Here is my take on alpha zero. I watch all its games. From what i have seen here is its tactics: it locks the middle, immobilizes a few of its opponent's minor pieces deep in their ranks, it sacrifices a minor piece or pawns to create files and mobilizes all its minor pieces for positional gains. It can be down one piece or two, but in reality it actually up cause some of its opponents pieces are immobilize or lock. Very clever. It creates multiple traps so that its opponents have an option to die slowly or die fast. Very naughty machine
Something I've noted from watching alphazero is it seems alpha has a tendency to use his knights to force its opponent in zugzwang. As bait, sacrificing the knight to (usually) a bishop, in turn taking bishop.
YEAH, when ever I see a deep mind video I insta-click. The last few days I sat down and learned about how deepmind actually works, and it turns out deepmind is learning even when it's playing games against stockfish. This answers your question about the refusing of the draw. Deepmind probably found a good continuation after bying some time with the repeat moves. Keep making deepmind videos please. Deepmind is now my new favorite chess player :)
Beerdy - Bruce Lee Central while it's certainly true that AlphaZero could learn while playing Stockfish, there are two reasons why I think they wouldn't have bothered: Firstly, it might learn bad moves from stockfish. AlphaZero learns through self play, because it gets the best game data from itself. Secondly because it would waste computation that could be spent focusing on the match. It is true though that after the match the game data could be fun through AlphaZero to make it better, but 100 games would be such a small contribution to its training set that I wouldn't see the point. In the four hours, it would have played over 700 billion games with itself.
I teach deep learning, and I agree that they likely didn't train deep mind while playing the matches against stockfish. If they did the games would have little impact in it's play vs the millions of games is has already played.
Hey, i have some questions about deep learning, just basic question to how to learn it. if you have time to answer them, i can give you contact details? shouldnt take more than tops 10 mins @Nathan Burnham facebook.com/omer.ulger.397
I'm not sure where you got 700 billion from, but I believe AlphaZero has played only a few million matches against itself. The preprint mentions only 700,000 games and claims that its performance exceeded Stockfish's after just 300,000.
There is a fitness function in ai that says how good is your solution. It can be as simple as winning = 1, losing = -1. If they set draw as a -0,1 or something similar it gonna refuse the draw by repetition. They should release the progression of learning from those 4 hours. Usually its funny stuff. First games random moves ending in perpetual checks. Then it learns not to draw and come up with some crazy attacks and trades, and eventually it would come up with openings, and basic strategies. Ending in this fish eating beast.
*Agadmator said,* "No more AlphaZero vs. Stockfish videos..we may not be able to appreciate human chess.." You may be a victim of your own words my friend haha.
I guess, Alpha0 was programmed to go for 2-time repetitions whenever possible, in order to induce the horizon effect in typical chess engines that examine the game tree to relatively small depths (usually up to 20 moves ahead in the middlegame). One of the ideas that Matthew Lai (one of the Alpha0 contributors) had expressed in his master thesis is to explore the tree deeper in those variations that are assessed (by a special neural network) as the most likely to be in 'the principal variation' (i.e. to be played if both sides play optimally), which is closer to how humans calculate variations, as opposed to usual chess engines that waste time on a lot of improbable variations and extend the search depth only in very specific 'violent' situations (like captures). Alpha0 uses this 'neural network for move probabilities' approach in its Monte-Carlo tree search (search for 'AlphaZero' on arXiv.org and read that preprint) and sees further than Stockfish in the critical variations that end up appearing on the board.
You know what would be interesting if they give alpha and stockfish different opening positions like nimzo or sicilian or some well known gambit positions like kings gambit etc and let them play
is not just that, in the second set of the three moves repetition, the rook and the bishop end up in the opposite position in comparison with the first set. which i thinks it's what alfa wanted. so you are right, the best move for black is the rook but not until the threefold, it's because of it. Saludos de un hincha cuervo desde huerta grande, córdoba
Hola cuervo, i don't think that AlphaZero ever hopes for a threefold. He just want to avoid it as long as he can find another node that's better than draw.
Albo's native language is Spanish. Since in Spanish Alpha Zero's word 'gender' would be masculine we automatically think of "he". It's a mistake we commonly make.
Cuervo 10 It's not what AlphaZero wanted, but rather it was forced to play a different move as playing the same move gave Stockfish the chance to draw.
My theory for the threefold repetition thing is that both of the bots are playing their best move, which just so happens to be a repetitive move. The thing is, when they get to the third time that move is no longer the best, because it leads to a draw, so it rethinks its move and does a better one
ah so they do the second best move as the "best move" would lead to a draw otherwise so basically the second best move becomes the best move. i think that kinda makes sense yea
This doesn't really make sense to me. If the second move is better than a draw then how could the first move be the best move if it leads to a draw? That is counterintuitive
@@liljackypaperThe draw only happens after 3 repetitions. The best move doesn't lead to draw in the first 2 times, so playing it in the tiny hope opponent plays less than optimal move, is worth it.
@@liljackypaper If they lose nothing from doing it, why not? Alphazero specifically, being MCTS, would treat the tiny chance of opponent playing wrong worth the extra move. Stockfish I think would treat the moves equally good, with or without mate-in-1 trap
If, back when he was the best player in the world, Paul Morphy had decided to never play 1. e4, that wouldn't have meant he had refuted it. If AlphaZero never plays 1. e4, that may be because it is less successful at that opening, but there are all sorts of reasons why that might be the case beyond it simply being a bad opening.
@@EebstertheGreat if A0 never plays 1. e4 and this is because it assesses it as less successful, this is exactly what a refutation is dude. after all, it never plays 1. e4 :o and in this video you can clearly see why.
@@Pintkonan That is not what a refutation means. A refutation does not mean the world champion is bad at that opening (or marginally less good than d4). A line is refuted if a refutation is found--that is, if a defense is found that proves the line is worse than another move. There is no defense to e4 that has been demonstrated to be successful, it's just that over many games, some engines win more with d4. If you want to be strict about it, by your logic, every opening is refuted except the single opening that Lc0 or Stockfish prefers. And if that ever changes in a better engine, suddenly the opening becomes unrefuted.
