|
< Earlier Kibitzing · PAGE 6 OF 15 ·
Later Kibitzing> |
| Aug-11-16 | | zanzibar: Some explorations of the effect of cache on Stockfish 7's evals: https://zanchess.wordpress.com/2016... I've noticed this very odd behavior often before when doing analysis with the engine. I think the above gives a pretty good idea of what the problem is. |
|
| Aug-16-16 | | WorstPlayerEver: @zanzibar
I wonder where your thoughts are based on, because after a few secs SF7 gives (20 ply) -4.13 for 35... Nf4. Galaxy Tab 3, Android. And really, since my 'comeback' I had to fight hard to reach 1400 lately, though I immediately noticed 35... Nf4 is winning. I puzzle a lot but that can't be the trouble here. Must be some personal system bug. |
|
| Aug-16-16 | | zanzibar: @worst - <Must be some personal system bug.> No, I'm running the off-the-shelf downloadable Windows executable from here: https://stockfishchess.org/download/
https://stockfish.s3.amazonaws.com/... Fairly generic setup, just a laptop, Intel i5 multi-core running W7. What's different is my steering program. I can issue a clear-cache (<ucinewgame>) whenever I wish before giving the engine a position. To clear, or not to clear, that is the question.
I don't think you can do that with the normal GUI's used to steer an engine, e.g. SCID. I suspect the cache doesn't need to be cleared when playing a game forward, or jumping to a specific position even. My suspicion is that the cache goes stale when playing a game backwards, or jumping around within a game. So, try jumping to the end of the game and moving backwards to the position vs. playing the game forward. Are the results different?
(I generally set MPV = 3. Also, in my steering program I can try to limit the depth of the search by issuing a stop.) |
|
| Aug-16-16 | | WorstPlayerEver: @zanzibar
I'm running an older version, I suppose. Sometimes it hangs while using Analyse This (free version). Scrolling back a few moves back and fro usually fixes that inconvenience. What GUI do you use? |
|
| Aug-16-16 | | zanzibar: <Worst> SCID (4.4) almost always with Stockfish (7). (Though I have a few other engines and GUI's). |
|
| Aug-16-16 | | WorstPlayerEver: @zanzibar
Didn't know about SCID. Hmm.. last update seems to be from March/April 2013. I'll check it out. I use Lucas chess for PC. Not so happy with it. Don't like Arena also. Planning on buying a DGT set; comes with Fritz. Quite expensive though. https://www.chess.com/groups/forumv...
https://sourceforge.net/projects/sc... |
|
| Aug-16-16 | | zanzibar: <WPE> Try out SCID - it's has the best design and functionality in my opinion. I rearrange the startup to make mine look like the 4.3 startup - I prefer to have more screen space. The tabs are easy to navigate through. I use 4.4 rather than 4.5 because they mucked up the crosstable views. I need the full set of options and 4.5 yanked a lot out for some dumb reason. SCID 4.5 also opens a new Game List for each DB as you select them - which has some virtues, but basically just ends up confusing me as I move between DB's (something my work leads me to do all the time). On the other hand, being able to sort the GameList window via the column headers is a great feature, only available on 4.5. Note- some people will recommend SCID vs PC, just ignore this advice. The basic SCID is what you want. (Actually, I've tried both, and you might like to as well. But I still like SCID best.) |
|
| Aug-16-16 | | zanzibar: BTW- SCID has a steep learning curve, as do all the GUI's. But it's well worth the effort. (E.g. Tree window, Search by position, Search by pattern, xtabs, Tournament Finder, Player List, Twin Deletion and Annotate are all features I use daily. Annotate can be used for blunder checking, but beware -- Annotate's <Mark tactical exercise> must be unchecked for it to work. Otherwise it hangs due to a bug.) |
|
| Aug-17-16 | | WorstPlayerEver: <zanzibar>
Thanks for the info! Certainly will give it a shot. |
|
Aug-17-16
 | | AylerKupp: <<zanzibar> Some explorations of the effect of cache on Stockfish 7's evals> That discussions doesn't make much sense to me. All that a cache (or hash table or transposition table) does is store previous positions and their evaluations to that, if the same position is reached by transposition that was reached earlier while investigating a different branch of the search tree, the position does not have to be re-evaluated, thus saving some time. And, as far as I know, all caches use the LRU (least recently used) principle; if space is needed in the cache to hold the most recent position and evaluation, the earliest position and evaluation is discarded and the freed space is used to store the latest one. So, yes, if an analysis is conducted and stopped, the cache will contain the N most recently evaluated positions, or "forward populated" as you said. If a new analysis is started and the cache