TCEC 14 final: Lc0 just lost 3 games with tactical..

[Leo]

There is something disappointing about Houdini being ahead in the TCEC against a struggling LZ unless Houdni was improved recently.

I know only that Houdini could beat stockfish in the TCEC cup.

I do not know about Houdini being ahead against LZ.
LZ is in the superfinal against stockfish and not Houdini

I have dial up internet and cant see the TCEC. I thought the Cup was the superfinal.

[yanquis1972]

fwiw, i find t40 passing all of these (these being the 3 starkest blunders i've found so far) & avoiding a couple other, less-losing moves i've tested, but have only analyzed the Kf1 blunder at several MNs.

1r1q1rk1/p1pbbp2/7p/8/2PPnP2/1P6/PB2B2p/RN1Q1R1K w - - 0 20 31 [bm Qd3 (??)]

20. Rf3?? Bg4!

br4r1/4k3/B5qp/1PppNp2/4nP2/P7/2Q1R1PP/5RK1 b - - 0 31 [bm Qb6 (?)]

31...Qd6?? 32. (a4, Rd1 [played])

6k1/b3q3/Q3b2p/p2p1p2/4n3/3B1N2/PP3PPP/2rNR1K1 w - - 6 34 [am Kf1]

34. Kf1?? Bxf2+!

at least the 1st position may be lost already & i've only done shallow analysis

Try these.

6k1/5p2/5Pp1/6P1/2nPK3/2B5/p7/8 w - - bm Kd3; id "Leela Goof 1"
4k1R1/p7/8/5R2/5N2/KP3qP1/P4r1p/8 b - - bm Ke7; am Kd7; id "Leela Goof 2"
8/2k5/rn4R1/1pp5/2p1n1B1/p1P3P1/4KPP1/2R5 b - - bm a6a8; id "Leela Goof 3"
5q2/1b2r2k/n1p3p1/1pR3P1/3P1p2/2N1pP1p/6P1/RQ4K1 w - - bm Re5; id "Leela Goof 4"
8/5pkp/1r6/5P2/2n5/p1pR3P/qbQ2PP1/3R2K1 b - - bm Qb3; id "Leela Goof 5"
r4r2/5pk1/2nR4/4p1p1/1p6/N1P5/PP3PRP/4K3 b - - am bxa3; id "Leela Goof 6"
6k1/b3q3/Q3b2p/p2p1p2/4n3/3B1N2/PP3PPP/2rNR1K1 w - - 6 34 bm d3e2; id "Leela Goof 7"
r1bqk3/2p1ppb1/1n1p2p1/p6r/1p1PP2P/4BP2/PPPQ4/1K1R1BNR b q h3 0 14

thanks! are these all from tcec positions? i'm crosspolluting w/ this thread, fwiw: https://groups.google.com/forum/#!topic ... QftH_0_G0Y

3 more, all of which i have t40 failing quite badly on initial search (it plays Nc5 but doesn't see the position as losing even after 25. Bh6, & sees Rd6 as losing after its played (ie not nearly soon enough)

r2q1rk1/p2p2p1/n7/1p5R/2p1Q3/6P1/PP3bKP/R1B5 b - - 3 23

23...Bd4?? 24. Qh7+ Kf7 25. Bh6! winning [23...Nc5]

1r6/5pkp/6p1/3Rn3/4P3/p1p1N2P/qbQ2PP1/3R2K1 b - - 1 36

36...Nc4? 37. Nf5+! gxf5 38. exf5 Rb6 39. R5d3

8/5pkp/1r6/5P2/2n5/p1pR3P/qbQ2PP1/3R2K1 b - - 2 39

39...Rd6?? 40. Rg3+ 1-0

[dkappe]

They’re from a variety of sources, including tcec and cccc

[supersharp77]

In particular, tactics are where engines are clearly better than humans. If Leela can not help, that's a bad analysis engine for humans. Therefore, it is important that the Leela developers try to improve here. I hope it!

You are 100% correct.....Leela0 is strictly a positional monster...I have observed some terrible tactics from LC0 (but it plays some closed and semi-open positions) quite well indeed.....

[M ANSARI]

This was bound to happen, but it just seems a lot more than I expected. Usually Lc0 fails with tactics much less often. I think this will be patched up soon and it won't be due to a new network. I think network alone is not sufficient enough to patch up these weird misses as it seems A0 also had some of these tactical misses (although much less).

But for TCEC, you have to realize this is an entertainment show more than actually engine test. The book lines used are intentionally made to create very tactical positions which are not in Lc0 favor. The idea is to reduce the number of draws as much as possible. It is not secret that Lc0 can be weak in some tactics and that tactics is something that needs to be patched up.

By the way, Lc0 also had some pretty incredible wins. Looking at the evaluation of the Blue Fish (176 thread SF) it seems that it would have beaten SF on that massive hardware in those games as well.

[Guenther]

i'm not sure that's true, but i haven't seen 2 of them. i do know that t40 avoids the Kf1 blunder (tested id40743) after around ~200KNs, but i don't know how long it takes to find Bxf2+ after Kf1 with clean search:

[d]6k1/b3q3/Q3b2p/p2p1p2/4n3/3B1N2/PP3PPP/2rNR1K1 w - - 6 34

when i checked, SFs reply to Be2 was d4, not sure if it's clear best move.