It would be so interesting to force these engines to a certain opening and let them continue from there, just to see how it would turn out. Gotta say I did not expect to like these engine games, but they are pure gold and I´m so excited when I see you have uploaded another one! Keep up the good work, cheers from Finland \o
At 6:41 how did black capture after white moved to f4? They were both on the 4th row next to each other (?) Am I missing something? Please help, I'm fairly new to chess, excuse my lack of knowledge
I wonder if there's a way to deconstruct scenarios and outcomes from various playing scenarios by forcing Alpha 0 to play itself and set it's opening sequences (Queen's Indian / Belgian v e4 or others) then release the analysis to discover why... Could be interesting either way it's incredible play and thank you for sharing
I really like the fact that these games are a real grind, and not just ownage in just few moves. For humans this game is great (its well balanced for both sides) and certainly has alot of future!!
In your variation, you can simply play Bc3 with white and you do stop both pawns for a while. It seems that the pound in H will fall but the black king will take d6 and from there is not dificult to win.
Agadmator I love your channel and have watched most of your videos. Am planning to watch the rest of the videos too. I would like you to do videos on Chess openings and discuss the various variations in each main opening. There are few good chess openings videos on RUclips and I am sure the entire community would learn from your videos. Please do chess openings videos.
Agreed, alpha repeats moves bc he thinks he has a better position and is trying to 'bait' his opponent or exhaust all chances that his opponent will make a different move and not repeat ... After alpha repeats twice he doesn't want the draw bc , idk, alpha thinks the position is still favorable or playable (in alphas mind!) So he avoids the draw and moves on vs. he always has a draw in the back pocket or he learns a new moves after the same position is repeated.
Hi agadmator. I just want to share, wouldn't it be interesting if Alpha Zero plays against itself? I mean, what if there are the same moves every time? What if white always wins? What would it mean?
What do you think would happen if Google let Alpha Zero train for more than 4 hours? What kind of God would they create if they let it train for weeks???? #suggestion if you can find a Alpha Zero vs Alpha Zero game that would be very interesting.
what i find interesting about the repition breaks by alpha is not the question whether it found a winning variation in the mean time or just trys to win and having the draw in its sleeve all the time but whether it *cares* about not losing a game and going on with a move it considers less optimal because it *wants* to go on.
I personally think that alpha chooses the best position, by repeating the position 2 times, and then doing it again, in the end it will be in the same position that it actually wanted to be (because after 2 repeatitions it will be 1 turn away from it's optimal position and after doing it twice it's exactly where it wanted to be) It circumwents the 3 fold repetition rule so to say to force the opponent to accept a position where they normally would want to draw.
I pray the algorithm is developed for commercial distribution because this method of search is beautiful to behold. Could go a long way towards training the next generation of chess players.
Amazing video, thank you! I have some questions: It seems to me from the AlphaZero games and paper that it's power lays in super advanced stratigique thought (or maybe stratigique calculation? Hard to chose words to describe this "entity"). Whereas, from my limited knowledge, Stockfish's (and chess engines in general) strength lays in brute force of calculation. So, added with opening books and different middle and endgame tables, Stockfish is merely "mimicking" stratigique thinking, but isn't actually considering positinal aspects, space usage, flexebilty, activity and synergy. It IS eventualy "taken into consideration" indirectly via brute force, because the consequences of such elements are evident in lines calculated by Stockfish. Against a human or an inferior engine, the force of calculation is enough to "hide" the inability to think/calculate strategy. But it seems this is how it is outplayed by AlphaZero.. Also, Stockfish is engineered by humans to evaluate a position not only by calculating possible lines of play but also through material numeric value. Maybe we, humans, "misled" stockfish by "teaching" it a wrong or incomplete evaluation of material and position process... Maybe AlphaZero can teach us a new way of thinking about material value. Either we will learn that a knight is actually worth 3.5 and a bishop is worth 2.7, for example, or that it's wrong to even go through that line of thinking. What's also interesting in my opinion, is that SF's brute force makes it a "god" of tactics, as tactics are based on calculation rather than "thought". (They could also be based on 'post-calculation'. A GM doesn't have to always calculate a full process to spot a tactical trap, he/she can train to see it by noticing patterns and structures, or known lines of "theory" based and calculation made by them or someone else (including engines) in the past). I believe Stockfish is bound to always calculate, and it can't develop these abillities that GM's can. Though, It probably doesnt mind (Pun intended ;) ). It is a preety f***ing good calculator. But is it possible that AlphaZero DOES develop (like a human would) to recognizes tactis without calculating all the time? Is it possible AlphaZero is "thinking" strategy in a broad and complex way? Is it possible Stockfish is yet superior in tactics? Would be interesting to present them both with very complicated chess puzzles to see who is better. (Though probably even AlphaZero's inferior calculation power of 'only' 80,000 positions per second can stand any chess puzzle we humans created, and the gap between SF's and AlphaZero's tactiacal quality - if indeed exsists such a gap - would be insignificant or impossible to notice unless both of them are given only fractions of a second to solve the puzzle.) I want to add that all the asumpstions I based my thoughts upon could be flase. I am new to chess and know almost nothing about computering and AI tech. Also, as some people find it somewhat depressing that AlphaZero belittled centuries of game development in 4 hours, I want to add an incourging thought: Even though AlphaZero outclassed us and our programs so effortlessly, it still isn't capable of INVENTING AND DEVOLPING the game of chess. Or even if it is, if instruced to come up with a game, it can't do so just because it WANTS to and INTRIGUED by it. We still have the ability of doing something for the sake of pure enjoyment going for us. For now. :)
HI AGADMATOR.... i am a huge fan of ur chess commentry...i would like to make a suggestion though... i think it would be very nice if u could make some space on the screen dedicated for the move (like b7 ,e4 etc) , for dead peices and for the name of the opening or attack(like kings indian defence or scicilian defence etc.) thank u
It may not be a good idea as it would take the attention away from the action on the board. If you want the moves displayed, then what's the need of commentary?
