GM Grigoryan on the "Myth" of Solving Puzzles

 

What GM Matthew Sadler said in his interview on the Perpetual Chess Podcast (starting at 00:25:00) was quite specific:  1) That the developers of Leela Chess Zero (Lc0) stated that most of the strength of the engine was in its positional evaluation and that only a modest additional increment came from looking ahead.  2)  That he played a series of games against Lc0 set to look ahead only to the next move and found that he won about 78% of them, putting its strength with these settings around FIDE 2500.

This means that yes, if you set up a two-move combination, the engine would not see it.  However, it's very possible that you would rarely get the opportunity.

(If you think about the type of pattern recognition that tactics practice teaches, though, it makes sense that an engine with sufficiently advanced static evaluation might be able to make moves that are optimal for foiling tactics without having to actually calculate what those tactics are.  In a sense, being able to look at the board and just "see" a tactic without calculating amounts to static evaluation.)

Wow. That’s interesting. Do other engines act like this, or is this exclusive to Lc0?

 

I don't really know.  However, his book, The Silicon Road to Chess Improvement, might offer more detail.  I haven't yet read it (though I plan to!)

Ah, thank you!

Pretty sure you are becoming rude now. Paleobotanical has already answered. I will add this example:

 

 

I am pretty sure that, with some effort maybe, any 1400 can find the solution to this puzzle. But if I set my LC0 (v0.27) to analyze this position at 1 node, its UCI output is:

info depth 1 seldepth 1 time 3670 nodes 1 score cp 2727 tbhits 0 pv c8e8

bestmove c8e8

Now, since I don't know what I'm talking about, will you explain to us the meaning of that output?  At least everybody can see that Leela with the above settings plays 1... Qe8 here, which loses (if another NN was used, maybe it could choose the right move in this puzzle, but it would fail in another one). But I invite you to play against it an entire game at standard time control, setting it to produce the same type of output as above every move. This engine at 1 node can "see" simple tactics like forks, pins, etc. "at a glance"; I have tested it right now, but no more than that, as showed by the example above. Its tactical strength is far less than a master.

I have already said that your challenge cannot prove anything, because you say that you use only tactics to play, but we cannot be sure of that. Besides, none can train a 500 elo rated player to become stronger than a 1800 in a day, even in a specific position (apart from some endgames, like K+P vs. K). So, I beg your pardon, but your challenge does not make any sense.

You go on repeating the same thing over and over, simply ignoring or denying any other argument, not disproving them. It has been pointed out that White wins more than 50% of the times at all levels; how do you explain that, if below 2000 being in a positional advantage is useless, like you say?

None has said that tactics and calculation are not important. In fact, LC0 with calculation and tactical strength is a top engine, much stronger than a GM.

Sorry, but you are acting like a troll, so I think DrJetlag is right: time to move on. I will not argue with you anymore.

Does Lc0 use Silman’s imbalance method as well?

It does not use any hand-crafted method to evaluate positions. It uses a self trained neural network, hence the "zero" in the name: no one taught it anything but the game rules.

Wow. That’s amazing. 

 

Now that I'm aware of the relationship between Lc0 and AlphaZero (which I learned just above in this thread), I can answer this: Lc0's static evaluation is based on a neural network that's trained by running many millions of games against itself.  The developers do not rely on implementing heuristics.  Instead, over this very large number of games, it's developed its own algorithm for evaluating positions.

Unfortunately, the nature of deep neural networks is that it's extremely difficult to extract useful heuristics from them.  They're complex, nonlinear, and do not lend themselves well to human analysis.  There is reportedly some work going on to try to improve the visibility of Lc0's internal process somehow, but I'm not familiar with the details.

It's important to note, in relation to the topic of this thread, as I did above, that training pattern recognition can turn calculation problems into static evaluation problems.  This is why the approach to tactics training that some coaches recommend is to focus on drilling simple tactical patterns repeatedly to build speed, rather than pushing to the limits of one's calculation ability to try to exercise that skill.  It's certain that Lc0's static evaluation includes a LOT of pattern recognition that amounts to stuff like "That position is obviously crappy because it leads to a fork" even if it doesn't explicitly calculate future moves.

Ah, thank you. Pardon my ignorance. 

 

I certainly didn't mean to call you out or make you feel bad!  I think my post crossed over that of haiaku.  Interesting stuff, though, and that's part of why GM Sadler has spent such time studying it.

Engines are more interesting than I thought!

A heuristic neural net is a rather different type of engine.

You have to know the concrete before the abstract … rules, moves, checkmate, stalemate, simple captures, and simple defense. Then tactics and endgames (positions where one can force the above regardless of opponents’ defense). Then strategy, then openings (methods of trying to achieve flexible positions where it seems like you might be able to eventually force the above or fluster your opponent into a bigger mistake that reduces the position to a forced tactical or endgame win).

