Outrageous Artificial Intelligence: (Game 1) DeepMind’s AlphaZero crushes Stockfish Chess WC

Outrageous Artificial Intelligence: (Game 1) DeepMind’s AlphaZero crushes Stockfish Chess WC


Hi all. I have a very interesting game
from the realms of computer chess to show – DeepMind
which was the company that Google bought from Demis Hassabis who was actually by
the way someone I knew from from school for a while who was himself a great
chess enthusiast but he left Chess to focus more on programming and
game writing and all sorts of things after that. Artificial Intelligence (AI) was a big interest so
he founded DeepMind that was bought by Google and now they defeated the Go world champion using some sort of AI Neural Network – that they developed and more recently
there’s a Research paper which has just been released the alpha0 algorithm developed
by google and deepmind apparently took only four hours of playing itself and it
taught itself how to beat the world computer champion Stockfish and it kind
of beat stockfish twenty eight to zero in a hundred game match but I think the
time limit might be pretty short like one minute each for these games but
nevertheless it’s pretty shocking news for the world of chess AI as I can tell
because I I believe that there wasn’t much scope for improvement outside of
the incremental brute-force search better Hardware was the primary driver
but I guess the heuristics that chess engine
writers have been evolving over the years and the other things they’ve been
doing to try and make them their chess programs better and better that’s been
just super surpassed by this approach of I think it’s called reinforcement
learning it just teaches itself with reinforcement learning and so a neural
network teaching itself but with it in more generic search algorithm anyway
let’s have a look at one of the games so deep mind’s playing black against
stockfish so e4 e5 Knight f3 knight c6 we have Bishop b5 Knight f6
so it does like the Berlin defence it hasn’t been given anything no opening
book no end games no nothing it’s just been taught itself and it’s arrived at
this conclusion this is a good opening so the super Grandmasters might be right after
all it’s kind of validated a lot of the popular Chess openings as well in what it
plays d3 Bishop c5 white took on c6 now White Castles and White is attacking
that pawn now that’s protected Nbd2 black castles Queen e1 so f6 some
pressure on e5 is released – reinforced here Knight c4 and now rook
f7 which gives f8 for the Bishop a4 the bishop goes back here rather than say on
this diagonal King h1 Nc5. The Knights preparing to reroute a5 Ne6. It seems you
know very human-like play at the moment now this sacrifice looks a little bit
dodgy you might think but it looks like one of the moves which can promise to
give white something because there’s a lot of pressure on the black position
and in fact in this position now so it’s 2 points for the Knights you know White has got this kind of rolling pawn Center potentially and here after Ne5 – a
repetition is avoided actually with a6 when we see another point
that this whole structure is being attacked
c5 f4 so there’s definitely some compensation for that piece sac Queen e8
a takes bishop takes a now Queen a5 and you can see pawns double attacked. Well there’s a lot of pressure on blacks position this protects at the moment
there’s a lot pressure on that position Knight d4 Queen drops back and now yeah
with this pressure here we see okay rook e6 protects that pawn bishop Eve free is
played if this comes back Bishop III rook b6 okay we have light c4
now rook b4 be free a five-star offering a pawn here actually and all the pauses
quite a few pawns Knight takes Bishop a sex now Bishop takes rotate so let’s
examine the material situation it says the bishop power versus night and four
pawns are they three four three four five six seven quite a lot so at night
in free pawns at the moment but these are just doubled pawns as well these are
not great pawns here so we see Knight see for routine ain’t G free h6 and
there’s a hint with h6 that’s something like g5 might happen later Queen a5 now
it becomes four pawns offs this yeah because it’s double attack on a six and
c5 and black doesn’t want to relinquish the bishop pair so an interesting
decision here what’s a bishop c8 it’s a it’s a lot of home-button dust so what
can the bishop pair do here while some escape squares for the King
and taken with this as well so eyeing those light squares now rook
d7 Queen e5 and then in fact with this the Queen’s just come off because blacks
got that iPhone out but Oh Kim black create some sort of attack here night c4
g5 so the bishop can use this diagonal now rook c1 Bishop g7 95 rook hates
waiting around there for a moment by a free Bishop b2 I wish it drops here so
you’re not committing the rock and the rock can use f8 square in some cases
like g1 Pusha drops back to d7 92 Bishop d2 yeah
Bishop e3 keeping the pressure on White’s pawn chain and now Bishop g4 so
the bishops are working well together there now Bishop d2 and you can see that
this will this infiltration black starting to put so much pressure up
that’s gonna win some material soon after H free yeah rook Bishop takes e to
rock have to black gives back piece here now if G takes them rook takes c2 for my
bishop d4 all the pawns are going to go so we have this one rook takes easy but
this leaves quite a dominating dark square bishop all these pawns on light
squares this looks a bit silly comical in fact you know this this extra bishop
is actually quite post on air and there looks quite aggressive King g7 g for
another pollen lights ground this dark square bishop all these pawns on light
squares so the bishop has a free rein on the dark squares and the King comes up
traveling across dark squares that looks like a desperate pawn sacrifice off to
e5 gets taken or okay one we have two ends yeah Black’s making inroads here
check and these poems are going to be dropping off the site should drop off
leaving black simply our Bishop up yeah simply a bishop up here and so it’s
just a matter of consolidating really the position and here the oppressor or
maybe the computer just resigns itself I think probably the operators resigned on
White’s behalf so this game is just an example of something gigantic basically
called reinforcement learning I had assumed wrongly that the kind of
techniques used to beat that go while champions wouldn’t be used in chess I
was wrong apparently self learning has got quite a lot of potential it seems
that stockfish was crushed is just remarkable
using a combination of genetic algorithm and more generic sort of searching
algorithm it’s phenomenal stuff it’s it’s another landmark
it’s another checkpoint in AI history generally not just chess AI but a AI has
used chess as a vehicle for many years for explorations as a perfect
information game so it’s remarkable I want to show other games from this match
and see what we can learn about opening choices middle game ideas etc from this
one it seems it echoes some advice told that pieces are often worth more than
pawns and you can see that the pieces have more potential is interesting that
black rot would rather give up upon than lose the bishop pair at one point only
one later it cashed out one of its bishops entirely to do a lot of damage
and stop that insert to get all the pawns back so I hope you got something
from this and it’s a fascinating landmark example gained from the recent
paper which I’ll put in the description of the video you might might want to
check out commerce questions like shares appreciated patronage

