End of Chess AB Engine Era ? | Leela beautifully punishes Stockfish's outrageous materialistic ways

♚COURSES kingscrusher.t...
📚 kingscrusher.t...
📚 kingscrusher.t... || kingscrusher.t...
#KCComputerChess
#KCLeelaChess
♚ Play turn style chess at bit.ly/chessworld
♚ Play Chess vs. Kingscrusher and others: www.chessworld...
♚ Subscribe to best KZbin Chess Video Channel : bit.ly/kingscru...
🎥 Leela Chess Playlist: bit.ly/leelachess
FIDE CM Kingscrusher goes over amazing games of Chess every day, with a focus recently on games of Neural Networks which are opening up new concepts for how chess could be played more effectively.
👕 Kingscrusher Store: teespring.com/...
♚Donating via Paypal: goo.gl/7HJcDq
♚Channel exclusive perks: / @kingscrusher
Game quality tags: amazing, awesome, astonishing, brilliant, classic, crushing, dynamic, elegant, exceptional, excellent, exciting, fabulous, famous, fantastic, finest, flashy, greatest, important, impressive, incredible, instructive, interesting, magnificent, marvellous.
Info about Leela Zero:
en.wikipedia.o...
...
Leela Chess Zero (lc0) is a free, open-source, and neural network-based chess engine and distributed computing project.
Leela Zero's algorithm is based on DeepMind's 2017 paper about AlphaGo Zero.[3][6] Unlike the original Leela, which has a lot of human knowledge and heuristics programmed into it, Leela Zero only knows the basic rules and nothing more.[7]
Leela Zero is trained by a distributed effort, which is coordinated at the Leela Zero website. Members of the community provide computing resources by running the client, which generates self-play games and submits them to the server. The self-play games are used to train newer networks. Generally, over 500 clients have connected to the server to contribute resources.[7] The community has provided high quality code contributions as well.[7]
Leela Zero finished third at the BerryGenomics Cup World AI Go Tournament in Fuzhou, Fujian, China on 28 April 2018.[8]
Info about Alphazero:
en.wikipedia.o...
AlphaZero is a computer program developed by the Alphabet-owned AI research company DeepMind, which uses an approach similar to AlphaGo Zero's to master not just Go, but also chess and shogi. On December 5, 2017 the DeepMind team released a preprint introducing AlphaZero, which, within 24 hours, achieved a superhuman level of play in these three games by defeating world-champion programs, Stockfish, elmo, and the 3-day version of AlphaGo Zero, in each case making use of custom tensor processing units (TPUs) that the Google programs were optimized to make use of.[1] AlphaZero was trained solely via "self-play" using 5,000 first-generation TPUs to generate the games and 64 second-generation TPUs to train the neural networks, all in parallel, with no access to opening books or endgame tables. After just four hours of training, DeepMind estimated AlphaZero was playing at a higher Elo rating than Stockfish; after 9 hours of training, the algorithm decisively defeated Stockfish 8 in a time-controlled 100-game tournament (28 wins, 0 losses, and 72 draws).[1][2][3] The trained algorithm played on a single machine with four TPUs.
...
Relation to AlphaGo Zero
Further information: AlphaGo Zero
AlphaZero (AZ) is a more generalized variant of the AlphaGo Zero (AGZ) algorithm, and is able to play shogi and chess as well as Go. Differences between AZ and AGZ include:[1]
AZ has hard-coded rules for setting search hyperparameters.
The neural network is now updated continually.
Go (unlike Chess) is symmetric under certain reflections and rotations; AlphaGo Zero was programmed to take advantage of these symmetries. AlphaZero is not.
Chess can end in a draw unlike Go; therefore AlphaZero can take into account the possibility of a drawn game.
AlphaZero vs. Stockfish and elmo
...
►If you like my chess videos please consider become a full member here at my Chessworld Site. www.chessworld....
#KCChess game tactics
♞ Challenge KC and others for turn style chess at bit.ly/chessworld
♚TEAM kingscrusher.tv...
♚TWITTER kingscrusher.tv...
♚DISCORD kingscrusher.t...
♚TWITCH / kingscrusher
♚REDDIT kingscrusher.tv...
♚QUORA kingscrusher.tv...
♚DONATE paypal.me/king...
♚COURSES kingscrusher.t...
♚COURSES kingscrusher.t...

