Leela Chess Zero traps Stockfish with BLACK
Vložit
- čas přidán 27. 09. 2023
- Featured is the round 1 semifinal game between Stockfish (3862) and Leela Chess Zero (3822) from Chess.com’s Computer Chess Championship 21. The opening is a Sicilian Defense, Rossolimo with play beginning in this Rapid (10+3) game on move 9. It is very rare that Stockfish loses with the white pieces. How did Leela Chess Zero manage to trap Stockfish? After Stockfish attacks Leela Chess Zero’s b-pawn on move 25, a trap is set. A convenient black setup, an optimization of pieces for flexibility, and key pawn breaks paved the way for Leela Chess Zero to win the trapped piece 68 moves later!
I'm a self-taught National Master in chess out of Pennsylvania, USA who was introduced to the game by my father in 1988 at the age of 8. The purpose of this channel is to share my knowledge of chess to help others improve their game. I enjoy continuing to improve my understanding of this great game, albeit slowly. Consider subscribing here on CZcams for frequent content, and/or connecting via any or all of the below social medias. Your support is greatly appreciated. Take care, bye. :)
★ CHESS.COM www.chess.com/member/chessnet... (affiliate link)
★ LICHESS.ORG lichess.org/@/Chess-Network
★ TWITCH / chessnetwork
★ TWITTER / chessnetwork
★ FACEBOOK / chessnetwork
★ PATREON / chessnetwork
★ DONATE www.paypal.com/cgi-bin/webscr...
Not many humans would have the patience to wait 60+ moves to capitalize on a trapped rook.
It would be really cool to see *two* evaluation bars: Stockfish's and Leela's. I wonder how long Stockfish thought it was ahead.
Sometimes it's only for one move because it's often unusual moves that gets pruned away quickly that's the winning idea's. If you prune away your opponents best and winning move before calculating deep enough, you obviously lose. After the move is played it has to be looked at and it's way easier to find why it wins.
Other times it's a positional squeeze where the advantage just increases slowly.
edit: Here it obviously takes stockfish forever to understand what even humans can understand. Stockfish took forever to understand that the rook was in jail and wasn't coming out. An example of the strongest engine struggling with a positional concept.
Well that's what the Lichess analysis is showing. If Gotham chess is to be believed Leela said the position with b4 was -4
I think tcec would have that info if you click the archived game. There would be 2 or 3 simultaneous eval graphs.
Leela played that rook trap like a true master. It was beautiful to watch.
Great choice of game! These chess engines are so strong now it's mind boggling. Stockfish trapping it's own rook on b3 was clearly a bad idea to a human, but it took an engine as strong as Leela to show how to punish it. Incredible!
As black, I would have accidentally undefended my b4 pawn about thirty times during that sequence.
97% accuracy no blunders VS. 98 accuracy and no blunders. What a story!
Well, Stockfish deeming Stockfish flawless isn't really surprising.
Rb3 is a game losing blunder from a slightly better position. Stockfish is obviously blind to it's own mistakes. Otherwise it wouldn't make them lol.
It's like watching two super villains fight for dominance over humanity...
Yes, great comment
Somehow, as there is no beauty on it
This is one of the best illustrations of the difference between brute force calculation and a neural network. Stockfish: in twenty moves I still have my rook, so I am OK😊. Leela: I can keep the rook in prison as long as I want 😂, so I am better. Only 45 moves later Stockfish realizes: oops, my rook is not safe anymore and will be taken in twenty moves😮. Leela: told you so!
Stockfish is far away from brute force. There is a ton of pruning going on.
but it still is brute force. why didnt it prune the losing rook b3 move? it coulsnt go deep enough to see thar its a move that loses.
Stockfish destryos leela
I believe SF uses a blend of lightweight pattern-savvy NN-based position evals + super-efficient calc-based tree pruning. Leela OTOH uses slow & deep NN-based pattern recognition for both eval and tree pruning. The problem for SF is that the worthlessness (and/or eventual loss) of "permanently trapped" pieces lies way beyond the calculation-driven tree-pruning horizon. Leela is trained such that move pruning and pos eval coordinate. Thus once leelas NN gets the faintest hint that trapped pieces are correlated with eventual wins, BOTH her move selection and pos evals will be driven to explore lines that sacrifice material or other positional advantages in order to set and stengthen such traps. And as she trains against herself, she gets better and better at recognizing and using the pattern.