is cleared ahead of time (or even if it isn't), the cache will still contain the N most recently evaluated positions. And, until the cache is full, these will be the first N positions evaluated, or "backward populated" as you said. But unless there are some very specific reasons to do so, there is no "optimization" done with regard to forward or backward population, the algorithm is the same. As why you might get different results if the cache is "forward populated" vs. "backward populated" it's probably because of the different timing. Chess engines, particularly multi-core ones (which they all are these days) are non-deterministic; if you run an analysis of the same position on the same computer, with the same engine, and to the same depth, you will get different evaluations and move rankings. Not MAY, WILL. Guaranteed. So if a position is already in the cache then the overall timing of the search tree pruning will be different and different branches of the search tree will likely be pruned, resulting in a different final position in that branch, and hence a different evaluation of the same initial move on subsequent analyses. Komodo used to have a mode that guaranteed consistent evaluations but this resulted in a slowdown of the engine so it was not very popular and has since been dropped. I don't know why you say that Stockfish 7 is designed to beat other engines in an actual game and is not optimized for analysis. That would make sense to me since chess engines, whether commercial or non-commercial, only get brownie points by beating other engines. Some engines actually have separate executable for analysis mode and game mode (their names escape me at the moment) but they are a very rare exception. As far as to whether to clear or not to clear the cache before giving it a position, that depends on whether you are planning to do some backward sliding. If you are, don't clear the cache before you start the backward sliding and use the largest cache that you can that does not cause disk thrashing (excessive page swapping). That maximizes the chances that the positions you investigate during the backward sliding are still in the cache, thus speeding up the process considerably. |
|
| Aug-17-16 | | zanzibar: <AK> the question remains why are the evals so radically different for the same position, as shown in the blog? The engine missed 35...Nf4 despite claiming a 22-ply depth search. <Why is that?>
I admit, after re-investigating a little, that it could just be a "bug" in Stockfish. If so, then any explanation offered is bound to be misguided. 
click for larger view
2q3k1 /Q4p1p/1p1r2p1/3n4/P7/6P1/5P1K/B2R4 b - - 0 35 Finding 35...Nf4 shouldn't be that challenging to an engine. The tactic isn't that deep, and the various threatened M-2's constrain the action. |
|
Aug-18-16
| | AylerKupp: <zanzibar> the question remains why are the evals so radically different for the same position, as shown in the blog?> I don't really know but I don't think that it's a bug in Stockfish 7. I thought that perhaps d=22 was too low since Stockfish usually needs to be run to higher search depths than other engines to have an equal level of confidence in its evaluations (I typically run it to minimum d=35 and try to run it to d>40). But I ran an analysis of the same position and Stockfish 7 found 34...Nf4 at d=10 with an initial eval of [-4.50]. So d=22 should have been more than enough to find 34...Nf4 in this position. And Gull 3.0 found 34...Nf4 at d=8 (the lowest value that I have my GUI (arena) display evaluations) with an initial eval of [-3.07] while Critter 1.6a also found 34...Nf4 at d=8 with an initial eval of [-3.12]. Another possibility (pure conjecture) is that <interlistchess> did not have Stockfish's 'Skill Level' parameter set properly. It's maximum value is 20 at which level it plays its strongest chess. The lowest value is 0, at which level it "will make dumb moves" according to the documentation. But the default value is Skill Level = 20, so I don't know why (and if) <interlistchess> would have changed it. So this conjecture doesn't make much sense even to me. Since I have your attention I would like to ask you about your new filter program. I have been considering writing a similar program, also in Python, that would automate various engine analyses. Could you provide me with some more details in either your forum or mine? Thanks. |
|
| Aug-18-16 | | zanzibar: I know you generally run at insane levels of depth(*) - but in this case there's a clear tactical. And, as you point out, that should be identified at much smaller depth. I can only figure that the position has been stored in the cache with the wrong eval - which causes it to be pruned off as a loss of the knight without compensation. I would characterize missing it as a "bug". I mean, what else to call it? It should show up at 16-ply, let alone 22-ply.