You're right, Sameer. In the last video someone had requested that the positions (Eg. Knight to c3 etc)NOT be mentioned. I didn't quite understand. I mean how are you supposed to analyze otherwise. But to answer brandybuck, I would suggest this 1. If you needed screen space because it's difficult to grasp notations along with the video: I think I understand the nature of your request. But trust me, everyone has these struggles in the beginning where they find it difficult to comprehend the positions on the board and the notation. It will come with time. Suddenly one day, you'll be talking to yourself in notations (atleast for the first few moves). Suddenly when someone says 1.e4 c5 you'll inherently know it's the Sicilian without even looking at the screen or the board. And that's a beautiful feeling. Most people have this difficulty in the beginning and that might even deter you at times but that's the beauty of chess. But that's exactly why chess games are immortal because we can recreate games from 200 years ago with just notations. They're really powerful but it takes a few hours to get used to them. I sincerely hope you get through this phase and appreciate them, if this was your concern. 2. If your request was just aesthetic in nature, I agree with Sameer in saying that it does take attention away from the board. The description of the video gives the exact moves played in the game. One thing that you can do is to import the pgn if you are analyzing or watch more and more videos so that you do not need any screen space for the moves.
If you haven't noticed already, the notations are there in the description. You don't find other popular youtube channels including notations with every move on the screen. And that is for the same reason. There is a strict correlation between what we see and what is perceived by our auditory senses at the same time. While watching a game, it's important to pay attention to the move on the board while the brain automates the position in the brain (as we hear the notation). If it is displayed along with every move, then it definitely becomes a distraction because our eyes would shift between what's happening and what's being displayed. Like I said, I understand it from a beginner's perspective for the sake of convenience but a little effort and patience can't do any harm. You don't find a lot of subscribers requesting this for the same purpose because most people already know that, it doesn't take ages to get used to notations. I do, however agree with you on the point of making a website. That would be cool. Might need Medo's picture as the background :D
Mozisi. ur right. Initially it was difficult for me but now i can understand easily and now i can visualize positions just by hearing them. like your explanation :-)
Hey agadmaster, can you offer general recommendations on the best chess books available with reference to particular authors and publishers.. i would also be interested to hear specific titles that you think are particularly good!
Hi, i study machine learning, i think the 3 move repetition has to do with how self learning neural networks trains themself -- it has probably learned that the same position happening 3 times means a draw -- but if all it was given is the rules of chess it would have never explored past 3 repeated positions because it would have considered that position "known" or solved for -- the end implication is that is is now forced to pick its second best move, which it also probably thinks is winning .
Deepmind Alpha Zero's tight knit structure and movement of play reminds me of a sliding tile puzzle, both interlocking and spiral. A snapshot of Stockfish seems a looser version of the same structure and movement relative to Deepmind Alpha Zero (e.g. 5:58).
To understand why the position repeats but the draw is refused we can look at the algorithm they used to train Alpha Zero. It uses a kind of Monte Carlo method which is a randomization procedure to decide which moves to try next. This means while training if you allow your opponent more opportunities to deviate from the best line you have a higher probability of winning in the position just because you get that extra roll of the dice. I would interpret this behavior as showing that alpha zero felt Stockfish's defense is optimal and that deviation from that line significantly improves Alpha Zero's evaluation of black's position.
6:45 I think that this Bishop x Knight exchange has multiple purposes: strategically that's a bad Bishop in a close position, so it's great to exchange for a centralized knight; also that knight was defending c4, which will now depend on the Queen; as she can't move now, the obvious Alekhine Gun that Stockfish was planning to create at f won't happen.
The reason AlphaZero plays for two repetitions is because it's designed to play the best move for the position on the board. In those situations, it really is playing the move it thinks is best. And only when the 'best' move would force a draw by three fold repetition does it consider another move. I think the takeaway you should probably have is that in those positions, AlphaZero prefers the move only if it does not cause a draw. The move it plays instead is its second choice, but still has winning chances.
I was reading about Google's AI was actually acquired from a startup company. The curious thing is that when AZ learned Go by playing against itself for a number of hours . That version actually defeated the AI version that learned from a feed of thousands of previous games. I wonder if AZ learned chess from zero also or it was fed games... Hope Antonio can clarify...
If you look in the PDF paper on this, they did put Alpha Zero against Stockfish in the Ruy Lopez in 100 games, which they did in several openings, however, it's not clear which position it started from, either 3. Bb5 - a3 as in the picture, or 7. Bb3 - 0-0 as in the "PV". Nonetheless, Alpha Zero as black won 6 games, drew 44 and lost 0 from which ever position. As white, won 27, drew 22 and lost 1.
#agadmator I'm trying to find the Leela game (or Alpha) were the white queen spent much of her time on H1 but am failing. Can you or someone else point me in the right direction?
I think you mean AZ vs SF SF played QID. One of the QID games. Several RUclips people covered it. Maybe you mean this game? : ruclips.net/video/NaMs2dBouoQ/видео.html
With respect to AlphaZero *almost* doing three-fold repetition twice in a row: It is very likely that the rook moves (in response to the bishop moves) are in fact the moves that are most likely to lead to a win, if threefold repetition were not a rule. However, AlphaZero is trained to recognize when a move will result in a win, loss, OR draw; the first two repetitions are simply AlphaZero taking the most winning move at the time, but the most winning move changes when it will directly lead to a draw. What is most impressive is that AlphaZero learned the threefold repetition rule; it was not hardcoded into the neural network as it would be for a classical engine. Considering how few games end in threefold repetition, it's truly amazing how DeepMind was able to generate enough games for AlphaZero to learn threefold repetition from scratch.
Would be interesting to see how well Alpha Zero played right after learning the moves... was it instantly a GM with it's calculating ability? I don't think I've seen it make a single bad move yet.
