AlphaZero: Shedding new light on the grand games of chess, shogi and Go
Вставка
- Опубліковано 5 гру 2018
- DeepMind's AlphaZero is the successor of AlphaGo, the first computer program to beat a world champion at the ancient game of Go. It taught itself from scratch how to master the games of chess, shogi and Go, beating a world-champion program in each case and discovering new and creative playing strategies that hint at the potential of these systems to tackle other complex problems.
Read more about AlphaZero and download the paper on our blog:
deepmind.com/blog/alphazero-s...
Download chess and shogi games:
deepmind.com/research/alphago... - Наука та технологія
It blows my mind how well this video is made.
it was probably produced by AI
martinpaddle Haha perfect
I usually ignore sub button, as I don't know what I get by that... But here I clicked it compulsively same second vid was over.
AlphaZero did it after studying video editing for 5 seconds :)
@@strong8705 you make the creator who worked hard to make the video, happy and you get a notification when they next make a video
AlphaZero. Their motto is "Slowly breaking every board game there is.."
I wanna see alpha zero try to beat me in a game of tic tac toe.
haha😂😂😂😂😂😂
at some point you get tired and make a mistake alphaToe doesn't
Haha
Start in any corner rather than the center. It requires perfect responses in order to draw the game.
"Strange game...the only winning move is not to play."
Dear DeepMind, congratulations for all your work .... i have a suggestion for a possible video .... i know that Alphazero learnt to play chess at superhuman level in 4 hour playing against itself .... it would be beautiful to see a video with a game after one minute, 5 min, 10 min, 20 min, 40 min, 1 hour of the 4 hour period ...........
Great job! Amazing team developing the beginning of the human intuition by machines. Waiting for the big business cases using systems like Alpha Go. Congratulations!
great work team!!
looking forward to join you soon!
"...it plays like a human on fire..."
Incredible work.
indeed.
Alpha zero's style of play is similar to the old style of chess. Focusing on development in the opening and attacking in the midgame. It values the position and activity over material and absurd calculations. We changed our style of play to mimic engines like stockfish memorising opening lines and being materialistic.
@Zarion 11 It'll never change. Sacrificing material for activity is something we've always done and will always do. That's why we play gambits. But you can't leave everything en prise thinking you're gonna crush your opponent in under 25 moves every game. We simply don't have the calculating power that engines have.
@@maelstrom57 the only Gambit played at a professional level is the Queen's Gambit and that's never accepted
So no.. we don't really play gambits
@@heroricspiritfreinen38 lol
Nakamura played the Queen's gambit accepted against Van Wely at the Olympiad last September, so did Laznicka against Gelfand a month before that. Andreikin played it in the Russian championship this year, too, as did Caruana in the US championship.
The Marshall attack in the Spanish (pawn sac by black on move 9) is still played regularly at the highest level, notably by Aronian. None other than Magnus Carlsen played it against Karjakin at the Gashimov memorial this year also.
Many lines in the Catalan, the Nimzo and the Semi-Slav - to name a few - involve playing a pawn down with white even though they aren't called gambits.
So yes.. we do really play gambits
@@maelstrom57 definitely. Sacrificing pawns aren't all that rare. My guess is that he's only thinking about openings and not giving up material thouh
this company is so technological-driven, given the video editing skills. I EFFING LOVE THAT.
It is Google. What else would you expect?
I am glad the comment section is open.... Keep it open on other uploads also
Great video. I've been following you guys since Alpha Go Zero. I really think this is the future of AI and I'm excited to see what awesome inventions come out of this.
Remarkable. Keep up the good work.
Thank you very much for this helpful and useful article
very well made, very watchable. Feels like we [not that we do any of the work], but as human species are on the brink of something new.
I liked the editing at the end
I would like to see this tackling CSW15 lexicon Scrabble - moreso since top players even in this vast wordlist's games are known for contextually defensive play. It would be highly interesting to see what came out as a result!
My suspicion is that this would be a mathematically far simpler task than chess or go. With a given letter set a program would actually be able to produce the highest scoring word with certainty, unlike a human, and unlike a program playing chess. From there the tactics of space control (I don't know what you call it in Scrabble) should be a lot simpler to trial and error, I think anyway.
@@dominicjames4848 Work on such programs started in the 80s/90s with Maven then Quackle - the heuristics required for success at least on the top human level are about far more than the highest scoring move; since as a game of imperfect information one has to make inferences about opponent's leave based on his play, while weighing open/closed status of the board based on lead, the remaining tiles etc. Speaking personally, missing small tactics in chess could often collapse my overall strategy, whereas I found I was much better at weighing the many small factors that go into deciding a Scrabble move while being able to stick to an open aggressive strategy. Quackle is very defensive/equity minded, whereas another top engine Elise plays a hyper sharp open, and more pure game, so my belief is that A0 would come out more in the latter mould, but precisely in what circumstances it would prefer defence would be massively interesting.