<Another possibility (pure conjecture) is that <interlistchess> did not have Stockfish's 'Skill Level' parameter set properly.> For the record, <interlistchess> on wordpress = <interlist> on chesstempo = <zanzibar> on CG. Basically the same guy on all. But as for the options, the only non-default option used during the run was MPV = 3. As I said, off-the-shelf generic.
As I said before, I should post the tacticals I found with the filter (after using the "ucinewgame" for each position). The engine/filter did a fairly good job given that it was patched together in about a week. (*) How long does it take the engine to reach the depths you're using, on average? Isn't it far longer than the average time/move for a 40-move/2hr game? Why does the engine need more time to "see" a position than what it takes to beat any human GM? |
|
| Aug-18-16 | | zanzibar: <AK> I left a note in your forum... as suggested. |
|
| Nov-01-16 | | scholes: Stockfish 8 releaased. Nearly 80 elo better than SF 7 https://groups.google.com/forum/?fr... |
|
| Nov-01-16 | | WorstPlayerEver: Happy birthday! |
|
| Nov-02-16 | | WorstPlayerEver: New version is ready to go! |
|
Nov-23-16
 | | Ron: Stockfish 7 evaluates the position below as +.85 for White. White is up two pawns but there's opposite colored bishops. This is really a draw.
click for larger viewThe positions arose from a game. I set up the opening as Queen's Indian Defense, then the rest of White's move were a combination of Stockfish's and mine, and Black's moves were solely Stockfish. This is the game: 1. d4 Nf6 2. Nf3 e6 3. c4 b6 4. Bg5 Bb4+ 5. Nbd2 Bb7 6. e3 Be7 7. Bd3 c5 8. O‑O O‑O 9. Re1 cxd4 10. Nxd4 Nc6 11. Nxc6 Bxc6 12. Nf3 h6 13. Bh4 Bb7 14. Qc2 Bxf3 15. gxf3 Rc8 16. Bg3 b5 17. Qe2 bxc4 18. Bxc4 Qb6 19. Red1 Rc5 20. Ba6 Qc6 21. Rd2 d6 22. Rad1 e5 23. Bh4 g5 24. b4 Rc3 25. Bg3 Rb8 26. b5 Qc5 27. a4 Qb4 28. Rc2 Nd5 29. Kh1 h5 30. h3 Rxc2 31. Qxc2 Nc3 32. Ra1 Kg7 33. Rc1 Nxa4 34. Qf5 Nc5 35. Bxe5+ dxe5 36. Qxe5+ Bf6 37. Qxc5 Qxc5 38. Rxc5 Re8 39. h4 gxh4 40. f4 Re4 41. Rxh5 Rc4 42. Kg2 Rc7 43. Kf3 Be7 44. Rd5 Bb4 45. Rd4 Ba5 46. Rd1 Bb4 47. Rg1+ Kf8 48. Rb1 Be7 49. b6 axb6 50. Rxb6 Rc2 51. Rb5 Kg7 52. Bb7 Rd2 53. Bd5 Rd1 54. Rb7 Kf8 55. Bc4 Rd6 56. Rb5 Kg7 57. Kg4 Rd2 58. Rf5 f6 59. Kf3 Rc2 60. Bb3 Rc5 61. Rxc5 Bxc5 62. Kg4 h3 63. Kxh3 |
|
| Dec-03-16 | | scholes: Stockfish wins TCEC season 9 and now tops all engine rating lists. |
|
| Dec-03-16 | | WorstPlayerEver: S8 almost 3400:
http://www.computerchess.org.uk/ccr...