No. There were tons of draws, though- the final count was 28 wins, 72 draws, and 0 loses, if I recall correctly. (However, DeepMind has only released ten games, in all of which AlphaZero won.)
that's go. "Japan Chess" would be shogi (which alpha zero also played). But in any case, the video is about alphago, a predecessor of alphazero. The publication on alphazero does say, however, that alphazero lost several games to the world's top shogi and go engines
Jonah Albertijn not in the main 100 game series but stockfish won some when alphazero started after a set opening. The results can be found in the paper. I think overall across 1,300 games (100 main games and 100 per opening for the 12 most played openings) stockfish won 4%.
honestly, the only people i watch when it comes to chess commentary and analysis is MatoJelic and Agadmator's chess channel! They give the most accurate analysis and they are entertaining to listen to and watch!
I first learned about the technology that DAZ uses (parallel distributed processing) in the early 90's and it was revelatory; AI is smarter when it tries to act like a real brain than a computer. I use some of the lessons from this in my personal work (making mistakes is GOOD, as it helps you learn, don't be afraid of making a mess of something, you might learn something!) and now seeing how DAZ makes such interesting, elegant moves, it's very cool to see.
I am glad there some discussion in the comments regarding the apparent handicaps Stockfish was dealt at the beginning of the match. Has anyone suggested a rematch yet?
For me it just looks like Alpha Zero doesn't lose anything playing Rf7, because if the opponent responds in the drawish way it can just go back and nothing changed in the position. It is just scouting SF to see if it will answer in the best way possible, cause if it don't maybe there might be some crushing lines behind it
What would be interesting is how fast AO could undisputably solve chess if allowed to play long enough and self learn to the point where it always picks the same move for any of the estimated 10^43 possible chess positions (there is an estimated 10^120 possible chess games). It constantly refines its strategy based on past learnings...it must already be close if not finished if it can beat Stockfish.
The key here is the Alphazero calculates move probabilities, and not just what is the best move for its opponent. So there is always a small but non-zero probability that white will play something else, and thus giving black an advantage. Since Alphazero is not penalized for playing cycles (until it leads to a draw), it is always better to play the cycle and see whether the opponent will make a suboptimal move or not.
Idea of repeating the move ... is to make sure wther yr opponent find the correct sequence .. I have seen many time stockfish doing this to me when i am analysing my games .. even with +4 advantage ..there is a chance that yr opponent might do something wrong that will increase yr advatage and finishing in quick moves ^_^
Here is my take on alpha zero. I watch all its games. From what i have seen here is its tactics: it locks the middle, immobilizes a few of its opponent's minor pieces deep in their ranks, it sacrifices a minor piece or pawns to create files and mobilizes all its minor pieces for positional gains. It can be down one piece or two, but in reality it actually up cause some of its opponents pieces are immobilize or lock. Very clever. It creates multiple traps so that its opponents have an option to die slowly or die fast. Very naughty machine
Something I've noted from watching alphazero is it seems alpha has a tendency to use his knights to force its opponent in zugzwang. As bait, sacrificing the knight to (usually) a bishop, in turn taking bishop.
It is a fan of anatoly Karpov
It's called ACTIVITY
It’s like watching killer whales force their prey to the surface of the ocean where there is no where left for the prey to swim to
It just seems like it just does everything perfectly well....
And stockfish resigned the game... I can never hear enough of this 👏
Alpha Zero has no ego as well.
I don't think he meant it for ego reasons, but rather because it's such a rare phrase. A situation that rarely happens.
it's a super CPU vs a laptop with outdated software
YEAH, when ever I see a deep mind video I insta-click. The last few days I sat down and learned about how deepmind actually works, and it turns out deepmind is learning even when it's playing games against stockfish. This answers your question about the refusing of the draw. Deepmind probably found a good continuation after bying some time with the repeat moves. Keep making deepmind videos please. Deepmind is now my new favorite chess player :)
Beerdy - Bruce Lee Central while it's certainly true that AlphaZero could learn while playing Stockfish, there are two reasons why I think they wouldn't have bothered:
Firstly, it might learn bad moves from stockfish. AlphaZero learns through self play, because it gets the best game data from itself.
Secondly because it would waste computation that could be spent focusing on the match.
It is true though that after the match the game data could be fun through AlphaZero to make it better, but 100 games would be such a small contribution to its training set that I wouldn't see the point. In the four hours, it would have played over 700 billion games with itself.
I teach deep learning, and I agree that they likely didn't train deep mind while playing the matches against stockfish. If they did the games would have little impact in it's play vs the millions of games is has already played.
What is probably happening with the refusals is that A0 is making the move with the next-highest probability of winning.
Hey, i have some questions about deep learning, just basic question to how to learn it. if you have time to answer them, i can give you contact details? shouldnt take more than tops 10 mins
@Nathan Burnham
facebook.com/omer.ulger.397
I'm not sure where you got 700 billion from, but I believe AlphaZero has played only a few million matches against itself. The preprint mentions only 700,000 games and claims that its performance exceeded Stockfish's after just 300,000.
No light square Bishop was harmed in the making of this video
underrated comment
not really. A light square bishop was captured
stockfish: e4?
alphazero: you lose.
lol savage.
@@dannygjk HA!
Just straight up “Lmaooo mate in 211 after e4”
@@deanaraula meanwhile me proving alphazero by getting mated in 7
That’s brutal.
So many alphazero videos despite you had said no more :D keep up the good work we love them all
Erkin Tunca very hard to resist the temptation!
Guess Alpha Zero forced the move :)
Taoh Rihze 😂😂
Agadmator you are an amazing channel thank you for all your work
Thanks Brian
agadmator's Chess Channel damn this was 2 years ago is alpha any better now
I really hope it can't improve further
There is a fitness function in ai that says how good is your solution. It can be as simple as winning = 1, losing = -1. If they set draw as a -0,1 or something similar it gonna refuse the draw by repetition. They should release the progression of learning from those 4 hours. Usually its funny stuff. First games random moves ending in perpetual checks. Then it learns not to draw and come up with some crazy attacks and trades, and eventually it would come up with openings, and basic strategies. Ending in this fish eating beast.
