< Earlier Kibitzing · PAGE 6 OF 6 ·
|Apr-08-19|| ||frogbert: <Magnus Carlsen, a player with a very strong positional style of play, has now added a highly dynamic edge to his style.>|
Well, he started out as a dynamic, attacking player, a style he refined and adjusted when his opponents became stronger and better at defending. Then his positional understanding and end game skills brought so much success that it became a trademark (and increased the average GM game length by 30-40% in the process).
However, as opponents improved further, also becoming relatively younger (compared to Carlsen), his game seems to have developed - again. The versatility has increased over the past 1-2 years, and the clearest "sign" came when Caruana's 1. e4 was consistently met with 1... c5 during the World Championship match. Carlsen again plays sharper openings and positions, unafraid of sacrificing material for initiative, as seen in games here and in Wijk this year.
That's how I see it, anyway. A simplification, of course, but possibly a more complete picture of his development.
|Apr-08-19|| ||keypusher: <Carlsen has got good, there must be a reason. Conclusion, he is using AlphaZero.>|
Sally, you’ve got to get over your tragic love affair with strawmen.
|Apr-09-19|| ||HeMateMe: <If Black plays ...Nd3 then White can immediately play Bxd3 and remove that horrible piece.>|
White will die a horrible death on the light squares without that LSB on the kingside.
|Apr-09-19|| ||Cedroke: You could say Karjakin got "Sveshnikoved" or perhaps even "Sveshnishoved". I'm so impressed by Carlsens play.|
|Apr-09-19|| ||Jambow: As to <frogbert> and <chancho> we are all seeing the same thing. |
I will note that Magnus's greatest success came as a positional maestro, with his dynamism actually arising from seemingly innocuous positions that even the elite would miss until a fatal wound was inflicted.
It is very unusual for a player to switch from positional play to dynamic play as they get older. Carlsen is still young but is already a multi time defending world champion. I hope he can continue this streak and rather enjoy his dynamic play.
I don't think initially Magnus was this strong after he altered his style a few years back and frankly questioned his wisdom in so doing. I didn't know if he could surpass the elite players in that dynamic style. I stand corrected...
|Apr-09-19|| ||Sally Simpson: ***
The claim that this game was the result of AlphaZero discovering new positional ideas (knights on the rim etc...) and Carlsen has picked up on them deserved such a strawman response.
|Apr-09-19|| ||AylerKupp: <<Sally Simpson> The claim that this game was the result of AlphaZero discovering new positional ideas (knights on the rim etc...) and Carlsen has picked up on them deserved such a strawman response.>|
I didn't read any posted comments that said that AlphaZero discovered any new positional ideas, although some of the more radical AlphaZero fanboys and fangirls might think so.
On the contrary; in "Game Changer" it's main author, GM Matthew Sadler, lists 18 historical parallels which show that AlphaZero emulated some of the best players in history (Kasparov, Tal, Larsen, Capablanca, Botvinnik, etc.) with regards to some of the ideas they had and some of the moves they made. But nowhere in the book is there an example of a new positional idea discovered by AlphaZero, just some very good exploitation of these "historical" ideas. I doubt that they are "new" to Carlsen and they probably would have gone unnoticed if not for the cult following that AlphaZero has achieved.
Try reading the book, I think that you might enjoy it. Sadler does the best job that I can remember in breaking down chess themes and giving examples of them. Just take the presumed superiority of AlphaZero over Stockfish with a grain of salt since no one, except for DeepMind, has provided any data that reflects the overwhelming computational performance advantage that AlphaZero enjoyed over Stockfish in their matches.
|Apr-09-19|| ||Dr Winston OBoogie: Just been through Magnus's last 3 games and this is the best but 33. Bg4?? was appalling, really confusing. |
click for larger view
And that knight on a4 is just useless. Great game to play through though :)
|Apr-09-19|| ||MKD: Why the ECO code for this game is listed as Aoo? Shouldn't it be B33?|
|Apr-09-19|| ||saffuna: When games are presented live they start with A00, as there are no moves, and it sometimes takes a while to make the correction.|
|Apr-09-19|| ||Sally Simpson: ***
This is the not the first time we have had an AlphaZero discussion. I predicted it's arrival back in 2014.
Zurich Chess Challenge (2014) (kibitz #1342)
I'm not anti-computer, I just don't think they have reached their full potential and are being programmed solely to beat other computers. The fact they can hammer humans is a by product.
If we can get them seeing and thinking like we do then they really will enrich the game. At the moment our 1400 players are displaying more imagination than them.
