|
< Earlier Kibitzing · PAGE 4 OF 15 ·
Later Kibitzing> |
| Sep-13-15 | | Appaz: The first game from Stockfish, against Ginkgo, in stage 2 of TCEC may be relevant to the discussion above. Check it out: http://tcec.chessdom.com/archive.php It's crazy chess from a patzers view point, but of course it won. Some weird looking moves there: the pawn on h4, the knight on h2, the backward c5 pawn, the knight on d3 etc. Stockfish at these ply depths is killing human positional play. |
|
| Sep-13-15 | | SChesshevsky: <Appaz: The first game from Stockfish, against Ginkgo, in stage 2 of TCEC may be relevant to the discussion above.> Thanks for the link. It appears from the analysis board that they had a 2-hour game control. Didn't they roughly use an hour and a half for each of their times? A couple of questions are why should a computer have anywhere near 2 hours and maybe just as important why did they need to use nearly that amount of time? |
|
| Sep-13-15 | | Appaz: The games are played at 2 hours + 30 seconds a move. In stage 2 they use a 2 move opening book. Each engine will meet each other twice, playing the same opening from each side. They play 24/7 and the first round has just finished http://tcec.chessdom.com/live.php The engines are running on some <heavy> hardware, reaching a speed of around 20 million evaluations a second. Stockfish typically reached 35 plies in 3 minutes and 12 seconds on move 3, after having evaluated more than 3.6 billion positions. The reason for the amount of time is just to get the quality up. |
|
| Oct-15-15 | | mrandersson: Right i would love some feed back on this game as i am quite sure white used a engine here. This was a 8+3 min blitz game my grade is 1760 online white was 1380. I had took quite a few months off from playing internet chess as i personal feel most people use a engine not at the start but at some point in the game and my game here i think is a prime case
1. e4 d6 2. d4 Nf6 3. Nc3 g6 4. Nf3 Bg7 5. Be2 O-O 6. O-O c6 7. Qd2 b5 8. e5
dxe5 9. Nxe5 Nfd7 10. Nf3 Nf6 11. a3 Bf5 12. Ne5 Nfd7 13. Ng4 Nb6 14. Rd1 N8d7
15. Nh6+ Bxh6 16. Qxh6 Bxc2 17. Re1 Nf6 18. Bf3 Rc8 19. Bg5 Bf5 20. Re2 Be6 21.
Rae1 Nc4 22. Ne4 Nd6 23. Nxd6 Qxd6 24. Bf4 Qd7 25. Be5 Bf5 26. h3 a6 27. b3 Qd8
28. g4 Bd7 29. h4 c5 30. d5 c4 31. g5 cxb3 32. d6 Re8 33. Bd5 Be6 34. d7 Nh5
35. Bf3 Rc5 36. Bxh5 Rxe5 37. dxe8=Q+ Qxe8 38. Rxe5 a5 39. Bg4 1-0 what dou you guys make of it? |
|
Oct-15-15
 | | OhioChessFan: 13...e5 would have been much more thematic. The right Pawn push at the right time. 19...Qxd4 would have been better. 22....Nd6 too much shuffling of your Knights. As for White, 26. h3 and 27. b3 are pretty suspicious looking. On the other hand, White had a couple chances at moves 32 and 33 to play h5 and didn't, and I expect the engines would have. |
|
| Oct-15-15 | | Mehem: Can you take a look at the recent Carlsen - Karjakin game:
Carlsen vs Karjakin, 2015
Why Stockfish (and the other chess engines) when analyzing position after 42... Qf3 displays only 43.Rg1 with 0.00? However, if played 43.Rg1 it advises immediately 43... Ra1 with +M6. |
|
| Oct-15-15 | | Chuckles: It's because 43.Rg1 is a repetition of the position after 41.Qb2, and engines always evaluate repeated positions as 0.00. This is to allow them to avoid repeating themselves in their search, but it causes this type of strange evaluation in situations like this. |
|
| Nov-12-15 | | thegoodanarchist: I looked through all 4 pages of comments. No explanation on where the name came from for this engine. Is it supposed to be a joke? Anyone know? |
|
| Nov-17-15 | | zanzibar: <tGA> you might have wanted to check the wiki page first: <The program originated from Glaurung, an open source chess engine created by Romstad and first released in 2004. Four years later, Costalba, inspired by the strong open source engine decided to fork the project. He named it Stockfish because it was "produced in Norway and cooked in Italy" (Costalba is an Italian, Romstad is a Norwegian).> Now I'm wondering where Glaurung came from. Ah...