100 thoughts on “Outrageous Artificial Intelligence: (Game 1) DeepMind’s AlphaZero crushes Stockfish Chess WC

  1. there is couple of things to consider, which certainly make the match conditions a bit unfair. Stockfish was set to 1GB hash size only, which is rather low. Also, it ran on 64 core machine, whereas the AlphaZero ran on 4 TPU, which is an equivalent of around 2000 Haswell cores. in any case, this doesn't take anything away from the fact, that the whole concept of machine learning works also for chess (and most certainly for many other disciplines).

  2. It took humanity 2000 years to shape and master chess but this program used a few hours to learn it and master it to crush Stockfish.

  3. This is a more momentous development than when we saw an IBM computer defeat Kasparov in 1997, and it's not appreciated yet, even by chessplayers.
    We are truly seeing the dawn of AI and yes, it's frightening.
    The performance of AlphaZero in this match against Stockfish was just incredible, astounding!!

  4. Chess playing AI-entities are way to important to leave entirely to Google and other tech giants. On a smaller scale some open source projects need to be set up so that these results can be replicated on a less ambitious scale and using hardware that can be freely bought in normal computer shops.

  5. As soon as I saw the news about AlphaZero I knew you were going to cover it, and was looking forward to this video… 😀

    I had followed the AlphaGo news with great interest about a year ago, and I was hoping that we'd eventually see such a strong "intuitive" AI for chess as well. It brings a different style of analysis compared to conventional tree-pruning engines, one that's both fascinating and more difficult (perhaps impossible) to interpret and learn from.

    I wonder how it'll change analysis for the pros.

    Thank you, Tryfon!

  6. Just a correction on the game timings – for the 100 game match, they set the time to 1 minute per move, not per game 🙂

  7. If what we're reading about Alpha Zero crushing Stockfish over the last 48 hours is true – and it looks like it is – then we have advanced to a whole new level in chess engines, to say nothing of other areas in life where AI may begin to take over.

    But as both a chess and a chess-engine fan, I think that history will record early December of 2017 to be the point when computer chess leapt forward into a whole new paradigm. If this kind of software can be united in the future with something like Quantum computing hardware, then it could ironically destroy chess as we've known it, because chess will simply be "solved". The more exciting possibility is that it will open up huge new avenues of evolution in chess, which human players themselves can leverage off, and perhaps super grandmasters of the future will be playing at well above 3000!

  8. I'll be impressed when it can win with classic time controls. This is a $40 million supercomputer vs a obsolete laptop version of Stockfish

  9. We can never play at this level but it seems to play in a more human way against stockfish. Maybe we can learn to play stylistically like this.

  10. Great Video! I'm not sure how to feel about the power of A.I on our future. After this demonstration, Stephen Hawkins warnings about A.I being a a danger to mankind seems a little more plausible!

