|
< Earlier Kibitzing · PAGE 11 OF 15 ·
Later Kibitzing> |
| Dec-07-17 | | zanzibar: <<frogbert - quoting <Dom> to start> < So the French, Sicilian, Caro-Kann and Spanish are all now defunct>
...
Seriously, we can't tell from the diagrams in the paper why AlphaZero stopped playing certain openings during learning; was it based on white's or black's results in the given opening? Or neither? :)> Don't forget, like I did, to mention the KID being busted too, pretty much right out of the starting gate (although we already knew that!). But yes, why are some openings intensely played and then fall "out of fashion" during training? Here's a hypothetical - suppose one opening always scored draws, and another scored 50% wins and 50% draws. Would both be treated the same during training?
. |
|
| Dec-07-17 | | WorstPlayerEver: Oh yeah living in the world of telecommunication it probably takes a pigeon a year to spread the rest of the 90 games. As has been claimed in public for a while. Or -option2: now things get extremely boring- there are no 90 other games Or -option3- Bob's my mother. |
|
| Dec-07-17 | | zanzibar: (BTW- I share many/most/all(?) of <frogbert>'s concerns - e.g. the fact that SF was tuned with opening book/endgame tb enabled) |
|
Dec-07-17
 | | WannaBe: <frogbert> The answer my friend, is blowing in the wind. Well, OK, it's not in the wind, but a quick search or looking at London Classic... No French, No King's Gambit, No Bird, No Petrov, No Dutch, No Scotch. Of Carlsen's game in database, 8 games have 1...e6 Nakamura have 44 So, does an opening fall out of favour? Or is it top GMs think it's not optimal? It doesn't mean it's not playable to lower level, it's (almost) like no NFL team uses wishbone formation, but some colleges still do. |
|
| Dec-07-17 | | Absentee: <Domdaniel: So the French, Sicilian, Caro-Kann and Spanish are all now defunct? I'm not too surprised, apart from the inclusion of the Ruy Lopez/ Spanish on the list. I thought the Berlin wall was still a draw. If AlphaZero is right, it isn't, which means that the Petroff is the best drawing line after 1.e4 e5.> It hasn't solved chess, only beaten another engine playing certain openings under very specific conditions. For the question of whether those openings are a win, a draw or a loss, we're every bit as clueless as before. Restricting the time controls to 1 minute/move might be a very limiting factor for Stockfish, or any traditional engine. AlphaZero seems to use its own self-played games more or less the way we use pattern recognition, although the paper doesn't quite explain how this knowledge is stored and accessed (does it have perfect recall of every game? does it produce its own variables and adjust the evaluation function at specific moments? etc). With a relatively low number of positions analyzed per second, I would expect this approach to hit a ceiling sooner, and longer time controls to favor brute-force calculation. I also might be entirely wrong. |
|
| Dec-07-17 | | frogbert: <The latter is an obvious blunder which S8 not even considers. It plays 21. Bg5 f5 22. Qf4 hxg5>
And what does S8 do for black after 23. Nxg5, <WorstPlayerEver>? 23... Qxh5 loses to 24. g4!
23... Qg8 loses to 24. h6!
Hence, we're left with the line 23... Qh6 24. Re8 Qf6 25. Rae1 Kg8 26. Rxf8+ Qxf8 27. Qd4! and white is again winning. Got any improvements for black here? |
|
| Dec-07-17 | | frogbert: (For some reason Stockfish has an issue with finding 24. g4! too - I don't know why.) In short, it certainly doesn't look like 22... Nc5 was a blunder - rather black's best option in the position. But Stockfish realized too late that it was in trouble after Bg5! - possibly due to the enforced time restriction of exactly 1 minute per move. |
|
Dec-07-17
| | WannaBe: I have only (quickly) played game 1 from the PDF file, so a quick glance at FEN of game 5 - move 20, I see similarity to game 1: 1. White have no center pawn(s)
2. (Somehow) black got its pieces in the A8 corner(?!!?) Don't think any human player would get him/herself into that kind of position... |
|
| Dec-07-17 | | WorstPlayerEver: 21. Bg5 f5 22. Qf4 hxg5 23. Nxg5 Qxh5 24. g4 and Black is completely lost. Again. And I see no improvement. But it proves my point; S8 does not play 21... Nc5 afaik And a lot of other moves. I assume that AlphaZero actually played these moves and it all looks very interesting and brilliant indeed. I am the last to deny. There's really something interesting going on. However, I question the presentation of the facts. Because it's possible to filter out these variations and apply a different approach to them while programming. A testcase. So to speak. Besides, if I were so brilliant, I'd be the first to come up with all info=100 games and not 10. Why hesitate? So I guess we are repeating moves ;) |
|
| Dec-07-17 | | frogbert: In game 9, AlphaZero plays a rather brilliant positional piece sacrifice, with 30. Bxg6!! It takes Stockfish several moves after the sac to realize that it has become (much) worse - even a piece up. There's a similar theme in game 7 where black's bishop ends up being utterly useless on b7 and a8, making black practically play a piece down in the deciding moments of the game. |
|
| Dec-07-17 | | frogbert: <But it proves my point> It proves nothing of the sort: There are various builds of S8 around, your cache-state may have an impact, there are lots of tweakable engine parameters, and its performance (and choices) also depends on the hardware configuration. By playing around, I've seen that Stockfish seems to struggle with finding some of the refutations of its own moves - while AlphaZero did. Only after I help Stockfish with the right idea, does it change its opinion. My only "beef" with the comparison in the article, is that they imposed restrictions on Stockfish that are particularly harmful for depth-first approaches; in order to do selective depth-first well, the program needs to be able to manage its own time (within normal time controls, of course). |
|
| Dec-07-17 | | WorstPlayerEver: <frogbert>
Sure enough S8 has trouble to find this moves: it goes beyond their evaluation. Which is traditionally still too based on materialistic grounds. It doesn't recognize the pattern of the position. Which comes down to King safety. My objection stands corrected though, I think. The Black S8 positions are often more than just passive. Withdrawn is an appropriate term, I suppose. The e5 vs c6/e6 ingredient/locked Black LSB/king side attack, White Queen has hands free/open d-file for White. All managed through nice pawn sacs.