They are on the right path with AlphaZero. (at last).
|Apr-11-19|| ||PawnSac: < Sally Simpson: ***
Was Carlsen playing mind games with Karjakin. Here:
All Karjakin had to do was play 14.Qa4, Carlsen had to play 14...Bd7 it is a three fold rep. >
yea BUT.. was there draw stipulation in this event? No draws under 30 moves, or similar? If so, then yes i would think Carlsen was deliberately using this to force a less than best 3rd time Qa4? and possibly a weakening or risky choice? I looked on the official page and didn't see rules posted, but curious anyway.
|Apr-11-19|| ||PawnSac: well the shortest draw i see in the game list was Grishuk/Giri 22 moves, so if there was a stipulation it would be 20 moves or less presumably. hmmm.|
|Apr-12-19|| ||cormier: |
click for larger view
Analysis by Houdini 4 <d 23 dpa done
1. = / + (-0.27): 30.f3> Nxf3+ 31.Rxf3 Bxf3 32.Rf1 Bf6 33.Qd2 c3 34.bxc3 bxc3 35.Nxc3 Rfc8 36.Rxf3 Bxc3 37.Rxf5 Bxd2 38.Bxd2 Rc2 39.Ba5 Rxe2 40.d6 Rxa2 41.d7 Ra1+ 42.Kg2 Rd1 43.d8Q+ Raxd8 44.Bxd8 Rxd8 45.Ra5 Rd6 46.Kf3 Rb6 47.Ke4 f6 48.Kf4 Kh7 49.Ra2 Kg6 50.Ra3 Kf7 51.Ra4 Ke6 52.Ra3
2. - / + (-0.84): 30.f4 Qg6 31.Bf2 Nd3 32.Bh5 Qxh5 33.Qxe4 Rae8 34.Qg2 Bf6 35.g4 Qh7 36.Nc5 Re2 37.Nxd3 cxd3 38.Qf3 Bxh4 39.Rad1 Bxf2+ 40.Rxf2 Rxf2 41.Qxf2 Qe4 42.Qf1 Qd4+ 43.Kh2 Re8 44.Rxd3 Qxb2+ 45.Kg3 Qxa2 46.d6 Qe6 47.d7 Rd8 48.Qb1 a5 49.Qc2 Qb6 50.g5 Kf8 51.Qd1 Qc7 52.Kg4 g6
|May-08-19|| ||keypusher: Incidentally, 26.d6 can be answered with 26....Bxe4! 27.dxe7 Qh3!, since 28.exf8+ Rxf8 29.f3 Qxg3+ 30.Kh1 Ng4 31.Rf2 Nxe3 is hopeless for White.|
|May-08-19|| ||AylerKupp: <<Sally Simpson> I'm not anti-computer, I just don't think they have reached their full potential and are being programmed solely to beat other computers.>|
Let's just say that you don't think that using computers is always in a player's best interest. And I would technically agree, there are times that you have said that when the game is lost even with best play on your part, you might as well play a move that is less than the best in the hope that your opponent didn't expect it and therefore will play less than his best.
While I agree that the player with the lost position has nothing to lose in that situation and that this approach might work against another human player, the odds will not be in your favor and it won't likely work against computers. Unless the option of PONDER=ON is set (when the engine uses its opponent's calculation time to calculate the best moves in response to its opponent's reply), an engine doesn't "expect" any particular move by its opponent. True, if the opponent doesn't deviate from one of the positions already examined, those position's evaluation are in its hash table, the engine won't have to recalculate them and would save some time, and the time saving might allow it to calculate its response to its opponent's next move a little deeper. And that additional search depth might allow the engine select a better line than it would have selected if it had to evaluate each position from scratch.
But would that lower search depth be a good tradeoff against its opponent not playing the best move? Again, technically possible but I doubt it.
While I agree that chess-playing computers have not reached their full potential, what has? I don't think that there are too many things about which you can say that with certainty. I also don't think that computers are being programmed solely to beat other computers, they are programmed to play the best chess they can. The fact that their best opponents are other computers and not humans is, as you said, a by product.
<If we can get them seeing and thinking like we do then they really will enrich the game. At the moment our 1400 players are displaying more imagination than them.>
It depends on your perspective and what you consider important. I personally want to see the best chess being played by both players. If you think that it's more imaginative to have computers play like 1400-level players and that this somehow enriches the game, I suppose that's your prerogative.