Assuming SF also trains its NN with self-play, the recognition of subtle patterns like this (where material payoff is way beyond calculation horizon) will not be reinforced, because efficient calc-based pruning is so quick to discard any branch that sacs material for advantages way beyond calc horizon.
What i wonder is: would Leela be a more formidable foe if she was able to learn that SF has this blindspot? Her self-training likely causes her to assume opponent also recognizes the trapped piece pattern (and would actively avoid it). So there are probably a lot of lines where leela could attempt to trap a piece (longterm advantage +3.0) with material sac (shortterm material value -2.0). However, she might also see that the opponont can pass up the sac and refute the trap with a line valued at -0.1. But if she trained against SF and learned its blindspot, she could learn when to expect SF to have no clue that the longterm value of the trap is -3.0 and would almost certainly grab the illusory +2.0 material advantage.
That was an amazing game! That Stockfish was on the wrong end of it makes it even more amazing; you don't EVER see Stockfish taken to the back of the woodshed like that!
Indeed engines are engines but the way you’ve reviewed the game, you make it look so easy for humans. This video is amazing. Thanks for your efforts & time you spend making such videos.
That’s one of the best games I’ve ever seen
I love setting 68 move traps
i laughed out loud
One of the deeper traps out there. 😎
One of the greatest chess videos ever and this is an understatement. Thank you for this wonderful job, Jerry!
Leela was just teasing Stockfish at the end: "Look, I can give away 3 queens and I'm still winning!"
Btw, love how they let the computers play these games to mate nowadays. I remember when the losing side would just resign, "sparing" some computational energy of the chess computers or something. Maybe it was just them being used to humans resigning and implementing that into computer chess.
They really wrote a small novel with this one huh? The 115 page story of "The Hostage Rook"!
Is Leila just trying to demoralise opponents by making two queens that just get captured. "Hey look. I've made two queens - you are just useless"
I enjoyed the pop quiz that had a 68-move answer!
A brutal squeezing of the trapped rook. That rook was about as trapped as any rook ever was.
I like it that the engines don't resign so we get to see the checkmate. This one wasn't interesting, but sometimes you'd love to see it to the very end, but players resign before it happens.
This game should be showed in every chess school.
Always loved your explanation of ideas in a simple way. And the choice of games as well 👌
👍
Jerry is the best
22:12 She wanted extra style points for 3 queen sacrifices.
Jerry is back and so are we ❤
Not sure if you mentioned it, but this is the first time Stockfish lost as white in two years. Agadmatot covered this game as well, and mentioned that.
After 25 Rb3, Stockfish had the evaluation at 0.7, but Leela already had it at -4.x. Leela knew all the way back there that the game was over. Stockfish sometimes has difficulty recognizing the importance of trapped pieces, and this is one of those times.
What is -4.x?
Is that -4?
@@nix4110Minus 4.[something] I don't remember what the second part was, so I put x instead as a variable.
These engines have endgame databases for all positions with up to 6 (or is it 7?) pieces. So if they see a line that reaches a solved position they will always go for it, even if they have to give up material to get there -- it's safer.
I dont think AI engines use endgame databases
@@Aca99100 Yes they do, these engines are made to win games and with endgame databases you immediatly have the best possible solution.
What a beautiful game, and excellent commentary. "The trapped rook game"
Jerry always delivers good stuff.
very satisfying to see one over arching theme ... where the rook is trapped and eventually taken. re the pawns promoting at the end, I don't really know, but I guess maybe it's making a bee line for some table base?
Correct !! It is still a database app :))
Fantastic analysis of an incredibly high level game, thanks Jerry.
"Kingside castles, no sides castles" got me :D
😎
Interesting case of Stockfish struggling with a positional concept. Goes to show that top engines are still capable of blundering.