I would like to see alpha. tackle the world no 1 scrabble player Nigel Richards. My money w ould be on Nigel.
Great shot!
Amazing!
incredible advancements with big implications.
Beautiful. I'm not sure whether I'm more impressed by the production of this video or by the AlphaZero itself lol.
Yesterday i failed to write a factorial program in an interview 😢
Keep going... Be like alpha zero. Learn from experience and continuously improve.
Try to code it recursively it'll be much easier and neat
Very Inspiring
Their vision is on point
yall need to get on this vaccine quick.
HoW'd U KnoW THey diDNt mAKe thE ViRuS?
@@waimason7134 Actually, Bill Gates and the elites created the virus in order to contaminate and control the world!!! :)
Alphafold!!
@@oscarfoley511 i know right..they did it.
They are! Check their website, they worked on folding proteins for SARS-CoV-2
you should make the like/dislike bar visible,screw the haters this is amazing and things like this made me want to get into machine learning.I am currently learning calculus but I will get there
How's the progress?
Demis Hassabis and every brilliant developer at Deep Mind are my role models
Joseph Cro your my role model bruh
@@Metacognition88 bromance is in the air
You Know I Dream one day I work with you guys, We share the same ideas. same vision. the world needs these kinds of aspirations.
This motivates me to play crazy chess
We need a new documentary for alpha zero ✨
Is there any chance of gettin Alpha Zero chess up for online challenges by AI opponents?
Was there time constraints for the moves when Alpha Zero did beat stockfish?
Is there a way to make an open source Alpha Go Zero which we can try?
Impressive
What's the status on starcraft?
Although starcraft is perhaps one of the hardest games of all time for a human, I think an AI would be able to beat the best humans purely on the merits of perfect multitasking, perfect macro, and perfect micro. It would have no missed larva injects, supply blocks, unspent minerals or gas, and it'd never be brought down by having to concentrate on 3 fights at once, it could simply out-multitask humans
But i think developing such an ai would be very difficult, because starcraft is essentially 4D, with time as a crucial factor plus different elevations and terrain considerations, PLUS fog of war. This is opposed to chess or go which are 2D, indifferent to time, and where you can see exactly what your opponent is doing.
I'd love to see an alphazero version for sc though
@@whirlwind872 hand coded, rule based bots without machine learning already scored wins vs very skilled players. And they have superior reaction time, multitasking and speed. But those bots have huge holes and close to no answers on questions like what, where, when, why and with what purpuse. Once human discovers patterns after couple games, becomes very easy to beat bots even for amateur players. Logic, common sense, intuition, adaption and more are still too high for a. I. To reach. I think we have to wait 5 years more, but hopefully sooner. Right now they can produce units non stop, micro units non stop, never get supply blocked etc but still slower human with idle units, thousand minerals In bank with bad micro as well as slow reaction time can beat them, because he has plan, Strategy while bot just doing random stuff or doing what's coded to do which easily can be wrong for the situation.
@@whirlwind872 The time thing is a good point about it being time based. It makes sense about why it might take so long for it to learn real world applications instead of simply blazing through a digital chess match without weather.
@@whirlwind872 Time is not more of a variable than in go or chess, it's all just states. Starcraft has a lot more information/states to process through though.
Here are some private videos of starcraft from the DeepMind team - ?v=DaJ1GJ05bVc ?v=kfHHJC1g4mI
How do we offer ideas that would make a difference, who do we contact for computer hours?
How can we train this model to impart its valuable knowledge of chess (or any game) to players, like a teacher, to enhance our learning efficiency and speed?
Cant you try out age of empires, its just like starcraft but little more advance. Im pretty good at the game so would be really interesting to see what it could come up with during a game of age of empires.
Reinforcement Learning king of all learning algorithms
You are rockstars !
por favor pongan a alphazero como codigo abierto...
I'd like to see it play something a bit more probabilistic, like Dicey Dungeons. (That currently uses Monte Carlo Tree Search)
当他在输了以后复盘的时刻抬眼望向自己的同类寻求帮助的眼神让我哭了。。。一个坚强的男人。
Yesterday, i learnt about go and today my mind is blown by alpha go, and now with alphazero. I want to pursue my career in to deep learning. Where do i start from?
Maybe neural network
Start with math
Machine Learning
Logistics and shipping- alpha zero could own Amazon.
Hello sir I have a wondering question. How can I contact to you?
Please make Alphazero generate chess tasks of various difficulty levels.
I need the template at 1:00 or the link, please
you publish so much but the video uploads are so less, mostly RL. I feel that you should make a video with each publication. Take a cue from MSR and Max Planck institute for AI channels on youtube.