https://en.wikipedia.org/wiki/Glaur... |
|
| Nov-17-15 | | zanzibar: Is anyone else struggling/suffering from Stockfish 6's unstable evals? I find evals going out to great depths are subject to change if I just walk forward a few moves (3-4) and then return to the position, with the evals radically changing even when re-evaluating out to just a few moves less depth. I don't see this when running an engine like Critter, at least not with such frequency. (Yes, some highly tactical positions can be extremely sensitive to depth, and cache loading. But I suspect Stockfish is actually making mistakes by narrowing too much - mistakes likely to mislead operators) C Ionescu vs Smyslov, 1986 Black's move 23. What's the best move?
Specifically, is 23...Rxc2, as played in the game, among the top 3 candidate moves or not? I have it, and 23...Rg8, both being draws (eval 0.00), at 22-ply. Yet, by 14-ply depth, Critter demotes it off the list of candidate moves. And if I do a walk-forward and back, so will Stockfish 6. (And it didn't work out so well for Smyslov either!)
What's going on? How can I trust an engine which re-evaluates so radically? |
|
| Nov-17-15 | | Retireborn: <z> In Informator, Ionescu evaluates the position after 23.Rd1 as won for White. Houdini (version 1.5ax64) doesn't quite agree and gives "big advantage for White". Both Houdini and Fritz 11 immediately identify 23...Rg8 as Black's best move and Fritz even thinks Black is better at first, although it changes its mind after 24.Rd7 Rxc2 (Ionescu only analyzes 24...Qc5+) 25.Rfd1 (not 25.Rxe7 Rxg2+) Qg5 26.g3. I suspect that whatever the engine used, the "walking forward" is necessary more often than you'd expect; only in tablebase positions would I simply trust an immediate eval. |
|
| Nov-17-15 | | Mehem: <zanzibar> My Stockfish-6-64 persistently doesn't subscribe to the Smyslov's choice: 1. 23... Rg8 @25: -0.10 @30: -0.18 @35: -0.21
2. 23... Rf8 @25: -3.45 @30: -3.58 @35: -3.64
3. 23... Rxd1 @25 -3.96 @30: -5.08 @35: -6.01
but when I play the strongest 23... Rg8 evaluated as -0.20 (i.e "equal") then Machine displays 24.Rd7 with +2.00 and indeed the won position for White. Anyway, it is heartening you still can't trust blindly the computers. |
|
| Nov-17-15 | | NeverAgain: If you go back a page or two up the thread you will see that a while ago AylerKupp and I came to the conclusion that d=35 is the lowest depth for trustworthy analysis with SF6. This position requires just a little more than that minimum: Ionescu - Smyslov 1986, 23.Rd1
click for larger viewStockfish 6 64 POPCNT:
 (0.45) Depth: 36/54 00:00:36 335MN, tb=19242 - finds 23...Rg8 24.Rd7  (1.62) Depth: 36/63 00:00:44 449MN, tb=32705 - evaluates it as a win for White This position is nothing special, there are dozens and dozens others that SF and most other engines can't evaluate properly at any depth. Go back to my earlier posts in this thread and you'll find the links to examples. As for this case, sure, d=36 is a bit high and may take a while on older hardware; but why are you running SF6 in the first place? Go to http://abrok.eu/stockfish/ and grab the last non-Lazy SMP dev build (dated October 15. although the internal stamp is 161015): Stockfish 161015 64 POPCNT:
 (-1.06) Depth: 12/23 00:00:00 289kN
 (1.34) Depth: 26/42 00:00:13 33681kN, tb=3207 Stay away from later (Lazy SMP) builds unless you're running on 20+ cores: Stockfish 161115 64 POPCNT:
(0.44) Depth: 36/63 00:02:57 2875MN, tb=728786 (1.90) Depth: 36/68 00:06:28 6558MN, tb=1954468 For comparison:
Deep Rybka 4 SSE42 x64:
23...Rg8 24.Rd7
(1.39) Depth: 15 00:00:43 29993kN, tb=1
(1.44) Depth: 25 00:20:54 1051MN, tb=37
Houdini 4 Pro x64 B:
= (0.07) Depth: 20/51 00:00:10 23829kN
(1.40) Depth: 31/92 00:13:55 17607MN, tb=63388 Gull 3 x64:
= (0.08) Depth: 17/51 00:00:02 7278kN
(1.62) Depth: 24/67 00:01:23 1123MN
Komodo 9.2 64-bit:
(0.72) Depth: 33 00:02:19 2080MN, tb=197637 (1.40) Depth: 43 00:17:25 16571MN, tb=5671631 As you see, SF is not doing so bad with this position - only Gull 3 outperforms it. Gull may actually be the best engine for analysis, even though its development has been frozen for a year and it doesn't support tablebases (however, it can handle endgames pretty well all the same). All the other engines (including the latest Komodo) fail to see that White is winning after 24.Rd1. The bottom line is: never trust any engine analysis 100%. Always have a close look and ask yourself "does this make sense to me?" |
|
| Nov-17-15 | | zanzibar: Thanks all. It will take me a bit to properly go through all the above. I appreciate all the comments. Does anybody have a collection of cases like this? As NeverAgain mentions, there certainly must be more than one. |
|
| Nov-17-15 | | NeverAgain: There are a few:
"My trusty engine tells me Chess Truth" thread on ChessPub:
http://www.chesspub.com/cgi-bin/che... "Hard talkchess 2016 beta 1" thread on TalkChess:
http://talkchess.com/forum/viewtopi...