4:26 you can smell the fear in Stockfish repeating his moves
we like stockfish vs alpha zero because it's like watching the chess equivalent of a dragon ball z fight
*Agadmator said,*
"No more AlphaZero vs. Stockfish videos..we may not be able to appreciate human chess.."
You may be a victim of your own words my friend haha.
I guess, Alpha0 was programmed to go for 2-time repetitions whenever possible, in order to induce the horizon effect in typical chess engines that examine the game tree to relatively small depths (usually up to 20 moves ahead in the middlegame). One of the ideas that Matthew Lai (one of the Alpha0 contributors) had expressed in his master thesis is to explore the tree deeper in those variations that are assessed (by a special neural network) as the most likely to be in 'the principal variation' (i.e. to be played if both sides play optimally), which is closer to how humans calculate variations, as opposed to usual chess engines that waste time on a lot of improbable variations and extend the search depth only in very specific 'violent' situations (like captures). Alpha0 uses this 'neural network for move probabilities' approach in its Monte-Carlo tree search (search for 'AlphaZero' on arXiv.org and read that preprint) and sees further than Stockfish in the critical variations that end up appearing on the board.
You know what would be interesting if they give alpha and stockfish different opening positions like nimzo or sicilian or some well known gambit positions like kings gambit etc and let them play
Debjyoti Bose i'd love to see alphazero play the kings gambit :)
IIRC in the paper we have alpha vs itself 100 times for a few popular openings. But they don't show any games only the winrates.
on arxiv
The king's gambit is a forced win for black.
Old man eating a cookie since when?
It's great to see that bully get a taste of its own medicine:)
Alpha tries to win the game every time playing as he knows is the best move. The rook is the best move until the posible threefold repetition.
+Albo Nice
is not just that, in the second set of the three moves repetition, the rook and the bishop end up in the opposite position in comparison with the first set. which i thinks it's what alfa wanted. so you are right, the best move for black is the rook but not until the threefold, it's because of it. Saludos de un hincha cuervo desde huerta grande, córdoba
Hola cuervo, i don't think that AlphaZero ever hopes for a threefold. He just want to avoid it as long as he can find another node that's better than draw.
Albo's native language is Spanish. Since in Spanish Alpha Zero's word 'gender' would be masculine we automatically think of "he". It's a mistake we commonly make.
Cuervo 10 It's not what AlphaZero wanted, but rather it was forced to play a different move as playing the same move gave Stockfish the chance to draw.
Alpha Zero overprotected the e5 pawn and maneuvered the knight at d6 for the blockade. Is Niemzovich correct all along? Even A0 used it.
Maybe Alpha read "My System" :D
My thoughts exactly.
@@agadmator Don't scare us man :P
Really strikes me how visually beautiful alphazeros development is.
I'm a visitor, VERY much enjoying the AlphaZero coverage. Thanks for doing these videos.
ALPHAZERO IS THE FUTURE.
Maybe... but if Alphazero is the future, we humans are the past...
My theory for the threefold repetition thing is that both of the bots are playing their best move, which just so happens to be a repetitive move. The thing is, when they get to the third time that move is no longer the best, because it leads to a draw, so it rethinks its move and does a better one
ah so they do the second best move as the "best move" would lead to a draw otherwise so basically the second best move becomes the best move. i think that kinda makes sense yea
This doesn't really make sense to me. If the second move is better than a draw then how could the first move be the best move if it leads to a draw? That is counterintuitive
@@liljackypaperThe draw only happens after 3 repetitions. The best move doesn't lead to draw in the first 2 times, so playing it in the tiny hope opponent plays less than optimal move, is worth it.
@@gJonii engines don't play like that though. They don't make sub optimal mate in one threats in hopes that opponents miss it
@@liljackypaper If they lose nothing from doing it, why not? Alphazero specifically, being MCTS, would treat the tiny chance of opponent playing wrong worth the extra move. Stockfish I think would treat the moves equally good, with or without mate-in-1 trap
*I'd treat AlphaZero to a drink*
for winning against stockfish as payback for beating me a lot.
Then alpha zero beats you 10x more than stockfish
I am late but the enemy of my enemy is my friend so he would still buy Alpha zero that drink 😂
"Is e4 a refuted opening?" Let's not go crazy, here.
EebstertheGreat Well, Alpha Go is probably the best chess player ever, and it never plays 1.e4
So there could be something to it.
If, back when he was the best player in the world, Paul Morphy had decided to never play 1. e4, that wouldn't have meant he had refuted it. If AlphaZero never plays 1. e4, that may be because it is less successful at that opening, but there are all sorts of reasons why that might be the case beyond it simply being a bad opening.
@@EebstertheGreat if A0 never plays 1. e4 and this is because it assesses it as less successful, this is exactly what a refutation is dude. after all, it never plays 1. e4 :o and in this video you can clearly see why.
@@Pintkonan That is not what a refutation means. A refutation does not mean the world champion is bad at that opening (or marginally less good than d4). A line is refuted if a refutation is found--that is, if a defense is found that proves the line is worse than another move. There is no defense to e4 that has been demonstrated to be successful, it's just that over many games, some engines win more with d4. If you want to be strict about it, by your logic, every opening is refuted except the single opening that Lc0 or Stockfish prefers. And if that ever changes in a better engine, suddenly the opening becomes unrefuted.
It’s an interesting question. To definitively refute any opening, one would have to solve the game
No matter how hard I try to define the game with a suitable word, I am heading for the same word over and over again - "Poetry".
It would be so interesting to force these engines to a certain opening and let them continue from there, just to see how it would turn out. Gotta say I did not expect to like these engine games, but they are pure gold and I´m so excited when I see you have uploaded another one! Keep up the good work, cheers from Finland \o
Isn't that what originally happened? I thought neither engines has opening books?
At 6:41 how did black capture after white moved to f4? They were both on the 4th row next to each other (?)