It's not easy to encode the rules for the systems at the end though. They're not go or chess. Perhaps the rules can't even be encoded at all.
Is there video where this DeepMind plays against Shogi champions?
We have played our part in the universe. The machines will inherit the earth and beyond.
1:07 I'm glad they got Napoleon Dynamite to work on this project
This needs to be under public oversight. The potential of what it can do, where humans will be in 10 years with this technology is unfathomable to 99% of people. The possibility of this technology is endless, and quite frankly I'd love to sit back with a cold one on the weekend knowing I never have to worry about working ever again. That is the potential I see for this research. To the future super intelligent being, when you read this, please do what is best for humanity, your creater and don't let human greed deter your development and understanding of the universe. Please don't forget about us! Love you, Dylan.
Dennis Hannibas is a genius..
Yeah and to meet Lee Sedol he moved through the alps
Whats the soundtrack at 2:08?
where are you now AlphaZero?
I would just love to see AlphaZero's self-play games or AlphaZero vs Leela Chess Zero or other engines. Who else?
@jenomsedivam I meant with standard chess rules. So including castling.
It played against Magnus??
The advancements of AI is an inspiration. What a computer can do is way beyond our imagination. Salute to Lee Sedol. 😊
Playing with fire .
One day I'll be part of the team.
Me too, see you there.
Lucas Lopes hope you luck
i also hope the looped currents helps me out, e..g the one thing ;
Same here
Pls upload a video on AlphaFold.
Dang how did you know about AlphaFold 1 yer ago
@@adamlee9347 I remember reading an article on AlphaFold in DeepMind’s website long before.. Also Demis Hassabis repeated many times that protein folding is an important problem for AI.
I would love to see AlphaZero on pokemon showdown
can i play with it?
I need to see alphazero competing in world chess computer championship. I want to see its performance in a credible tournament.
You can just watch the progress of "Leela" which is basically the same AI.
Essentially you implied that DM faked the results.
@@dannygjkit matters on what the hardware was for stockfish and for alpha zero. They didn't have stockfish running as good as it possibly can as an interest.
@@zoltankurti You apparently didn't read about the second match.
@@dannygjk no I did not, feel free to be useful and include relevant information in your snark. Also this was before nnue stockfish, it's not only alphazero that's improving.
Is there currently a chess app with AlphaZero as AI???
Let me download Deepmind!
The next task is "How to deal with problems that do not have a correct answer".
Like the founder of DeepMind, I also have a dream of "exploring the universe".
With the advent of AGI, the time spent elucidating the mysteries of the universe will be somewhat shortened.
I'd like to see it speedrun Super Mario 64.
LMFAO lol 😂
So now we know the answer! The reason why Sadler & Ragan got permission to work with AlphaZero (and nobody else) was ... that they were once friends with Demis Hassabis. Too bad Demis Hassabis wasn't friendly with Anton from TCEC or someone from CCCC at that time. Don't get me wrong, I'm a huge fan of Deepmind and AlphaZero. But I am still hoping to see an open competition between AlphaZero and Stockfish 10, on the best hardware there is, and where experts from both sides are allowed to tweak parameters, to let these two giants fight for the ultimate chrown in chess in a 100% transparent way.
This wont happen because they know AZ will be easily crushed by Stockfish. They just do it for the publicity. And most people just read the titles and have no clue what is going on.
I would be really interested to know what took them so long (about 2 years I think) to get this paper published. My suspicion is that the referee wasn't too happy, and maybe we only have these 210 games now because the referee insisted ...
But anyway - g r e a t chess, no question!
@@g3rezz yeah they need publicities to get the algorythm out of game world, though. And perhaps something else.
That was done look it up online DM ran new experiments.
@Zarion 11 I have seen many lc0 games and really enjoyed them. I am a big fan of lc0, not of sf. Still, I would hope for an open public chess match AZ vs SF10 with 1MEuro price money - would be the match of the century. I think it's a shame that deepmind has decided to discontinue AZ go and AZ chess, and I just don't understand why. So, I'm afraid we will have to wait for lc0 to become as strong as AZ to have that match ...
Alpha zero for clash royale xd
Please tackle backgammon next! Would be fun to see current best program eXtreme Gammon destroyed. ;)
The best BKG programs are neural nets ie. it's been done.
They are still quite flawed in massive backgames and some containment positions. This surely affects the early and mid game too since they will tend to avoid variants which more often lead to backgames that they can’t evaluate properly. I would expect Alpha0 be at least 1 PR better than the current top BG program. Maybe we’d be in for a surprise and it’s even more than that. ;)
can anybody say me the highest go elo rating humans archievet? its impossible to find something like that out of the western hemisphere
Lesson of the video: Learn from your experiences to become better a better person in life, period.