<They are on the right path with AlphaZero. (at last).>
If by the "right path" you mean having a chess engine play more like humans, that's hardly a new thing. Earlier versions of engines would make many unintentional mistakes and even I could beat them on occasion. In later versions of chess engines you could sometimes reduce the playing strength of the engine by having it occasionally play a move selected at random (which would likely not be the best move) from the list of all the legal moves possible from a given position. The weaker you wanted the engine to play, the more moves it would select at random. In that case the engine would play very human-like, so chess engines have been on the "right path" for a long time. Is this what you are referring to?
|May-09-19|| ||Sally Simpson: Hello Again AylerKupp,
" Earlier versions of engines would make many unintentional mistakes..."
No I do not mean or want that.
I want them to play how humans do, by taking a path that sets more OTB problems for the human to solve and if possible a pitfall on the way. (not a trap that can rebound - a pitfall).
Take the number 6. 3+3 = 6 but so does 5+1, 4+2, 10-4, 12114/2019 = 6. Millions of different ways to reach the same number and if we relate that to chess a smaller number of different ways to reach the same position or the same evaluation number give or take a 0.0 fraction.
Instead of the computer asking what is 3+3 of a human, I want it ask what number do you divide 12114 by to get 6. Sometimes they do but if the eval number is 0.01 in favour of 3+3 then that is what it will choose when the move 0.01 away is more difficult for the human to see and an error may occur.
It looks at the second choice line 20 moves deep and sees the 12114 = 6 problem on move 16. It knows there is, let's call it a move' a move that needs precise calculation. That is how I want them to 'think'. If move 16 solved then no harm done.
They cannot think like that (I keep adding here 'yet), though they are perfectly capable of seeing 20 moves deep, sometimes looking at positions that a human would find irrational and difficult to play
and in some cases the player would not even recognise that there is a 12114=6 problem to solve.
Very hard to tell a computer to 'think' like that. Some would say impossible. I say it will happen.
|May-09-19|| ||AylerKupp: <<Sally Simpson> I want them to play how humans do, by taking a path that sets more OTB problems for the human to solve and if possible a pitfall on the way.>|
So instead of being programmed to play the best chess in order to beat their best opponents, which at the moment happen to be other computers, you would prefer them to play like humans do and possibly risk losing games? Or simply set more problems for human opponents even though their opponents will likely not be humans, at least not among the top human players, and positions that might present problems for humans may not present problems to computers? To me whether chess engines play like humans or not is irrelevant, I'm only interested in their ability to play better chess, regardless of how they do it.
<Instead of the computer asking what is 3+3 of a human, I want it ask what number do you divide 12114 by to get 6. Sometimes they do but if the eval number is 0.01 in favour of 3+3 then that is what it will choose when the move 0.01 away is more difficult for the human to see and an error may occur.>
Why? What possible difference does it make whether a computer asks what is 3 + 3 or whether it asks what number do you need to divide 12,114 by in order to get 6 so that it can play better chess? Why would you want them to play like that if it doesn't enhance, and may diminish, their playing ability. I know that I'm taking that too literally but I just don't see the purpose of it.
In the early days of chess engines having them play like humans seemed to be the way to have the computers play better chess. It turned out to be a disappointment to many, particularly AI proponents, that the way to get the chess engines to play better, at least at that time, was not to attempt to have them play exactly like humans, but to take advantage of their strengths, aided by human knowledge. At least that's the way until fairly recently when neural network-based chess engines, when supported by the overwhelming computational capability of GPUs and TPUs, have shown that they can defeat classic chess engines operating with a substantial computational capability disadvantage.
But consider the way classic chess engines work. In the opening they follow an opening book just like a human memorizes opening variations. They have a hand crafted evaluation function comprised of factors that top human players think are important (material, king safety, space control, mobility, etc.). Each of these factors are given a relative importance (weights) that can in some engines, like Stockfish, vary depending on whether it's the middlegame, the endgame, or somewhere in between. The weights are initially established by the engine developers and then refined by having the engine play many games against itself and other engines and changing them according to the games' results. You might even call this a type of "reinforcement training".
And engines are aided in pruning their search tree by using both algorithms (like alpha-beta pruning) and heuristics such as Null Move Pruning, Late Move Reduction, Razoring, and many others. These heuristics differ from algorithms in that they do not guarantee the best solution, or even a solution, but merely represent an educated approach to solve the problem.
So how does this significantly differ from the way humans play? Humans assess the position, select the subset of moves that seem the most promising (candidate moves), reinforce that assessment by detailed calculation of the various candidate moves, and select the move to play based on the candidate move that seems to be the most promising. Of course, humans may also take into account things that the computer can't, such as a subjective assessment of their opponent, their position in a tournament or match, and whatever other intangibles they might consider important.