Am I missing something? Please help, I'm fairly new to chess, excuse my lack of knowledge
Google En passant
Love these alpha zero videos man! keep up the good work :D
2017 - DeepMind AI
2018 - By the time DeepMind became self-aware...
It did. Now it's developing a master plan to conquer Earth.
Such great analysis of Deep Mind's maneuvering of the bishop
Hooray! Alpha zero returns :)
I wonder if there's a way to deconstruct scenarios and outcomes from various playing scenarios by forcing Alpha 0 to play itself and set it's opening sequences (Queen's Indian / Belgian v e4 or others) then release the analysis to discover why... Could be interesting either way it's incredible play and thank you for sharing
I really like the fact that these games are a real grind, and not just ownage in just few moves. For humans this game is great (its well balanced for both sides) and certainly has alot of future!!
In your variation, you can simply play Bc3 with white and you do stop both pawns for a while. It seems that the pound in H will fall but the black king will take d6 and from there is not dificult to win.
Agadmator I love your channel and have watched most of your videos. Am planning to watch the rest of the videos too. I would like you to do videos on Chess openings and discuss the various variations in each main opening. There are few good chess openings videos on RUclips and I am sure the entire community would learn from your videos. Please do chess openings videos.
When refusing a draw, probably, Alpha Zero considers itself better in both positions so the draw is worse than going for the "less good" position.
Yep, this is so obvious and I find it strange that so many people don't get it.
Agreed, alpha repeats moves bc he thinks he has a better position and is trying to 'bait' his opponent or exhaust all chances that his opponent will make a different move and not repeat ... After alpha repeats twice he doesn't want the draw bc , idk, alpha thinks the position is still favorable or playable (in alphas mind!) So he avoids the draw and moves on vs. he always has a draw in the back pocket or he learns a new moves after the same position is repeated.
Hi agadmator. I just want to share, wouldn't it be interesting if Alpha Zero plays against itself? I mean, what if there are the same moves every time? What if white always wins? What would it mean?
That has already be done...
like 1 million times.
Sheppard. That’s exactly how Alpha0 learns.
Draw.
Hey can you provide us with a download link to the chess engine you are using on your computer? Thanks
i would love to see alpha zero playing black with various openings like nimzo defense, scillian, caro-kann, and against some gambot positions too.
What do you think would happen if Google let Alpha Zero train for more than 4 hours? What kind of God would they create if they let it train for weeks????
#suggestion if you can find a Alpha Zero vs Alpha Zero game that would be very interesting.
what i find interesting about the repition breaks by alpha is not the question whether it found a winning variation in the mean time or just trys to win and having the draw in its sleeve all the time but whether it *cares* about not losing a game and going on with a move it considers less optimal because it *wants* to go on.
+agadmator , plz keep making more AlphaZero videos, we need all the games against stockfish
I personally think that alpha chooses the best position, by repeating the position 2 times, and then doing it again, in the end it will be in the same position that it actually wanted to be (because after 2 repeatitions it will be 1 turn away from it's optimal position and after doing it twice it's exactly where it wanted to be)
It circumwents the 3 fold repetition rule so to say to force the opponent to accept a position where they normally would want to draw.
Imagine playing stockfish with black and you're offered a draw.
But, you just think: "Nah, a4 is winning."
I pray the algorithm is developed for commercial distribution because this method of search is beautiful to behold. Could go a long way towards training the next generation of chess players.
Amazing video, thank you! I have some questions: It seems to me from the AlphaZero games and paper that it's power lays in super advanced stratigique thought (or maybe stratigique calculation? Hard to chose words to describe this "entity"). Whereas, from my limited knowledge, Stockfish's (and chess engines in general) strength lays in brute force of calculation. So, added with opening books and different middle and endgame tables, Stockfish is merely "mimicking" stratigique thinking, but isn't actually considering positinal aspects, space usage, flexebilty, activity and synergy. It IS eventualy "taken into consideration" indirectly via brute force, because the consequences of such elements are evident in lines calculated by Stockfish. Against a human or an inferior engine, the force of calculation is enough to "hide" the inability to think/calculate strategy. But it seems this is how it is outplayed by AlphaZero.. Also, Stockfish is engineered by humans to evaluate a position not only by calculating possible lines of play but also through material numeric value. Maybe we, humans, "misled" stockfish by "teaching" it a wrong or incomplete evaluation of material and position process... Maybe AlphaZero can teach us a new way of thinking about material value. Either we will learn that a knight is actually worth 3.5 and a bishop is worth 2.7, for example, or that it's wrong to even go through that line of thinking.
What's also interesting in my opinion, is that SF's brute force makes it a "god" of tactics, as tactics are based on calculation rather than "thought". (They could also be based on 'post-calculation'. A GM doesn't have to always calculate a full process to spot a tactical trap, he/she can train to see it by noticing patterns and structures, or known lines of "theory" based and calculation made by them or someone else (including engines) in the past).
I believe Stockfish is bound to always calculate, and it can't develop these abillities that GM's can. Though, It probably doesnt mind (Pun intended ;) ). It is a preety f***ing good calculator.
But is it possible that AlphaZero DOES develop (like a human would) to recognizes tactis without calculating all the time?
Is it possible AlphaZero is "thinking" strategy in a broad and complex way?
Is it possible Stockfish is yet superior in tactics? Would be interesting to present them both with very complicated chess puzzles to see who is better. (Though probably even AlphaZero's inferior calculation power of 'only' 80,000 positions per second can stand any chess puzzle we humans created, and the gap between SF's and AlphaZero's tactiacal quality - if indeed exsists such a gap - would be insignificant or impossible to notice unless both of them are given only fractions of a second to solve the puzzle.)
I want to add that all the asumpstions I based my thoughts upon could be flase. I am new to chess and know almost nothing about computering and AI tech.