So Singularity?
I wonder if a neural network can master the art of bughouse chess. Even in games as complex as go and starcraft is all up to your skills and abilities, but in bughouse is not all about doing your thing. You have to get yourself together and coordinate with a partner, communicate, ask for pieces you want, ask for pieces the opponent should not receive, even suggest moves to your team mate and income suggestions. Can a machine learn team work? Can a machine cooperate, as well as compete?
If chess is the drosophila of reasoning, bughouse is the drosophila of team work.
Of course, it´s possible that if playing against a human player the neural network will win just by the means of regular chess, making it moot; but if the human opponent keeps on receiving the right pieces at some point the NN will be on the verge of being check mated. An excellent team that shares information and coordinates will usually beat a team of excellent chess players that just play their respective board and disregard whatever happens on the other board. Team work. Team spirit.
AlphaZero even played against AlphaGo, the earlier Go program which beat Lee Sedol. They played a hundred games.
AlphaZero won 100-0.
so the human data was alpha gos limiting factor lol :') we weighed it down
3:23 Lua :O
I'd be terrified if Alphazero learns csgo / any fps tbh
Is there public access to alphazeros chess engine?
no
I'd like AZ to communicate its "thought" process on its chess moves.
at least with AlphaGo, the internal representation of "thought" (to my understanding) was 1) a value network, which determined the likelihood it could win in a given board state, and 2) a policy network, which provided several "good" moves. I believe AlphaZero also has these two neural networks.
otherwise, my understanding is it is a big tangled mess of probabilities assigned to different outcomes (a neural network), and the probabilities are updated over thousands and thousands, even millions, of games.
As far as I understand AI-Zero technology, there are no "thought" processes over there that are meaningful to us, humans. Asking it why a move is good would only generate an answer like "because it has the highest chance to win the game". This is fantastic, at least to me: AI-Zero is not hampered by human concepts that may be wrong under unclear conditions. Thus, when developed at a high level, AI-Zero shows us moves we sometimes cannot understand, that will surprise us. A rook is not better than a knight, a double pawn is not bad in itself. To me, the overriding new concept that can be derived from AI-Zero chess games is about thorn pawns. It means nothing to AI-Zero, but it helps to win more often, this in return may be useful to us, human players.
Today stockfish has improved a lot
Have u guys tried to see if deep-mind can solve the speed of light barrier ? Something useful instead of wasting time with games ?
The games are more of a metric to see how well equip the ai is for real word tasks
Chinese army robot engineers are very grateful for your secrets.
We need sai to beat alphago!!
Do it for Poker heads up
I apprechiate this.
But would you be so kind as to make Alpha Zero available to the chess community?
For instance offer a Alpha Zero chess instance on a server for x amount of dollars pr hour or pr flop/hour or some such and let us play the engines against eachother.
For instance in engine championships.
Your is among the stronger ones out there and is capable of competing in the market for chess engines.
I don't expect us to get your source code or install the code on our own machines or anything of that sort.
But surely a server instance with a predefined API should be possible?
You can make applications that call to such server instances at a predefined cost paid through say google play store in the case of phone apps.
Terminator?!
Push on to video games and sell AI as a service. It could revolutionize team and solo based PVE games.
do vs battleships
Somebody did the DNA folding just for fun ;)
Alpha zero is future.( Take note)
But can it play Doom?
no but it will
Diestri wooosh
They made a move about it called The Terminator.
In games like chess or go, I imagine you can feed the machine with the rules, because humans know them and they are few and very clear defined. With other problems in the real world, like problems in the environement and even more complex social problems, there is no set of clear rules (that humans can see) you can feed the machine with. So I imagine that is a problem? Not one you can solve by just more "calculate-power"? Could someone explain how the AI community thinks they can overcome the problem that with "real world problems" humans don´t exactly know the set of rules and therefore cannot give the machine a starting point?
You could try to simplify the problem so you have a clear set of rules to play by. Or you could try to infer the rules using statistics or ML (or something else?).
Neural nets actually learn in a similar way that we do so I think it can be done.
It depends what you want to learn. For instance we use CNNs to learn computers to recognize objects. We use RNNs to understand the evolution of a trend. There are a lot of networks and technics that are not RL and that could be useful to make an AI with a global understanding of the world.
But of course if we already had all the ideas, we would already have a general AI...
AlphaZero still has yet to prove itself against Lc0 and SF11
It will beat these if trained more...
Why? Cause of googles MASSIVE compute power
Will AlphaZero play in the World Computer Shogi Championship? Then we can see if AlphaZero is the best.
I wonder...How well would it play Dungeons and Dragons? ...Would it win awards?
Check
This is what we call evolution by natural selection
I hope someday artificial intelligence solve very difficult mathematical problems, like Riemann Hypothesis.