I'm totally amazed by it, total different beast than S7. |
|
Dec-03-16
| | AylerKupp: I'm not sure if it qualifies (literally) as a totally different beast but a major overhaul to be sure, with every single file modified at least somewhat (1,683 source lines added and 1,628 source lines deleted according to http://blog.stockfishchess.org/). But it certainly is stronger than Stockfish 7. And if you <really> want to get an idea of what was changed (at a high level), you can check this site: https://github.com/official-stockfi.... To give you an idea of Stockfish 8's size, the CLOC counter determined that it consists of 40 files with 8422 source lines of code, 2007 comment lines, and 2722 blank lines. For comparison, Stockfish 7 consists of 41 files with 8386 source lines of code, 2039 comment lines, and 2723 blank lines. From the numbers I would suspect that about 20% of Stockfish 7 was modified to yield Stockfish 8 since I think that a modified source line of code is counted in http://blog.stockfishchess.org/ as one line added and one line deleted. To get a more accurate count you would have to use a diff utility like WinDiff and then you could see exactly what was changed between Stockfish 7 and Stockfish 8. |
|
| Dec-03-16 | | zanzibar: I'm just trying it out tonight. It better have far less of the "bad" behavior I observed Stockfish 7 exhibiting... Only time will tell.
Some probably have more capable rigs than I, so I wonder if your Stockfish 8 also suggests the French Defense as best for Black in answer to 1.e4? |
|
Dec-04-16
| | AylerKupp: <zanzibar> A good question, and the short answer is yes, Stockfish 8 still considers 1...e6 to be Black's best response to 1.e4, at least at d=36. But out of curiosity, and because I was going out to breakfast, I let Stockfish 8 analyze the position after 1.e4 and here were its evaluations, top 5 lines, and the ending position in each of those lines, with opening book OFF: 1. [+0.12]: 1...e6 2.d4 d5 3.Nd2 c5 4.Ngf3 cxd4 5.exd5 Qxd5 6.Bc4 Qd6 7.0-0 (I would say that White has a definite lead in development as compensation for its pawn deficit. Does anyone still think that engines are materialistic?) 7...Nf6 8.Nb3 Nc6 9.Nbxd4 Nxd4 10.Qxd4 Qxd4 11.Nxd4 Bc5 12.Rd1 Bd7 (and now that material is even, so is development) 13.Bf4 Ba4 14.b3 Bd7 15.Nf3 0-0 16.Ne5 Be8 17.Be3 Bxe3 18.fxe3 Bc6 (a strange move, giving up the bishop, accepting an isolated c-pawn, and hindering the pressure against White's Pc2) 19.Nxc6 bxc6 20.Kf2 Rfd8 21.Ke2 Re8 22.g3 Red8