Also, as some people find it somewhat depressing that AlphaZero belittled centuries of game development in 4 hours, I want to add an incourging thought:
Even though AlphaZero outclassed us and our programs so effortlessly, it still isn't capable of INVENTING AND DEVOLPING the game of chess. Or even if it is, if instruced to come up with a game, it can't do so just because it WANTS to and INTRIGUED by it. We still have the ability of doing something for the sake of pure enjoyment going for us. For now. :)
You have won the TL; DR award, my friend יונתן ריבק
Hey Agadmator, did you find any ratings for AlphaZero??
HI AGADMATOR.... i am a huge fan of ur chess commentry...i would like to make a suggestion though...
i think it would be very nice if u could make some space on the screen dedicated for the move (like b7 ,e4 etc) , for dead peices and for the name of the opening or attack(like kings indian defence or scicilian defence etc.)
thank u
It may not be a good idea as it would take the attention away from the action on the board. If you want the moves displayed, then what's the need of commentary?
You're right, Sameer. In the last video someone had requested that the positions (Eg. Knight to c3 etc)NOT be mentioned. I didn't quite understand. I mean how are you supposed to analyze otherwise. But to answer brandybuck, I would suggest this
1. If you needed screen space because it's difficult to grasp notations along with the video: I think I understand the nature of your request. But trust me, everyone has these struggles in the beginning where they find it difficult to comprehend the positions on the board and the notation. It will come with time. Suddenly one day, you'll be talking to yourself in notations (atleast for the first few moves). Suddenly when someone says 1.e4 c5 you'll inherently know it's the Sicilian without even looking at the screen or the board. And that's a beautiful feeling. Most people have this difficulty in the beginning and that might even deter you at times but that's the beauty of chess. But that's exactly why chess games are immortal because we can recreate games from 200 years ago with just notations. They're really powerful but it takes a few hours to get used to them. I sincerely hope you get through this phase and appreciate them, if this was your concern.
2. If your request was just aesthetic in nature, I agree with Sameer in saying that it does take attention away from the board. The description of the video gives the exact moves played in the game. One thing that you can do is to import the pgn if you are analyzing or watch more and more videos so that you do not need any screen space for the moves.
No I agree 29th brandy. If anything have it notated. To his website. Btw need a website lol
If you haven't noticed already, the notations are there in the description. You don't find other popular youtube channels including notations with every move on the screen. And that is for the same reason. There is a strict correlation between what we see and what is perceived by our auditory senses at the same time. While watching a game, it's important to pay attention to the move on the board while the brain automates the position in the brain (as we hear the notation). If it is displayed along with every move, then it definitely becomes a distraction because our eyes would shift between what's happening and what's being displayed. Like I said, I understand it from a beginner's perspective for the sake of convenience but a little effort and patience can't do any harm. You don't find a lot of subscribers requesting this for the same purpose because most people already know that, it doesn't take ages to get used to notations. I do, however agree with you on the point of making a website. That would be cool. Might need Medo's picture as the background :D
Mozisi. ur right. Initially it was difficult for me but now i can understand easily and now i can visualize positions just by hearing them. like your explanation :-)
Your dog is back! Great! =)
At 3:59 & 4:25, does Stockfish really think it can do no better than a draw? Or does it simply anticipate that the opponent will refuse the draw?
Is Bg4 a valid attack in your opinion (6:55)?
Alphazero just had a fish for dinner. loving those videos. his type of thinking is astonishing!!
Hey agadmaster, can you offer general recommendations on the best chess books available with reference to particular authors and publishers.. i would also be interested to hear specific titles that you think are particularly good!
The Berlin defense is better, provided you're a strong enough player to consistently beat Stockfish.
Stockfish is the Soviet Red Army. There is no Berlin defence.
Hi, i study machine learning, i think the 3 move repetition has to do with how self learning neural networks trains themself -- it has probably learned that the same position happening 3 times means a draw -- but if all it was given is the rules of chess it would have never explored past 3 repeated positions because it would have considered that position "known" or solved for -- the end implication is that is is now forced to pick its second best move, which it also probably thinks is winning .
Deepmind Alpha Zero's tight knit structure and movement of play reminds me of a sliding tile puzzle, both interlocking and spiral. A snapshot of Stockfish seems a looser version of the same structure and movement relative to Deepmind Alpha Zero (e.g. 5:58).
To understand why the position repeats but the draw is refused we can look at the algorithm they used to train Alpha Zero. It uses a kind of Monte Carlo method which is a randomization procedure to decide which moves to try next. This means while training if you allow your opponent more opportunities to deviate from the best line you have a higher probability of winning in the position just because you get that extra roll of the dice. I would interpret this behavior as showing that alpha zero felt Stockfish's defense is optimal and that deviation from that line significantly improves Alpha Zero's evaluation of black's position.
6:45 I think that this Bishop x Knight exchange has multiple purposes: strategically that's a bad Bishop in a close position, so it's great to exchange for a centralized knight; also that knight was defending c4, which will now depend on the Queen; as she can't move now, the obvious Alekhine Gun that Stockfish was planning to create at f won't happen.
The reason AlphaZero plays for two repetitions is because it's designed to play the best move for the position on the board.
In those situations, it really is playing the move it thinks is best. And only when the 'best' move would force a draw by three fold repetition does it consider another move.
I think the takeaway you should probably have is that in those positions, AlphaZero prefers the move only if it does not cause a draw. The move it plays instead is its second choice, but still has winning chances.
I was reading about Google's AI was actually acquired from a startup company. The curious thing is that when AZ learned Go by playing against itself for a number of hours . That version actually defeated the AI version that learned from a feed of thousands of previous games. I wonder if AZ learned chess from zero also or it was fed games... Hope Antonio can clarify...
If you look in the PDF paper on this, they did put Alpha Zero against Stockfish in the Ruy Lopez in 100 games, which they did in several openings, however, it's not clear which position it started from, either 3. Bb5 - a3 as in the picture, or 7. Bb3 - 0-0 as in the "PV". Nonetheless, Alpha Zero as black won 6 games, drew 44 and lost 0 from which ever position. As white, won 27, drew 22 and lost 1.
what is the difference between engine approach and alpha0 approach.