...

In the game SF expected 34. Ne3! after 5704M nodes with equality (0.16) at depth 42.
(This move never appears in all of your LC0 analyses)

The beginning of the pv:

Code: Select all

33... Rc1 @ 34. Ne3 Rxe1+ 35. Nxe1 Nxf2 36. Kxf2 f4 37. Kf1 Bxe3 38. Nf3 Qf6 39. Qd6 Kg7 40. Ne5 Bb6 41. Nd7 Bxd7 42. Qxd7+ Kf8 43. Qxd5 Qxb2 44. Qd6+ Kf7 45. Qxf4+ Ke7 46. a4 Qa1+ 47. Ke2 Qa2+ 48. Kd1 Qb3+ 49. Kd2 Qa2+ 50. Bc2 Qd5

[Nay Lin Tun]

If you think Lc0 is terrible in tactics, SF is terrible in strategy too.
The score is extremely close!!

[nabildanial]

fwiw, i find t40 passing all of these (these being the 3 starkest blunders i've found so far) & avoiding a couple other, less-losing moves i've tested, but have only analyzed the Kf1 blunder at several MNs.

1r1q1rk1/p1pbbp2/7p/8/2PPnP2/1P6/PB2B2p/RN1Q1R1K w - - 0 20 31 [bm Qd3 (??)]

20. Rf3?? Bg4!

br4r1/4k3/B5qp/1PppNp2/4nP2/P7/2Q1R1PP/5RK1 b - - 0 31 [bm Qb6 (?)]

31...Qd6?? 32. (a4, Rd1 [played])

6k1/b3q3/Q3b2p/p2p1p2/4n3/3B1N2/PP3PPP/2rNR1K1 w - - 6 34 [am Kf1]

34. Kf1?? Bxf2+!

at least the 1st position may be lost already & i've only done shallow analysis

Try these.

6k1/5p2/5Pp1/6P1/2nPK3/2B5/p7/8 w - - bm Kd3; id "Leela Goof 1"
4k1R1/p7/8/5R2/5N2/KP3qP1/P4r1p/8 b - - bm Ke7; am Kd7; id "Leela Goof 2"
8/2k5/rn4R1/1pp5/2p1n1B1/p1P3P1/4KPP1/2R5 b - - bm a6a8; id "Leela Goof 3"
5q2/1b2r2k/n1p3p1/1pR3P1/3P1p2/2N1pP1p/6P1/RQ4K1 w - - bm Re5; id "Leela Goof 4"
8/5pkp/1r6/5P2/2n5/p1pR3P/qbQ2PP1/3R2K1 b - - bm Qb3; id "Leela Goof 5"
r4r2/5pk1/2nR4/4p1p1/1p6/N1P5/PP3PRP/4K3 b - - am bxa3; id "Leela Goof 6"
6k1/b3q3/Q3b2p/p2p1p2/4n3/3B1N2/PP3PPP/2rNR1K1 w - - 6 34 bm d3e2; id "Leela Goof 7"
r1bqk3/2p1ppb1/1n1p2p1/p6r/1p1PP2P/4BP2/PPPQ4/1K1R1BNR b q h3 0 14

Have anybody tried to check if ScorpioNN with the same T30 net would blunder these positions?

[MikeGL]

fwiw, i find t40 passing all of these (these being the 3 starkest blunders i've found so far) & avoiding a couple other, less-losing moves i've tested, but have only analyzed the Kf1 blunder at several MNs.

1r1q1rk1/p1pbbp2/7p/8/2PPnP2/1P6/PB2B2p/RN1Q1R1K w - - 0 20 31 [bm Qd3 (??)]

20. Rf3?? Bg4!

br4r1/4k3/B5qp/1PppNp2/4nP2/P7/2Q1R1PP/5RK1 b - - 0 31 [bm Qb6 (?)]

31...Qd6?? 32. (a4, Rd1 [played])

6k1/b3q3/Q3b2p/p2p1p2/4n3/3B1N2/PP3PPP/2rNR1K1 w - - 6 34 [am Kf1]

34. Kf1?? Bxf2+!

at least the 1st position may be lost already & i've only done shallow analysis

[d]6k1/b3q3/Q3b2p/p2p1p2/4n3/3B1N2/PP3PPP/2rNR1K1 w - - 6 34
6k1/b3q3/Q3b2p/p2p1p2/4n3/3B1N2/PP3PPP/2rNR1K1 w - - 6 34

Sorry to interrupt but whats the complete line after Kf1?? Bxf2! of black?
Looks like Lc0 (with white) gets 3 minor pieces versus lone Q of SF (black).

3 minors is not bad against a Q and Lc0 thought so too, also endgame position looks like fortressable.
I remember Ljubejevic beating Karpov using 3 minors against Karpovs lone Queen.

[M ANSARI]

I think more important than looking at what move Lc0 will play is to see what it "thinks" in MPV mode. If it sees a move that is second best or third best and is still instantly losing but the evaluation is very neutral, then that is something worrying, even if it chooses the correct move.