O-O-O - Queenside castle
O-O - Kingside castle
O! - no-side castle :)
I like O-no :)
The last couple of pawn sac moves was just Leela taunting Stockfish change my mind
most likely that is just how leela has learned to end games, because the entire goal of ML is just to simply win games and not lose, how you do it is irrelevant
Thanks for the video! Please make more videos of computer vs. computer analysis, Jerry. They're always so interesting and enjoyable.
👍😎
What a game, and what an analysis, Jerry! Thank you.
These high level computers REALLY like suffocating their opponents by restricting activity. My perception of how the computer views this strategy is that it seeks to bait the opponent into inadvertently trapping one of his pieces in an attempt to maintain equality. Then it patiently applies pressure with tactics all over the board, understanding the opponent will eventually be a little worse due to a hindered piece. And of course, at some point the position will lead to a decisive material imbalance after the trapped and constricted piece becomes too much of a liability and is either captured or fails to contribute to the game in a meaningful way.
Thank you Jerry, I really enjoy your content! Keep it up.
It's one thing to move like a computer, but to analyze why they moved where they did...BRAVO!
On the ending, that seems to be a general problem in neural networks. Theres no more reward than a guaranteed win, so it can start messing around.
There are solutions, but they usually dont impact actual performance so no point. Ive seen it pop up independently in 6 mediums and multiple architectures, for example rocket league's nexto bot would stop trying when far enough ahead.
Thank you Jerry .. you are super cool!
As soon as I saw Levi post his review I was really hoping to see you cover this! Your leela game videos are some of my favorites
That was a fun analysis, thanks!
No need to rush taking a trapped piece
Excellent analysis
Great video. Loved the rook and pawn being trapped
Thank you for the sharing all those insights on this game, Jerry. These analysis videos you're making really boost my chess IQ. It's particularly interesting how stockfish ran its rook aground on b3. I guess even engines can make ill advised decisions.
oh dope another Leela x Jerry vid I can watch 30 times
Leela trolling stockfish
so high level... can't make a single mistake or you will be slapped down.
Amazing game, thanks for sharing it!
Intelligent analysis as always, Jerry sir! ♥️
Your commentary was excellent and relaxed. As far as the tactical complications of the chess-playing machines is concerned: whoosh!
this is most complicated analaysis which i know from Jerry
.. Greetings from Bosnia and Herzegovina
I enjoy watching your analyses of these more than those of human games
World Class analysis. Congrats.
Thank you! A supper interesting game.
My personal favorite move is the 54..f3+ clearance sac. Guaranteeing white has two long-term weaknesses, beautiful.
we are back with neural network chess computers boys! man, i missed these. leela now feels like proper successor to alpha zero, who also used to wreck stockfish. i love it.
Funny thing is Leela is already far stronger than AlphaZero ever was.
What a weird ending? Maybe the computer wanted those squares available for its queen? Maybe it just knew it was winning so it stopped caring?
that rook was out of the game so early...
A dumb theory on 22:10
I could imagine Leela being lazy since the position after 104. Rxh1 is down to 7 pieces and an "obvious" table base position for her.
On the other hand, she has to calculate any faster mates with 8-pieces and beyond by herself.
Leela was flossing at the end there lmao
The difference between winning and losing is 1% accuracy.
yes, if you make the slightest mistake, computer will key on it and pry open your position till you collapse.
incredible game, goes to show what a beast stockfish is that it takes a 60 move deep strategic idea to beat it
leela won dude
@@hqcart1 leela sure did. your point?
GREAT JERRY. THANKS
Hey hotrod, who da funk is Jerry and what does he have to do with the price of tea in China?
The rook trap literally no human could lay out in advance.
At the end i guess Leela is taking tablebase in 3 as better than mate in 6
Please give us more leela v stockfish
basically the engine is not able to see its blunder past some risdiculous number of moves. brute force conputation doesnt work in many positions
A true masterpiece by Leela.
Wow.
Amazing squeeze by Leela. SF was better out of the opening, Leela reached a drawing position and then managed to see that she was better where SF was blind and suddenly Black was winning.