#agadmator I'm trying to find the Leela game (or Alpha) were the white queen spent much of her time on H1 but am failing. Can you or someone else point me in the right direction?
I think you mean AZ vs SF SF played QID. One of the QID games. Several RUclips people covered it. Maybe you mean this game? :
ruclips.net/video/NaMs2dBouoQ/видео.html
I wonder if all the drawn games were drawn because of threefold repetition.
7:55 Never has a doubled pawn looked so powerful -- Alpha's pawn fortress...wow.
With respect to AlphaZero *almost* doing three-fold repetition twice in a row: It is very likely that the rook moves (in response to the bishop moves) are in fact the moves that are most likely to lead to a win, if threefold repetition were not a rule. However, AlphaZero is trained to recognize when a move will result in a win, loss, OR draw; the first two repetitions are simply AlphaZero taking the most winning move at the time, but the most winning move changes when it will directly lead to a draw. What is most impressive is that AlphaZero learned the threefold repetition rule; it was not hardcoded into the neural network as it would be for a classical engine. Considering how few games end in threefold repetition, it's truly amazing how DeepMind was able to generate enough games for AlphaZero to learn threefold repetition from scratch.
Would be interesting to see how well Alpha Zero played right after learning the moves... was it instantly a GM with it's calculating ability? I don't think I've seen it make a single bad move yet.
No it was nowhere GM strength until it had played millions of training games.
This is really funny because I have not lost with the Berlin defence jet. 15 matches.
Is there a game where stockfish won against alpha zero? If so, could you show it?
No. There were tons of draws, though- the final count was 28 wins, 72 draws, and 0 loses, if I recall correctly. (However, DeepMind has only released ten games, in all of which AlphaZero won.)
ruclips.net/video/XZotQte2HbY/видео.html alpha lost here. But at japan chess
that's go. "Japan Chess" would be shogi (which alpha zero also played). But in any case, the video is about alphago, a predecessor of alphazero. The publication on alphazero does say, however, that alphazero lost several games to the world's top shogi and go engines
Jonah Albertijn not in the main 100 game series but stockfish won some when alphazero started after a set opening. The results can be found in the paper. I think overall across 1,300 games (100 main games and 100 per opening for the 12 most played openings) stockfish won 4%.
honestly, the only people i watch when it comes to chess commentary and analysis is MatoJelic and Agadmator's chess channel!
They give the most accurate analysis and they are entertaining to listen to and watch!
I first learned about the technology that DAZ uses (parallel distributed processing) in the early 90's and it was revelatory; AI is smarter when it tries to act like a real brain than a computer. I use some of the lessons from this in my personal work (making mistakes is GOOD, as it helps you learn, don't be afraid of making a mess of something, you might learn something!) and now seeing how DAZ makes such interesting, elegant moves, it's very cool to see.
It was given with clear instructions during the initial programming. It works "Minimum defense and Maximum Attack". Stockfish has it opposite way.
That is not how deep learning works.
I am glad there some discussion in the comments regarding the apparent handicaps Stockfish was dealt at the beginning of the match. Has anyone suggested a rematch yet?
For me it just looks like Alpha Zero doesn't lose anything playing Rf7, because if the opponent responds in the drawish way it can just go back and nothing changed in the position. It is just scouting SF to see if it will answer in the best way possible, cause if it don't maybe there might be some crushing lines behind it
What would be interesting is how fast AO could undisputably solve chess if allowed to play long enough and self learn to the point where it always picks the same move for any of the estimated 10^43 possible chess positions (there is an estimated 10^120 possible chess games). It constantly refines its strategy based on past learnings...it must already be close if not finished if it can beat Stockfish.
4:37 Maybe alpha zero is giving up a move to reposition the White bishop, since it placed the knight in a square blocking the bishop.
8:17 After knight to F3, why doesn't black offer a queen trade by picking up the pawn on E4?
LoowheezeBreeze You lose the bishop as it’s attacked by the queen
i don't understand why white would ever go c3 with the pawn, can someone explain. Why would you block your knight like that ever
Thanks for analyzung these Aloha Zero videos. They're quite entertaining.
It's not Aloha zero but rather Alpha Zero
It's fkn unbelievable! AI is going to rule the World and this game shows how!
ur the best chess youtuber :)
Yes, we want them all.
This is now the DeepMind Alpha Zero chess channel.
Curious how many 3-fold repetition draws there were out of the 100 games
I was expecting a bit more at 6:44 regarding the lines with Qxe3 capture and Re2xe3
When alphazero played itself did it ever beat the BD and win?
I thought you'd decided to let us keep enjoying human chess.
Won't complain, these videos are awesome.
7:04 how many moves did alpha make with the king already
7:05 this position reminds me of Alpha's French Defense game, just reversed with the black pieces.
does anyone know which stockfish this is ........like 8 or 9 or 4 or 5....
8
at 2:04 why not black go pawn f6?
i was looking forward to see where it is going to blow today
Helrlo everyone! Love your videos and accent.
I sent $10.
10:57 why would u move the bishop cause the kind can't capture cause lose 2 pawns
i fucking love your dry humor 0:09
The key here is the Alphazero calculates move probabilities, and not just what is the best move for its opponent. So there is always a small but non-zero probability that white will play something else, and thus giving black an advantage. Since Alphazero is not penalized for playing cycles (until it leads to a draw), it is always better to play the cycle and see whether the opponent will make a suboptimal move or not.
"And in this position, stock-fish resigned" Kreygasm
what i want to know is, does Stockfish dispute AI?
more alpha and stockfish matches pls
Idea of repeating the move ... is to make sure wther yr opponent find the correct sequence .. I have seen many time stockfish doing this to me when i am analysing my games .. even with +4 advantage ..there is a chance that yr opponent might do something wrong that will increase yr advatage and finishing in quick moves ^_^
At 3:57 it was not Rh4 it was Bg4.
I mean Bh4.
Great video!