  11. Everyone dramatising as hell. They played alpha zero on supercomputer against stockfish which runs on 64 bit laptop. I dont really care but if stockfish was played on good comp as well this would end differently.

  12. One thing to point out though is that you need a supercomputer to run AlphaZero (even for just playing. Training requires even more). Stockfish also works fine on a PC.

  13. Self teaching ai in four hours?? I dont feel amused, i feel terrified! Is there a security precaution about this machine?? Careful in developing such things you guys!

  14. Whats pretty interesting is that Alpha Zero is using strategys like grandmasters used from before 100 years and gets away with it…sometimes it even moves the same piece multiple times in the opening and ignores for a while the development of his pieces but inflicting structural damage around the Kings castled pawn defense.

    The question is how it will fare against other engines or with a bigger time window

  15. This is amazing. Chess playing programs normally use exhaustive search and piece evaluation. This system can not work by assigning values to pieces because there is nothing in the rules of chess which says this.

  16. I want it to play against the best ICCF players. That organization allows the use of engines, and engines guided by humans have proven to also be far stronger than an engine working on its own.

  17. We have to assume that all deterministic games, which can be defined in a language like GDL, are utterly pwned by Google's Deep Mind AlphaZero … or will be within a year. https://en.wikipedia.org/wiki/Game_Description_Language
    Am sure that Prisoner's Dilemma and Poker-like games can be fed en-mass to AphaZero by some sort of language as well. What happens when the CCP's 50-cent army get this tech?

  18. Wasn't Stockfish heavily handicapped computationally? A0 (Alpha Zero) had access to 4 TPUs (approximately 400-2000 CPUs) while SF had access to 64 CPUs and 1 GB of hash. Not to mention the version of SF they used is not the most powerful version. (asmFish/Stockfish9Dev holds that title) Anyways if google decides to go up against the most powerful chess engine why not use the most powerful version on equal terms?

  19. I've followed computer chess for a long time. We've always known that the search was too wide, but a wide search ruthlessly orchestrated mapped so well onto commodity PC hardware, it was hard to win a competition taking any other approach, and without winning (or reasonable prospect thereof), where does your creative energy come from? Success breeds success, and collective ruts are not unexpected.

    We could have created a chess tournament where competitors were allowed to propose their own custom chip designs which were simulating to run at the speed of real hardware. (And then the whole show would have been taken over by VLSI wonks contesting the true values of simulation handicaps, driving out anyone else with a different passion.) Such an exercise would have, however, come closer to revealing electronic chess truth.

    Google actually made real TPU chips—very expensive to engineer, though hardly more so than a standard Xeon chip—because the scale of their data reached a tipping point where it justified the investment. The problem with simulating a TPU is that your chess tournaments now run at 0.01 × real time (i.e. every second on the chess clock corresponds to 100 seconds of wall clock time in simulation, for everything except Stockfish-style solutions optimized for real hardware). We didn't go this route. Instead, we ended up exploiting what a Xeon is really good at. What is a Xeon really good at? Cache agility, aka enormous transposition tables. That immediately biases the communal approach toward search over evaluation. And from there, as I said once already, success breeds success.

    I've always understood that reinforcement learning could optimize gradient locally. The trick is pushing the gradient from the final outcome all the way back to the opening move in a principled way. Man. Problem solved. The second trick is narrowing search without remorse (e.g. filtering out the killer rebuttal). In so far as chess experts can introspect, we've always suspected that chess had to amenable to fairly narrow search without remorse, as no earthly human has ever claimed to do more than that.

    It remained plausible (until yesterday) that human experts were exposing themselves to far more remorse than we realized—being mere patzers of flesh and blood—and that the lack of remorse of a wide search would put just enough on the table to keep wide search in the pole position.

    The one stunning revelation here is that AlphaZero seems not to suffer much—if any—remorse when playing a wide-wide-wide program like Stockfish (which traditionally eats remorse for breakfast).

    Apparently, remorse in chess—with sufficient pattern recognition—does not have a fat or long tail. That's new, and a super-important empirical discovery that will change the shape of computer chess forevermore—unless a surprising counter-punch develops which restores balance to the force.

    How will that play out? Well, Stockfish will punch back, but with the AlphaZero learning curve seemingly beyond ferocious, this is shaping up to be Stockfish pissing into a gale-force willy-nilly wand. (You found a tiny edge? No problem—I can change the global search-breadth hyper-parameter again after lunch—tada!—and then reboot chess knowledge from scratch, without missing the first crescent drip of happy hour wetting my prim coaster.)