[@kingscrusher]

Q. Are we approaching the end of Chess AB Engine Era ? 🤔😀😎🌍 Video: goo.gl/xy8rvE Replayable game: goo.gl/TDX4Ag #chess #chessgame #ArtificialIntelligence #AI #MachineLearning #DeepLearning #Stockfish

[@pnachtwey]

Are you sure that LC0 does not use an alpha-beta search? I thought the AI was only for the evaluation routines at the end nodes of the search tree. Searching deep doesn't help except in tactical situations. Search smart is good as long as the extra smarts don't affect the depth of search. It is a trade off. BTW, your videos inspired me to replace my 9 year old computer with a very fast one with an I9 cpu, 2.66 GHz memory and a GTX 1080 with lots of CUDA cores so I test the best chess programs.

[@julioandresgomez3201]

An interesting question, when a well trained neural network dismisses certain opening, or favors an opening, does that mean that the opening is objectively stronger (or feeble), or that it´s picking what better suits its strengths and talents?

[@Jaydelrey]

@@pnachtwey Lc0 uses MCTS.

[@julioandresgomez3201]

Still have Stockfish 11 as last stand. Nonetheless, it´s predictable that any progress in refining evaluation and in-depth calculation will be severely outmatched by reinforced learning systems. Tactical imperfection is the only hope but in practice there´s no difference between never blundering and blundering once in a million.

[@quag443]

Actually SF has been the reigning champ for only one year. It didn't win TCEC 10. The top spot is always being contested which I think is a great thing as it keeps things interesting and encourages rival engines to keep improving. At the end of the day, everyone wins. One thing we can do with less of though is the bashing. I hope hardcore Leela fans will one day respect the hard work that goes into Stockfish, and vice versa. Both engines will co-exist and benefit each other for quite some time so it would help if the fans got along too. Anyway, another great game by Leela. Let's hope she keeps it up!

[@jtmv8915]

Leela is being developed by one of the core Stockfish developers. Other developers of top AB engines will also try to create their own NN engines. To partake in fandom bashing like these are rival sports clubs is absurd. (But it'll probably happen anyway)

[@u.v.s.5583]

End of the end of the end of the beginning of the end of the beginning of an era.

[@viktorpti3946]

It's wrong contraposition: AB-engine vs. NN-engine. Since these are different orthogonal aspects: search algorithm and evaluation function. Orthogonal means that all combinations are possible: AB can be used together with NN, while MCTS can be used with hand-crafted eval.

[@jtmv8915]

Quite true, although the NN evaluation function takes much more time to calculate than current state of the art handwritten ones. Hence why it's - at least for now - tethered to the current Monte Carlo methods.

[@Bobby-fj8mk]

Wow - White didn't castle & black had 2 passed connected pawns and still lost. This is breaking all conventions. Great game.

[@kingscrusher]

Cool comment, K

[@Brandon-a-writer]

when Leela becomes sentient and enslaves the world, she'll remember you and spare you the indignities to befall the rest of humanity. :D *tear* our girl's all grown up. so proud

[@ternerito]

Amazing game by leela. I thought that the previous game she won was also quite interesting. Any chances KC will find the time? :)

[@kencusick6311]

Very impressive win by Leela. The advantage gained by the pawn sac isn't even remotely on SF's search ply horizon. A classical position sacrifice. Tal is applauding from somewhere.

[@kingscrusher]

Cheers, K

[@douglasquaid7550]

L E E L A , we , the fans of her, told the fishbois that she would champ of S14 6 months ago, and well look at that now she is leading the superfinal, GOOOO LEEEEELAAAA !!!

[@kingscrusher]

Fun comment - Cheers, K

[@Askhat08]

That's true! ;)

[@carrottoponcrak]

the fishbois lol

[@dilyarm9477]

TCEC needs your nice evaluation, thank you so much

[@kingscrusher]

Cheers, K

[@Hellstaff01]

I also expect to see a revision in opening theory. But what would be the outcome, if leela plays all mentioned openings against herself, finding better defensive resources or avoiding deep traps? Some of these openings could be rated as objectively inferior and may disappear in high level chess.

[@kingscrusher]

Cheers, K

[@AndrewBackhouse1]

I agree this is groundbreaking stuff.

[@CarlosRodriguez-ru8em]

Silly question most probably already know the answer. What will happen when Alpha Zero plays Leela? Alpha Zero smoked Stockfish yet Leela is winning slightly vs the fish. Also have two AI Artificial Intelligence projects played before? My guess is Alpha Zero will be stronger than Leela but only slightly maybe in 100 matches come down to lets say 59 to 41, but eventually Leela will match and perhaps exceed Alpha Zero maybe a few years down the line.