"Difficult for stockfish?" thread on RybkaForum:
http://rybkaforum.net/cgi-bin/rybka... "A position that needs a physician" thread on TalkChess:
http://talkchess.com/forum/viewtopi...
My current favorite is the position after 23...Rc6 in Y Gusev vs E Auerbach, 1946:
click for larger view
24.Qxe5!! fxe5 25.Rf1!
Stockfish 161015 d=46, Komodo 9.1 d=52, Gull 3 d=32 and Houdini 4 (in regular mode) d=36 all think the position is = (0.00) and want to play <24.Qa3 Rxe6 25.Ra7>. Top engines rated 3000+ cannot find a move made by an obscure player who was still five years away from his master title. |
|
| Nov-18-15 | | zanzibar: Thanks <NeverAgain>, I've bookmarked your post for later reference. Just curious, when you run an engine on a given position, how do you quote the eval? By that, I mean, do you run for a fixed time, or a fixed depth, or do you watch the MPV's and wait for them to stabilize? Also, you sometime quote seldepth. Do you ever run SCID? If so, how do I get that number? BTW- here's a post I can across which seems to identify a similar problem on the Stockfish discussion boards: http://support.stockfishchess.org/d... |
|
| Nov-18-15 | | NeverAgain: I usually run single-PV in Infinite Analysis mode with the Let's Check window open for reference. Typically I wait until at least d=35; more for endames. I never ran SciD but in Arena the seldepth display option for analysis is under UCI options for the engines, so you should look for something similar in SciD's settings. Note that Komodo apparently doesn't support showing seldepth. The post you linked to covers the same issue as the more recent "A position that needs a physician" thread linked in my previous post. The extensive analysis I did on the Gusev - Auerbach game in the past two days (also linked above) showed that the issue remains in recent dev builds and just how ridiculous this issue is. From what I read on TalkChess and on SF support forums my hunch is that the devs don't care about things like these. As long as there's no significant ELO regression they won't bother. My impression is that they (and most of the fandom) don't really care about chess per se, only the epeen size of their engine compared to that of others. |
|
| Jan-08-16 | | scholes: Stockfish 7 has been released. It is rated 58 elo higher than stockfish 6 on ipon rating list. stockfish 6 was released in Jan 2015. If engines keep improving at this rate then after 4-5 years we will only have sea of red moves in chessbomb. |
|
Jan-08-16
 | | WannaBe: < scholes: Stockfish 7 has been released. It is rated 58 elo higher than stockfish 6 on ipon rating list. > Greaaaaaat, instead of me resigning after move 14, I can now resign after move 12. |
|
Jan-08-16
 | | AylerKupp: <NeverAgain> Correct me if I'm wrong, but I thought that the value listed for <seldepth> is the highest search depth reached at that particular ply due to quiescent search extensions, etc., not necessarily the highest search depth reached in that particular line. If that's the case then the number is pretty much meaningless as far as I'm concerned. But elsewhere I've read that it represents the highest search depth reached for the PV, in which case it is possibly meaningful for the PV, but not for the other lines if MPV>1 is specified. And I've also read that "search" is interpreted different by Rybka among other engines. Do you know whether any of these <seldepth>s is correct? BTW, I revisited Y Gusev vs E Auerbach, 1946 and your comments were right on. I was obviously too lazy and/or lacked the time and/or motivation to do any forward sliding to validate the engine analyses I did. My bad. So I'm going to try it again and see what difference 2 years of engine development make. Maybe Stockfish 7 can do better in either identifying 24.Qxe5 as a win for White or find a better defense for Black. So far I'm just letting Stockfish run starting with the position after 23...Rc6. I started by setting MPV=12 and at d=8 (the lowest ply for which I have Arena report results) 24.Qxe5 was ranked as the 10th best