This was awesome
thx Jerry 😊
Giving the promoting pawns away freely was really weird. Neural nets must be watching some hyper bullet time scrambles or something.
most likely because it had already calculated the win without the extra material, so it saw no need to overcomplicate, after all, the only thing neural nets care about is the win, point being neural nets dont care if they win by material at all, the material is just a means to an end
Weird how stockfish didn't just sacrifice the exchange to at least have some counter play. It really liked that rook.
When did it have the chance to sacrifice the exchange?
@@codegeass7162Wow, 56 minutes ago. Cool.
There were plenty of times. I can't remember when, rewatch the video.
Lc0 is not really optimized for fast endgames.
It gives away material as long as it is not losing
Tanks for that. I have watched this game before with Gotham chess and was terrible. I guess we can call this strategy " bad beshop x terrible rook" it's amazing how every pawn it's in a dark square....crazy game
Possible ending explanation: 50-move rule with pawn leads to the neural net having a slight bias toward getting rid of the pawns. Or favors having open boards, or something. Whatever the case, it isn’t important to play the fastest win if you’re guaranteed to win regardless, so the training could’ve been a bit wonky for completely winning end games
So weird when the machines play...
Good job
I'll always click on a Leela vid.
You spelled Jerry wrong. ;)
These kids are so good!
that was rad
Giving away the pawns at the end was just to reach a tablebase position and from there it's just lookups. Not the fastest conversion but the easiest computationally.
Perhaps. 😊
this was very cool
I'm still amused by the bizarre moves the engines make in obviously won endgames.
Thanks for another great video, Jerry. You really do a great job on these engine games.
I think it is possible that the neural network Stockfish uses for evaluation (NNUE) has no way of learning about trapped pieces, given its design. There might be room for improvement by defining some hand-crafted features (e.g., the number of squares each piece can move to that are not attacked by a less valuable enemy piece) and feeding them as part of the input to the neural network.
LCZero's network architecture is much more complex and probably has figured out having trapped pieces is a problem.
The reason for dumping the material at the end might just be that it found a book winning endgame with the fewest number of pieces first and just beelines for it rather than searching deeper for a quicker win.
There's a game where Alpha gives up the Lopez Bishop for a pawn (just left the B there for a pawn to capture) in an Italian type position against Stockfish in order to trap the Black light squared Bishop and Rook in the corner.. I can't find the game again though I would love to see you analyze that one.
as far as chess videos go this one is rated 3862
you know the commentary is brilliant when you yourself come up with the key moves of a 3800 engine while going through the game. thank you jerry!
This game really left an impression on me. The first thing I thought was to state a bad Bishop strategy against a dead Rook, but she's not dead, it's worse than that, she still needs protection, time and resources. The bishop is fighting with a rook in agony. Was there not a time to accept this sacrifice in A3?
Man this reminded me of that game in 2018/2019 when A0 locked up SF8's queen in the corner. Honestly I still think that one was more aesthetic, but this one is very close (not to mention played at a much higher level of course).
Love these vids. All of them. But for me personally, I get much more value from human games than computer games.
22:10 Leela gives pieces and pawns away when it's already completely winning because it's then less to calculate
How? At that point It is probably a forced checkmate in a couple of moves. It can't get any simpler than that.
Clearly leela isn't coded to pursue the fastest mate.
If it found a winning line, it stops calculating?
Forcing moves y"see. Opponent makes queen you gotta capture it
It's mate in 6 and Leela calculates that almost instantly so it's not that.
@@Halibut86 leela is coded to know the rules of the game. Other than that, it is hard to tell how leela "ticks" as pretty much everything else is the result of machine-learning.
Feeling sad for the saddest warrior on the board, the b pawn😅😂
10:15 What would happen if black would play dxc3 instead of the queen move? The move looks nice (due to opening the diagonal file attacking the f2 square) but complicated to calculate all the following moves.
At the very least you're also freeing the rook tho
I think they have programmed it to floss if it's completely winning 😂😂
Well I think I spotted 3 of the moves, but probably got the deeper reasoning behind all of them wrong!