[@klieu90210]

A0 played SF8 with suspicious conditions. This is SF10. Also no signs of A0 being brought back any time soon, since it's not among Google's priorities.

[@Norpan506]

Is d4, c4 or Nf3 Leela's favourite opening?

[@Durian_Defense]

When was the last time that Stockfish played a match where the last five decisive games went the opponent's (in this case, Leela) way? The last five decisive games (including this one) were won by Leela. Simply crushing!

[@tomyman]

0:22 you meant last few seasons not years, in november 2017 season 10, houdini won the superfinal against komodo

[@Th3Wick3dOn3YT]

Yes KC, its not something to question. Yes the rook wants to participate in this game. :)

[@ig2d]

I prefer fianchetto to fianchetto

[@kingscrusher]

Fun comment - Cheers, K

[@BramCohen]

For these TCEC games you should include more comments about how the computers thought they were doing in the middle of the game, which is both shown in real time during the games and included in the history replays

[@dubsphinx]

That's the one thing I'm missing here too, show us some evals KC!

[@zhu05107]

Great game, Mr Crusher. I started playing chess a couple of years ago and just find this move towards self learning programs inspiring and just love your enthusiasm for it too.

[@kingscrusher]

Cheers, K

[@alexmagor7538]

I watched this game live. What struck me was Leela’s evaluation was over +2 from her first out of book move. And even more interesting was in the following game when she had to face the same opening from Black was how she managed to equalize and finish with a draw.

[@kingscrusher]

Cheers, K

[@glaucosaraiva363]

I could understand why Leela droped the second pawn, but the reason that made she drop the first one is unclear for me.

[@paulbloemen7256]

Great, great, great! And I mean: GO Leela, GO, GO, GO!!!

[@mustafaunal1834]

End of era? Not yet...

[@nicholasperkins4655]

Good to hear that Leela is holding the other side with the Slav. If Leela lost with the Slav and beat it with white then that would make the Slav a suspect opening and that would destroy chess theory as we know it. Slav defense is the most popular against d4.

[@davids2448]

My thoughts too. Opening theory revision only needed where both AI & AB style engines struggle with a given line.

[@kingscrusher]

Cheers, K

[@Pintkonan]

now, since kc has learned about correct pronounciation of fianchetto, my life becomes a bit more perfect =b

[@davids2448]

Now we just need KC to tell us that the C in Kingscrusher is silent...;)

[@yotoober1]

I would not discount SF just yet...with these losses, will not SF have to update its hash/tablebase to determine which moves resulted in a negative evaluation which resulted in a losing position? At some point, SF will equalize, so SF also learns from its losses too. Thats why I say that eventually, the ELOs will equalize within a sample error of games, say up to 50 ELO points +/- each. I would be very surprised if at some point wiht a new SF version, if SF does not equalize to a series of draws again. With these losses by SF, what has happened is that Leela has exposed a flaw in SF alogrithm which can be exploited indefinitely given the same position, so the coders/programmers of SF will have to adjust the SF algorithm to correct and equalize this computational error.

[@grolich]

yotoober1 SF doesn’t learn between games at all. Everything in its search/veal update is being done manually by the devs and not on a per position basis (that has been found to not be very effective years ago). Hash table update doesn’t affect next game, evaluation function doesn’t change as well. Engines that were using persistent hashes as an alternative to “better” opening books were found to be ineffective as without generalizations, getting just 1 position (or a few) doesn’t help at all in a measurable way, so it’s not really done usually anyway. Stockfish May equalize because as we have seen, leela can blunder, but unfortunately, the way you were thinking about sf learning from its mistakes just doesn’t really apply.

[@yotoober1]