click for larger viewStockfish's last few Black moves have been incomprehensible to me. 2. [+0.21]: 1...e5 2.Nf3 Nc6 3.Bc4 Nf6 4.d3 Bc5 5.0-0 d6 6.c3 0-0 7.Re1 a5 8.h3 Be6 9.Bb5 Ne7 10.d4 Bb6 11.Bd3 Ng6 12.Be3 c6 13.a4 d5 14.Nbd2 exd4 15.Nxd4 Bd7 16.exd5 Nxd5 17.Nc4 Bc7 18.Bd2 Ngf4 19.Bc2 Re8 20.Rxe8+ Qxe8
click for larger viewAn opening line that would not have been out of place in the recently completed WCC match. 3. [+0.21]: 1...c6 2.d4 d5 3.exd5 cxd5 4.Nf3 Nf6 5.Bd3 Nc6 6.c3 Qc7 7.h3 e6 8.0-0 Bd6 9.Re1 0-0 10.Bg5 Nd7 11.Nbd2 b6 12.Bh4 Bb7 13.Nf1 Rae8 14.Rc1 Bf4 15.Bg3 Bxg3 16.Nxg3 h6 17.Bb5 Rb8 18.Qe2 Rbe8 (I am not sure what the maneuver 17...Rb8, 18...Rbe8 was supposed to accomplish other than losing 2 tempi) 19.Bd3 e5 20.dxe5
click for larger viewAnd after 20...Ndxe5 Black has an easy game except for its IQP. 4. [+0.22]: 1...c5 2.Nf3 Nc6 3.Nc3 e5 4.Bc4 Be7 5.d3 Nf6 6.0-0 d6 7.Re1 0-0 8.a4 Be6 9.Bg5 a6 10.Bxf6 Bxf6 11.Nd5 g6 12.Ne3 (I guess Stockfish didn't evaluate too highly the Carlsen-like sequence 12.Nxf6+ Qxf6 13.Bxe6 fxe6) 12...Bd7 13.h3 Rb8 14.Bd5 Nd4 15.a5 Nxf3+ 16.Qxf3 Bg5 17.Nc4 Be6 18.Bxe6 fxe6 (but now it does!) 19.Qg3 Qe7 20.c3 Kg7
click for larger viewI prefer Black's position here, B vs. N plus pressure along the f-file. 5. [+0.40]: 1...h6 (rally?) 2.d4 e6 3.Nc3 d5 4.Bd3 Nc6 5.Nf3 Nb4 6.Ne5 Bd6 7.Bb5+ c6 8.Be2 Ne7 9.a3 Na6 10.Nf3 0-0 11.0-0 c5 12.exd5 exd5 13.Bxa6 bxa6 14.dxc5 Bxc5 15.Bf4 Bb7 16.Na4 Bb6 17.Re1 Re8 18.c3 Bc7 19.Be3 Bd6 20.Nc5 Qc7 21.Nxb7 Qxb7
click for larger viewAfter 22.Qc2 I would say that White's position is substantially better because of its superior pawn structure and it's outpost on d4. But these should be taken with a grain of salt (5. 1...h6?) since what Stockfish 8 considers the top 5 moves change from ply to ply. And no engine, including Stockfish, can search sufficiently deeply in a reasonable length of time following White's first move to reach results with any high degree of confidence. For example, at d=35 Stockfish 8 considered the top 5 moves to be 1...e6, [+0.20]; 1...c5, [+0.25]; 1...d6, p[+0.28]; 1...e5, [+0.30]; and 1...a6 (huh?), [+0.31]. And until d=33 1...Nc6 was typically listed among its top 5 move. And I'm sure that if I had let Stockfish 8 analyze longer, it would have come up with different evaluations and move rankings. |
|
Dec-04-16
| | AylerKupp: Out of further curiosity I let Stockfish 8 analyze the position after 1.d4, also until d=36, and see how it evaluates things. It evaluated Black's best response to be, surprise!, 1...e6. Mais oui! But, again, these evaluations should be taken with a grain of salt (5. 1...a6?). 1. [+0.06]: 1...e6 2.c4 Nf6 3.Nf3 b6 4.a3 Bb7 5.Nc3 d5 6.Bg5 Be7 7.cxd5 Nxd5 8.Bxe7 Qxe7 9.Nxd5 Bxd5 10.e3 0-0 11.Rc1 c5 12.dxc5 bxc5 13.Bb5 Rd8 14.Qe2 Nd7 15.e4 Bb7 16.0-0 Nf6 17.Rc4 Rac8 18.Rfc1 h6 19.h3
