends on the status of 1.g4, if it's lost what you'd do to find the elusive draw would be a waste of time, and if it's drawn then you should be trying to find a high scoring line that can be saved.

[Zenmastur]

Anyone ever do this before or have any suggestions?

I've done it. I have such a backsolved tree for all of chess (which was used to build my chessmap.)

My suggestion is to resist the temptation of quantity over quality, all those 15ply games will provide a lot of noise, and variety isn't a good thing (the only reason you want variety is that there's some mainline you don't want to miss, so playing more lines would catch it; if you'd somehow miss it there's something else to fix before starting this.)

I'm only using 15 ply to vet human games. everything past the first error gets lopped off. I'm also am lopping off any position that isn't seen at least "x" times. Since the database is only 4,200 games the stub-games I'll be left with will be pretty short (i.e. mostly 2, 3, 4, 5, and 6 plies). If there are too many position when I'm done vetting then I'll select the shortest "x" of them and use those.

The I'll play some games of moderate depth (say 21-25 plies). I plan on doing several batches at different depths to get a good selection of continuations.I'll use both one-core and two-core games because two-core is less deterministic. Then I'll extract lines and positions that appear several times. Then repeat at longer time controls. This will take some time, maybe a week or two for the first run depending on how much comp time I give it. Then a last set with with longer time controls and more cores.

What has worked best is doing it the other way around, you build trunks of analysis lines of the highest quality moves that you can find, then you search for alternatives to those moves at decreasing depth. Starting with low depth will only make you realize later how much time you wasted in some lines (say, having 2000 lines for a poor move and 100 for the best move when it should be the other way around.)

I thought of this, but there are few high quality games AND how do you know which are which? Human games mostly suck and can't be trusted. So, I'll start with just short ply openings and build longer ones from engine games. They may not be the strongest games but they will be much more reliable than human games. I do have a bunch of correspondence games but until I get a "basic" book I don't want to mess with them. They are too narrowly focused to make a general purpose book from and there isn't enough of them. After I built a basic book, I can go back to my data base and extract just the high quality games. AND then use the corr. games and some analysis to extend the book lines. There are actually quite a few lines that can be played by black. This makes the tree very bushy and I don't think I can do that much analysis in a reasonable period of time.

Another thing is game adjudication, you shouldn't bother with playing games to completion, depending on how badly scored a line is you'd only play games from it until they reach a certain score, and if they do you look for an alternative line.

Yes, I plan on early adjudication using something like +/-3.5 pawns etc. I'll do some testing on this first, of course.

Unfortunately the optimal way of carrying it out depends on the status of 1.g4, if it's lost what you'd do to find the elusive draw would be a waste of time, and if it's drawn then you should be trying to find a high scoring line that can be saved.

I'm going to assume it's lost. I've seen no evidence to the contrary. This makes making the book a bit easier as I can throw out any game where white has an advantage because to get an advantage black had to screw up.

Regards,

Zenmastur

[Ovyron]

If you assume it is lost then you just need to refute 1 black mainline per each white line and don't need to check the others as this one is sufficient (a black winning book would just need this line.) It seems you're planning on doing a lot of stuff that will be irrelevant in the end (you can do without the human games, you can do without the low depth games, most correspondence games from before today's top engines will prove useless, etc.) but I guess it'll work on giving you the experience about knowing what is unnecessary for books you'd do for other openings on the future.

[mmt]

Anyone ever do this before or have any suggestions?

I made a manual book (based on LC0's and SF's evals) and played something like 80 games of LC0 384x60 vs LC0 62xxx and SF vs SF with half the time. But this was mostly to identify more promising openings to look at deeper with SF or LC0 playing black with long time controls. So for this, I only included 1-3 possible black moves per position. I was planning to finish this and run LC0 vs SF games too. If SF could draw using half the time and LC0 drew also with 62xxx and 384x60 at different time controls, that would seem like a promising opening. You could do it like I planned - start at short time controls and double them for the next tournament. It will be interesting to compare results.

[mmt]

Another thing is game adjudication, you shouldn't bother with playing games to completion, depending on how badly scored a line is you'd only play games from it until they reach a certain score, and if they do you look for an alternative line.

Yes, I did this also. I set it to something like resign if <-4.0.

[Zenmastur]

If you assume it is lost then you just need to refute 1 black mainline per each white line and don't need to check the others as this one is sufficient (a black winning book would just need this line.) It seems you're planning on doing a lot of stuff that will be irrelevant in the end (you can do without the human games, you can do without the low depth games, most correspondence games from before today's top engines will prove useless, etc.) but I guess it'll work on giving you the experience about knowing what is unnecessary for books you'd do for other openings on the future.

Well... When I'm done I would like a full book. Or at least as full as I can make it with the time I can spend. I want human games because humans have a way of finding all kinds of strange lines. Most are crap, but with a little engine help, maybe something can come of them. What I don't want is a book that has these massively long computer lines that have never been played and never will be played. Not that many of these lines will get a lot of use anyway.

Just looking at what I have now it's surprising the number of different lines and the amount of transpositions there are. I was deleting a lot of lines until I realized that other lines were transposing into them. So, by themselves they are bogus BUT, they gain in significance when a lot of different lines transpose in to them.

I'm about a quarter of the way done with my analysis of the games. I want to start games soon so maybe tomorrow I'll create som FENs from what I have. Then I can do a test run of games from FENS that are from positions that are 2 to 4 plies into the opening.