@@grolich So how will the developers correct the computational error that SF has made in its losses if its not using a hashtable? Otherwise, there would actually be no way to correct the error at all, as a human will not be able to find the computational eval error. Obviously, the exact position may never occur, agreed, but I meant the approximate position where an incorrect evaluation was made, or perhaps it is better to say where SF negative evaluation began to increase in an accelerating manner, ie, similar to a breakthrough position. I dont know how large a hash they are able to accomodate, but nonetheless, use of a hashtable is the only way to expose the error by SF, thus without use of it, I dont know how the devs will adjust for the errors in SF evaluation and computation. Taken to an extreme, to point out the validity of the hashtable, with an unlimited hash size, I can guarentee a draw every game no matter who the opponent is, as each move has been logged for its reply for positional equality. I dont know how or who made the decision that hashtables were not effective, obviously, I did not (lol). They are 'not effective' given the limited size of memory to accommodate larger and larger hashtables. I would expect as with everything else, just to prove the point, that a still larger and larger computer memory will be made available to finally prove the point of NN vs. brute force in chess, ie, that with virtually unlimited memory, that brute force calculation will draw or win everytime. It must, as all NN are trying to accomplish, is to deduce etherically (ie, 'intuit') what brute force does mathematically. Ultimately, all Leela will be proving by beating SF in the Finals, is that in chess, a NN or whatever they wish to call it, is superior to brute force calculation with a 'limited' hash table size. It will prove nothing else...and weve already known that for quite some time now, havent we?

[@yotoober1]

@L Max You have said something which is similar to what I have been saying in a different way. But the key word is "appoximate", as it can never duplicate exactly what a tablebase can do, at least in chess, as chess does have a finite, although very large, amount of moves that can be made. So, NNs would be more useful where there would be an unlimited amount of possibilties, but even there, it can only learn what it is 'coded' to look for and nothing else. Google is trying to do the same with self driving cars, but safe self driving cars in my opinion are a decade or so away at least, as again, an NN cannot cover all possibities taht can occur while driving, eg what would it do if while driving a metor appears to be falling out of the sky; so does the car keep driving, slow down, stand on the brakes, etc...it would not know what to do as it is not coded to look for such a situation and this situation rarely if ever occurs.

[@nicholasperkins4655]

Awesome! I get to see how Leela destroys the Slav!

[@kingscrusher]

Cheers, K

[@pazdziochowaty]

Leela is going to become the champ. However I would not say this is the end of AB Engine era. For that hardware has to become cheaper. Stockfish will remain "folks king" because it can run on regular computer and have great strength. Neural network require high-end GPU which is expensive as hell. On my notebook I am able to get quality analysis from SF (depth 50+) but to get anything above depth 12 from Leela I would need to wait for a week because it runs very slow on regular GPU. Lower depth means you do not get leela's full strength. So Leela can be used by chess professionals who are willing to spend $10k or more on a strong GPU. I am not willing because that GPU would not serve me for anything else but running leela (I am not a CAD designer or cartoon maker). I would rather buy a stronger CPU for that sum because CPU can be utilised by different kind of software, not only chess engines

[@Jaydelrey]

$10k seems to be a stretch.. a 2080ti is "only" $1300 ish and TCEC uses a 2080ti + 2080 for leela. (2080 is around $800 USD)

[@pazdziochowaty]

@@Jaydelrey Even $800 for piece of hardware only for chess engine is way too much for someone who treats chess as a hobby. On my laptop, SF gets 60MNPS while leela struggles at 300NPS (not mega, not kilo, 300 nodes:)

[@mrrobot790]

What's AB stand for?

[@An-ht8so]

Alpha Beta

[@gJonii]

Alpha beta pruning. It's a method to make brute force chess engines much faster, by eliminating large number of tree branches you need to check(that's where the pruning part comes from). The alpha beta comes from the way this pruning is done. You assign one node a label alpha, and another beta, and given some conditions, you don't need to check what happens with the tree branch starting from beta. This allows you to skip like 99.999% or more of all the tree in large enough trees, thus making it thousands or millions times faster than just your average brute force search.

[@mrrobot790]

Thanks you two!

[@mamago92]

The future has come too early?

[@Stockfish1511]

personally i think the hardware sf uses is weak compared to what leela uses. Would be fun to see the same net playing sf on ccc instead.

[@youuuuuuuuuuutube]

The difference is more about the algorithm than the hardware. Some people want SF to play with BF hardware (4 times as many cores), but if you think about it, to get 1 full extra depth in nodes calculation, you need around 30x the hardware power, since there are on average 30 possible moves, so a 4x won't give you that much extra.

[@Stockfish1511]

@@youuuuuuuuuuutube im not very into details. But seems like sf on 90 threads at ccc is much more sharper than the one in tcec

[@tomyman]

@@youuuuuuuuuuutube it doesn't work that way because the engines prune the moves, with x2 hardware you get about +40 elo (it has been tested for years)

[@Askhat08]

This hardware setup is perfectly balanced.

[@Askhat08]

And SF is having a hard times against army of Leelas in ccc right now ;)