    As Mark Twain once said, never get into a turf war with a man who souses his lush clover by the sweet alembic sweat beading down the twisted bonsai stem of a gaudy algid goblet.

    https://www.easternleaf.com/v/vspfiles/photos/801620-03-2.jpg — bonsai with helical stem
    https://images.promotionsonly.com.au/product/twist-148ml-martini-glass.jpg — helical stemware
    https://www.merriam-webster.com/dictionary/algid
    http://drinks.seriouseats.com/2013/09/print/cocktail-glossary-parts-of-a-pot-still-distillation-terms-dictionary.html

  20. Soon, AI will beat us in everything we can think of. Every game, every sport, and anyone using this in a war will be unbeatable. But of course it will not be afraid to sacrifice a few million humans in order to win.

  21. Thanks for the video. Just a minor nitpick – Alpha Zero actually trained for 9 hours before these games with stockfish, not 4 hours.

    See Page 15: https://arxiv.org/pdf/1712.01815.pdf

  22. Deep learning is gonna be huge in a couple of years. Computer (or machine) are going to be better than human in a lot domains….

  23. This is an incredible. And I have seen some information about AlphaGoZero and its programming. The implications here for the development of artificial intelligence are remarkable. If there is a Nobel Prize for computer science these guys should win it.

  24. Alphazero playlist: https://www.youtube.com/playlist?list=PL9JCz2Gsbqe6bfqKQOFXDLJIsZcjHWD8-
    Replayable game link: https://www.chessworld.net/chessclubs/ltpgnviewer32/ltpgnboard.asp?GameID=4771516&v=M3B318ybLqA – Cheers, K

  25. Great, thanks KC, I was hoping you'd lend your insightful and passionate commentary to these games. Excited to watch them. You are hands down the best chess commentator.

  26. wtf, i have been watching youtube for 10 years, i am glad i found your channel. You are just a guy which is really awesome in terms of passion and discovery. What amazing commentary. THIS IS WHAT I HAVE BEEN LOOKING FOR

  27. It wasn't surprising to me. I was surprised it didn't happen earlier because the potential of neural nets was known long before the Deep Mind Alpha series. It's just that in the past we didn't have enough computer power to do the neural net training phase in a reasonable amount of time.

  28. I think the reason why Alpha Zero beats Stockfish 8. Is One they took away its Book Openings .Second No hash,Stockfish 8 was given 1 GB hash table size When 32 or 64 or More .The hardware could make a great difference . You need to have at least 1024 of hash in order for Stockfish 8 to play very strong . .No endgame table bases. The pounder is off. You need to run Stockfish 8 on at least 64 cores . Mean while Alpha Zero is running on a super
    computers. Stockfish 8 did not . This game is not a fair game at all .

  29. I believe that A0 uses Morphy-esque moves because he thinks way more deeply than humans, in fact he is not influenced by our way of thinking at all and thus can understand the game as a whole instead of thinking about "this move leading to this one in 10 turns" or "this piece being worth 5 points"… Stockfish loses all these encounters because he is inevitably mistaken in the value of each piece just like we would be against A0. Indeed, in order to have the exact value of each piece it is needed to find the optimum strategy from a given configuration. For Stockfish this value is hardcoded whereas for A0, this value is approximated throughout a very high number of iterations.
    If you analyse most of A0 vs Stockfish encounters, A0 almost always wins due to a really surprising position advantage. I think that for him what might happen in 10 or 20 turns if one piece is taken becomes obvious. Instead he is way more focused on getting a better position in 20 or 40 turns than we are. When less and less mistakes are made, the slightest shift can win the game.

    And to get a better position you need to take the initiative as soon as possible to get that better position in the future. That is why I think A0 and all of the future machine learning based algorithms will keep playing seemingly losing moves in order to eventually get a better position several turns later. So yeah Romantic chess will revive, but only for AIs.

    Well actually let me take it back, I really hope that there will be more obsessed geniuses like Bobby Fischer who will be able to develop this kind of playstyle against other humans. However, in order to be realistic they would have to work extremely hard on a few openings, thus being able to consistently win games with the white pieces.

  30. If stockfish would stop making desperate pawn runs it would fare much better. I see this time and again. It looses because of one illogical stupid move, maybe the creator placed into its DNA? A bit like satan with his corrupt fallen nature that will destroy himself in the end, or the illuminati.

  31. According to chess.com alphazero calculate 80.000 positions per second, while stockfish calculate 70 millions of positions per second. Wow, amazing, this is one of most glorius moments of AI

  32. It was engineers and mathematics, the human mind that generated the proper algorithms to force this program to teach itself. It doesn't have a Will of it's own, it's following a set of instructions, nothing more. Credit should be given to the geniuses who wrote the incredibly complex programming.

Leave a Reply

Your email address will not be published. Required fields are marked *