It's almost like sac sac mate ;) |
|
| Dec-07-17 | | john barleycorn: I have a strong feeling that the whole thing was manufactured in one way or the other. Has Kasparov been involved? |
|
| Dec-07-17 | | frogbert: <Hence, we're left with the line 23... Qh6 24. Re8 Qf6 25. Rae1 Kg8 26. Rxf8+ Qxf8 27. Qd4! and white is again winning.> And here, my Stockfish doesn't find the important move 27. Qd4 within one minute - and Qd4 actually works due to a second piece sacrifice (Bd5+). :) I hope everyone realizes that AlphaZero also does a tree search during play - MCTS or Monte Carlo Tree Search simulations - which is more costly for each node in the tree; hence, the huge difference in number of positions examined per second between AZ and S. |
|
| Dec-07-17 | | Marmot PFL: Komodo seemed listless and out of form in some games against Houdini too. Even computers have bad days I suppose, and the programmers go through the opponents games, looking for weaknesses to exploit. |
|
| Dec-07-17 | | WorstPlayerEver: <frogbert>
You are right. I proved very little today.
But hey, except from rather cramped opening ideas, I always stated there's -far-
more to chess than everyone thought. My infectuous optimism may well have contributed to this h.. I mean revolutionary developments. Who knows. Or maybe I am a little on the narcissistic side? |
|
Dec-07-17
 | | Domdaniel: <Worst> That's good. It's only when you reach the *very* narcissistic point, where optimism is quashed, that things get problematic. |
|
Dec-08-17
 | | keypusher: < WannaBe: I have only (quickly) played game 1 from the PDF file, so a quick glance at FEN of game 5 - move 20, I see similarity to game 1:
1. White have no center pawn(s)
2. (Somehow) black got its pieces in the A8 corner(?!!?)Don't think any human player would get him/herself into that kind of position...> Kasparov vs Marjanovic, 1980 |
|
| Jan-10-18 | | scholes: Current sf dev version now 50 elo stronger than sf8. Halfway to alphazero ☺️ http://www.sp-cc.de |
|
| Jan-26-18 | | chesslab0: Dear chessfriends, the 19th (already) FICGS correspondence chess WCH waiting list is now open, all tournaments should start on March 1st, 2018... Stockfish and Houdini are kind of required here :) ... and yes, after all these years GM Eros Riccio is still FICGS correspondence chess champion! If you're IM or GM (basically rated over 2350) and used to chess engines, you may get the opportunity to enter the knockout tournament to reach the candidates final. We really need the participation of the strongest corr. chess players to shake the crown, join the fun :) http://www.ficgs.com/waiting_list_f... You can replay the games of the 14th WCH final match here: http://www.ficgs.com/tournament_FIC... By the way, for more convenience FICGS now has applications for Android that you can find here: https://play.google.com/store/apps/... |
|
| Feb-01-18 | | scholes: Stockfish 9 released. 60 elo stronger than sf8 in self test. Would be more in rating list due to new contempt in sf9. Stockfish has 20 as default contempt. Use 0 contempt for analysis unless you what it does. |
|
| Feb-03-18 | | zanzibar: SF download page:
https://stockfishchess.org/download/
Apparently an early release got out, somewhat mistakenly: https://www.reddit.com/r/chess/comm... (Thanks <scholes> for that tip - but aren't the parameter settings the responsibility of whatever UCI interface one is using?) |
|
| Feb-03-18 | | zanzibar: Oh, wait, the engines can be run stand-alone too, though I rarely do that. Was the contempt setting then changed between v8 and v9? |
|
| Feb-03-18 | | scholes: Yes in jan 2018 it was changed from 0 to 20. Largest contempt value which doesn't causes regression. No you a chess hui does not change uci parameters by default. Only a user can change it. |
|
| Feb-03-18 | | zanzibar: <scholes> I run SCID, and it's gui has a UCI configure engine dialogue box. When I run it with SF9 it comes up with Contempt = 0. Is there any method for testing what Contempt setting an engine really is running with other than looking at the UCI configure window? PS- I disagree with the statement that only a user can change it - as the gui's can store settings and send them to the engine at startup. I haven't looked at the details of the code - so I don't exactly who does what, when. |
|
 |
|
|
< Earlier Kibitzing · PAGE 11 OF 15 ·
Later Kibitzing> |
Kibitzing