The main difference between classic chess engines and humans, as I see it, is that because of their superior ability to calculate chess engines can take into account a larger number of promising candidate moves than the human can, and can search deeper into the nuances of the position to determine the refutation of tactically-oriented moves. I suppose that's why many consider that classic chess engines operate by "brute force" when in reality they incorporate a lot of previously obtained human knowledge and it's how they combine that knowledge that to a large extent determines how well chess engines play.
<Very hard to tell a computer to 'think' like that. Some would say impossible. I say it will happen.>
And I would say that if telling a computer to 'think' like that does not improve its ability to play then it is not worth telling them to do so.
|May-09-19|| ||john barleycorn: I am not sure whether I can follow all this. What is a "classic chess engine"? One of those from "ancient" times where the human player would open 1.e3, keep the position closed, and finally outmanouver the engine? that dream was over as fast as preparation to utilize the "horizon effect" of the engines. Fact (for me) is that the "horizon effect" is now a "human" issue.|
|May-09-19|| ||Diademas: <john barleycorn: I am not sure whether I can follow all this. What is a "classic chess engine"?>|
The Turk (Automaton)
|May-09-19|| ||john barleycorn: <Diademas>, thank you.|
|May-09-19|| ||Sally Simpson: ***
Probably not explaining myself correctly.
Yes they can beat humans and have algorithms and heuristics with Null Move Pruning, Late Move Reduction and Razoring,
But let's take this simple position. (white to play)
click for larger view
With or without a tablebase it will not play 1.Ra7+ because it knows it is lost and can see no way out of it and it cannot anticipate a bad reply.(a human error)
1.Ra7+ Kd8 2.Rxa5 draw.(2...Rh5+ 3.Kd6)
And good humans fall for that one.
O Bernstein vs Smyslov, 1946
click for larger view
Black played 59...b2 60.Rxb2!
Shirov vs Morozevich, 2001
click for larger view
White played 55.Kb2 ? Black failed to see and play 55...Rxh4.
I'm not wanting the computer to dumb down I want it to give itself a chance and play what it knows is not the best move by seeing a way out if it can get a plausible error out of it's opponent.
(in the first diagram all it will do is play the moves that staves off mate for the longest number of moves like a cold robot without setting any trick or trap. It will just roll over. A one K ZX81 with it's penchant for giving worthless checks had more chance of saving that position than a modern computer - especially if it's stuck inside a Tablebase...with the Memphis Blues Again. )
If we can get it thinking along the lines of looking for a plausible error and going for it WHEN IT IS NOT LOSING and if the trick is spotted it is still not losing then it is starting to play Chess.
And if we can get it doing these things 10-20 moves deep then it's treats galore. At the moment they are what they are. Adding machines with a fancy name. Of no practical value at all to the average player often doing more harm than good.
'Yes they can beat humans and have algorithms and heuristics with Null Move Pruning, Late Move Reduction and Razoring...but with all it's hardware, softeware and all-singing, all-dancing gimmicks it still cannot set a basic two move trap and in this game (human v human) the two move trap is King.
|May-10-19|| ||AylerKupp: <<john barleycorn> I am not sure whether I can follow all this. What is a "classic chess engine"?>|
Sorry about that. I have been using this term for a while now and have described it, but it's not yet part of the vocabulary.
Basically what I call a "classis chess engine" is a chess engine developed along the lines of Claude Shannon's landmark 1949 paper, "Programming a Computer for Playing Chess", https://vision.unipv.it/IA1/aa2009-.... In it he describes things like:
(1) An evaluation function taking into consideration things like material, mobility, king safety, pawn structure.
(2) Applying the evaluation function only in quiescent positions, and using different evaluation functions in the opening, middlegame, and endgame.
(3) Use of the minimax algorithm to find the best line for both sides assuming that each player would play the move that either maximizes his advantage or minimizes his disadvantage.
(4) Using and pruning a search tree (although he didn't call it that) in order to reduce the number of search tree branches that need to be investigated.
These are still basic to constructing non neural network-based engines which don't use either evaluation functions or the minimax algorithm, instead relying on a probabilistic assessment of the various moves using the estimated scoring % for each move.