click for larger viewFairly balanced. White's pressure against Black's c-pawn is balanced by Black's pressure against White's e-pawn. 2. [+0.09]: 1...d5 2.c4 e6 3.Nc3 Nf6 4.Bg5 Be7 5.e3 0-0 6.Nf3 h6 7.Bxf6 Bxf6 8.Qc2 c6 9.Bd3 b6 10.0-0 Bb7 11.Rad1 (I would think that this rook belongs on c1, but maybe not) 11...Nd7 12.e4 dxe4 13.Bxe4 Qc7 14.Rfe1 Rfd8 15.b4 Be7 16.b5 Rac8 17.Qb3 Nf6 18.bxc6 Bxc6 19.Bxc6 Qxc6
click for larger viewA typical fight against hanging pawns. If White's rook was on c1 then 20.d5 would seem to give White an advantage. 3. [+0.11]: 1...Nf6 2.c4 e6 3.Nf3 b6 4.a3 Bb7 5.Nc3 d5 (this would seem to transpose into line #2 but ...) 6.Bf4 Bd6 7.Bxd6 Qxd6 8.cxd5 exd5 9.e3 0-0 10.Be2 c5 11.0-0 Nc6 12.Rc1 Rad8 13.Bb5 Rfe8 14.dxc5 bxc5 15.b4 cxb4 16.axb4 a6 (if 16...Qxb4 then 17.Qa4 Qxa4 18.Nxa4 Rc8 19.Nc5 and Stockfish evaluates that White has a very slight advantage in spite of its pawn deficit) 17.Bxc6 Bxc6 18.Nd4 Rc8 19.b5 Bxb5 20.Ndxb5 axb5 21.Nxb5
click for larger viewAnd after Black's queen moves White's knight will have a great outpost on d4. Maybe Capablanca or Carlsen can win this endgame but not very many could (or would bother to try). 4. [+0.27]: 1...c6 2.c4 d5 3.cxd5 cxd5 4.Bf4 Nc6 5.e3 Nf6 6.Nc3 a6 7.Bd3 e6 8.a3 Bd6 9.Bxd6 Qxd6 10.f4 b5 11.Nf3 0-0 12.0-0 Bb7 13.Ne5 Qe7 14.Qe1 Rac8 15.Qg3 b4 16.Nxc6 Rxc6 17.axb4 Qxb4 18.Rab1 h6 19.Rfc1 Rfc8
click for larger viewA (very) good bishop against a (very) bad bishop but Black has some q-side counterplay to even things out and White's potential k-side attack is not very far along. I would prefer Black after ...a5 and ...Ba6. 5. [+0.32]: 1...a6 2.e4 e6 (Stockfish 8 seems to love this move early!) 3.Bd3 d5 4.Nc3 Bb4 5.Nge2 (it almost turned into a Winawer French) 5...dxe4 6.Bxe4 Nf6 7.Bf3 0-0 8.0-0 Nc6 9.Bg5 Bxc3 10.bxc3 h6 11.Bxf6 Qxf6 12.Ng3 e5 13.d5 Na5 14.Re1 Bd7 15.Ne4 Qd8 16.Nc5 Re8 17.Nxd7 Qxd7 18.Qe2 Kh8 19.Rad1 b5 20.g3
click for larger viewIf I were Black I would move that Ra8 very soon. |
|
| Dec-04-16 | | zanzibar: <AK> wow, a fair bit of work, and a bit of fun. I wouldn't exactly trust any given line out to the full depth due to horizon effects, but worth a look. I like the 1.d4 e6 facet - either trying for the French or steering into a QID. For the record, I should point out the lack of symmetry after 1...e6 between 1.e4 e6 2.d4 ... (which is White's best opening) Then why not transpose after 1.d4 with
1.d4 e6 2.e4 ... ? |
|
 |
|
|
< Earlier Kibitzing · PAGE 6 OF 15 ·
Later Kibitzing> |
|
|
|
Kibitzing