move with an eval of [-5.35]. I am up to d=32 and 24.Qxe5 is ranked as the 7th best move with an eval of [-1.68]. So at least Stockfish is heading in the right direction. I'm going to let it run overnight to see whether it can both improve on 24.Qxe5's move ranking and its evaluation. But my computer is fairly slow (32-bits, 4-cores, and 2.66 GHz clock rate) so I don't think that it will be able to get too much deeper by morning. Re your response to <zanzibar>. I also typically run in Infinite Mode with MPV=3 unless I'm on a fishing expedition when I want to see what moves are reasonable, then I run with MPV=5 or higher as needed (like in this case). I usually let the engine analyze until I run out of time or patience, whichever comes first. But I think that running at a low MPV (2 or 3) is a good idea so that you can see how the evaluation of the PV is changing relative to other lines. If after a reasonable depth the first move of the PV is unchanged and the difference between the evaluation of the PV and the second ranked line is increasing, then I think that the analysis is stable and can be stopped with good confidence in the accuracy of the result. But if the first move of the PV is changing and/or the difference between the evaluation of the PV and the second ranked line is decreasing, then I think that the analysis is uncertain and should be allowed to run longer before its results can be trusted (as much as any engine results can be trusted). Of course, in the majority of these situations I run out of patience first. |
|
Jan-10-16
| | AylerKupp: <scholes> Stockfish 7 has been released. It is rated 58 elo higher than stockfish 6 on ipon rating list.> The latest (Jan-09-16) CCRL 40/40 list has Stockfish 7 rated at 3334 and Stockfish 6 at 3304 for a 30-point elo difference. Stockfish 7 still slightly trails Komodo 9.2 (3342) and Komodo 9.3 (3337) as far as rating is concerned but the gap has narrowed considerably and all 3 should probably be considered equal in strength for all practical purposes. And Stockfish 7 is way ahead of either Komodo version in terms of price. :-) The latest (Jan-10-16) CEGT 40/20 rating list has Stockfish 7 rated at 3312 and Stockfish 6 at 3302 for a 10-point elo difference. And Stockfish 7 still slightly trails Komodo 9.2 (3342) and Komodo 9.3 (3325) by a small amount. IPON (Jan-08-16) on the other hand rates Stockfish 7 at 3238 and Komodo 9.2 at 3211 (no rating for Komodo 9.3), a 27-point elo difference in favor of Stockfish. Of course, all the tournaments use different time controls (IPON's is based on a 5 minute blitz game + 3 secs/move extension) and equipment (IPON uses 1 core/engine, CCRL uses 4 cores/engine and CEGT uses 2 cores/engine), so this might affect each engine's performance. But it might be an indication that Stockfish 7 is better at the faster time control and with less resources (cores) and Komodo is better at the slower time controls with more resources, but that's just a guess on my part. And I don't know what to say to <WannaBe> other than he is toast no matter what he does, the only difference being how light or dark the toast is. :-) |
|
Jan-31-16
| | AylerKupp: This link indicating what's new in Stockfish 7 should be of interest to Stockfish users: https://www.chess.com/forum/view/ge... |
|
Mar-02-16
| | AylerKupp: This link should be of interest to anyone that visits this page and in particular <SChesshevsky>: http://rin.io/chess-engine/ The title of the article says it all: "How Stockfish Works". Unfortunately the author did not indicate which Stockfish version he examined and there is no mention of the date of the article so I suspect that the information might be dated, perhaps grossly so given the rapid progress on Stockfish, but interesting nevertheless. |
|
| Mar-02-16 | | Appaz: Thanks <AylerKupp>, a very interesting article - "How Stockfish Works"! |
|
| Mar-02-16 | | tbentley: The author mentions Stockfish 2.3.1, which was released in September 2012. |
|
 |
|
|
< Earlier Kibitzing · PAGE 4 OF 15 ·
Later Kibitzing> |
Kibitzing