If you can think of a better term than "classic" for these types of engines, I'm all ears.
|May-10-19|| ||AylerKupp: <Sally Simpson> I think that I understand what you would like, we've discussed it before. In this case you would like the player at a disadvantage to play a move other than the objectively best move (which would not change the most likely outcome of the game, just delay it) in favor of an alternate move that might change the outcome of the game if the player with the advantage makes a mistake, particularly if that move "encourages" the player with the advantage to do so.|
<If we can get it thinking along the lines of looking for a plausible error and going for it WHEN IT IS NOT LOSING and if the trick is spotted it is still not losing then it is starting to play Chess.>
I fail to see that. Oh, it might succeed if it was playing a weak human opponent or a top-level opponent in time trouble, but it will not succeed against a computer opponent because the computer opponent will spot the trap and avoid it. It is not going to make plausible errors (whatever that might mean to a computer) since all 3 of those traps would be easily seen by a computer with only a 6 ply deep search.
Perhaps you are advocating that computers (and humans) should start all games with 1.e4 so that if 1...e5 2.Bc4 Nc6 3.Qh5 Nf6?? 4.Qxf7#? After all, the computer wouldn't be losing at that pont and if the trick is spotted it would still not be losing. Would you call this "starting to play Chess"
And the computer will not be playing against human opponents or, perhaps more accurately, human opponents will not play against computer unless they are given substantial material or time odds. So, given that there would be no advantage to implement this kind of logic in a classic engine (see my post to <john barleycorn> above), why would the human programmers spend any time incorporating this logic into their chess engines? I've mentioned that before also.
Maybe neural network-based engines will prove me wrong. After all, there is no explicit human effort required to implement this special logic in case the situation ever comes up. But it won't in an engine vs. engine game.
<And if we can get it doing these things 10-20 moves deep then it's treats galore.>
How is it a "treat" to have a computer, who will never get into a disadvantageous position against a human, try to set up a situation where it can place the human (who must be pretty good to get into a position where he has an advantage against a computer) fall for that trap?
<At the moment they are what they are. Adding machines with a fancy name. Of no practical value at all to the average player often doing more harm than good.>
Well, if the average player doesn't know how to use computers in order to learn something from it, and are not willing to spend the time to learn to do so, that's hardly the computer's fault, is it?
<'Yes they can beat humans and have algorithms and heuristics with Null Move Pruning, Late Move Reduction and Razoring...but with all it's hardware, softeware and all-singing, all-dancing gimmicks it still cannot set a basic two move trap and in this game (human v human) the two move trap is King.>
I think that's it's more of a case of not setting up a basic two move trap because there is no point in doing so rather than it being unable to do so. In the examples that you listed above, did it ever occur to you that by playing the best moves to extend the game as much as possible a player is maximizing the number of opportunities for its opponent to make a mistake? After all, no one ever won (or drawn) a game by resigning. The more moves are played the greater the chances that the player with the advantage will mess up. And the greater the likelihood that better opportunities will arise to set up less obvious traps that its opponent will more likely fall into, particularly if time is a factor.
|May-10-19|| ||Sally Simpson: ***
You are correct we have done all this before.
I am not interested in computer v computer games (I hesitate to use the word 'game' - neither knows it is playing a game.) nor computers v humans. Pointless.
I want it use it's undoubted ability to aid the average human player.
Might as well use this again. (White to play)
click for larger view
A computer will play 1.Qf3 because although the eval is over -3 it does not want to play any move that makes this very very slightly worse.
The 1400 Human in this position played 1 Qh3.
Same idea, get the Queens off to save the Rook, the difference between Qf3 and Qh3 is mere fractions, The knight re-takes on f3 on h3, yet 1.Qh3 is the best practical try and in the actual game it worked because Black took the h1 Rook - 2.Qc8 mate.
The 1400 human wins a game from a position a good computer would have lost simply because the machine cannot anticipate a human error.
I want them to start looking for practical moves expecting a plausible error.
A switch would do it. Keep the machine as they are now but flick a switch and it goes into human analysis mode (HAM) where it looks for moves that will give it a 0.25 increase in it's eval. (that should take out gross blunders and you can set parameters maybe up to 0.75) It will shoot down the 0.25 line using it's full playing strength seeing where it goes. That 0.25 may turn into a + 2 or even better.
That is what I meant by treats galore. It is looking for a bad move then it start show a lot more bang sac lines that it undoubtedly sees but refuses to show us because it all comes from one 'not the best move.'
At the moment we have to nudge it towards interesting looking 'not the best moves' positions, not all players have that skill. That is why these things are of practical value only to strong players.
This will, with a bit re-programming and tweaking, give ideas closer to human v human games where errors, tiny to massive, are the norm in every game.
I want it to think dirty, Is that really too much to ask?
< Earlier Kibitzing · PAGE 6 OF 6 ·