Perhaps it is not completely linear, and you don’t need to be a 4300-level tactician before beginning to study other phases of the game (it is easier to repeat opening principles and strategic rules of thumb than to consistently solve 8-move checkmate compositions), but perhaps it is much more important than most players think.

But perhaps players would get more bang for the buck working to recognize and calculate forced winning and drawing sequences that are 4-6 moves deep rather than trying to assess the needs of a position that may require 15-20 moves in multiple branching lines to properly assess, as is typical of GM analysis and preparation.

Thousands of endgames have shown that so called "bad strategic moves" actually end up winning. Analyze the famous endgame of carlsen vs karjakin i think it was?(world championship game) where you had to purposefully restrict your knight and put yourself in zugzwang in order to win... sounds strategically horrible but works tactically...

Watch kingcrusher's video on it as well... where in one of his games he played all strategically correct moves but later found out that his opponent had a crushing attack...

Pretty sure i am the only one who actually has gone through several games of lower rated players since i find that its a good way to train tactics. They allow tactics frequently and i aim to find out the winning shot. By going through several games i can clearly see that they don't lose games due to "missing a deep strategy"... nah, they lose as they allow combinations. This is a common, constantly occurring theme.

It seems to me that no one here has spent any time actually going through the games of lower rated players. If you go through carlsen's games then u would obviously think that strategy is terribly important since carlsen does not allow small tactics to ever show up. However we are talking about lower rated players and there biggest weakness lies in there tactical ability and in them allowing so many combinations. That's where they are lacking.

Not trolling at all, just speaking from experience. I urge you to go over the games of some 1000 rated players in this website and u will see for yourself... they don't need lectures about pawn structures, they need to just "see" moves... that's all which is needed

I think their comment comes from a good place but they are just ignorant of what actually happens in games of lower rated players. They are suffering from kind of blindness of the issues of low skill. It's common in other games and is common in chess as well.

I am sure people who are putting too much weight on strategy have never actually studied the games of lower rated players. People study games of Super GMs and those games will make anyone think that strategy is the soul of chess since those super GMs don't allow small tactics to be executed so the game turns into strategy.

Any study of the games of lower rated players immediately reveals the issue which they have, the issue is not of a strategical nature. There is no use in learning pawn formations when u are hanging a piece every 3 moves. 

I once had a discussion with him where it became clear he didn't understand how to play the white side of a classical KID. He insisted white only opens the c-file when white has a target there and in fact, white should open it quickly, get the major pieces off and then infiltrate with the minor pieces. So his version of "strategically correct" might not be entirely accurate.

I liked how u ignored the carlsen karjakin example... 

Anyway, i am talking about a single game which he played. In that game his moves all followed principles and logic... yet analysis from computer later revealed that his opponent had a crushing tactic to destroy white's position. He made a video on it as well on how the brute force of engine is quickly destroying any strategical stuff.

The simplest way i can explain my case is this... chess, in it's purest form, is just interaction between various pieces just like how football in it's purest form is just men manipulating the movement of a ball with their feet. Unless a guy is good in seeing various interactions he cannot be a good chess player just like how unless a guy is good with his foot control with a ball he cannot be a good football player.

Strategy is needed when your opponent won't allow you to just take his pieces.... u NEED strategy to win in that situation. However lower rated players allow MANY combinations, many tactics and frequently hang pieces... therefore to beat them u DO NOT need strategy...

I hope my point is clear. Please go through the games of lower rated players and see for yourself. 

 

I analyze a good number of lower-rated player games because I'm one of them.  Yes, simple tactical errors (in my case usually failure to see 2-3 move combinations rather than single-move blunders, though they happen too) decide the majority of my games, and working on that one thing is a reasonable strategy to improve.

But, in certain games, reasons for the outcome are less obvious.  I'll make a series of mistakes that do things like reduce coordination of my pieces, break up stable pawn structures, and ignore strategic rules like improving king safety and getting rooks on open files.  None of these moves lead directly to loss of material, but my engine evaluation and board position just gets worse and worse until my opponent has all the options and I have few.

On the Perpetual Chess Podcast, guests with titles and coaching experience have offered a range of opinions, from:  low level players should focus mainly on tactics, to: get basic strategic ideas into low-level players' heads early so they can start thinking about how they interact with and lead to tactical options.

There are few, however, who take the hard-line you do, which is to suggest that low-level players' tactical problems render all discussion of strategy categorically pointless.  A lot of those "strategic" guidelines and heuristics pretty directly help improve one's own options at the expense of one's opponent's (for example, getting one's rook to the open rank first.)