Alpha Zero is HISTORY! - Stockfish NNUE vs Leela C Zero - TCEC 19 Final - ROUND 74 - Slav Defense

TCEC OFFICIAL SITE:
tcec-chess.com/
Join the channel and become a MEMBER!
kzbin.info/door/Qcn...
Support my channel and subscribe.
You can support my channel on www.paypal.me/jozarov.
Follow me on:
/ jozarov
/ jozarov
/ jozarov
lichess.org/@/jozarov
Thanks for watching. Chess is the best!!
Top Chess Engine Championship, formerly known as Thoresen Chess Engines Competition (TCEC or nTCEC), is a computer chess tournament that has been run since 2010. It was organized, directed, and hosted by Martin Thoresen until the end of Season 6; from Season 7 onward it has been organized by Chessdom. It is often regarded as the Unofficial World Computer Chess Championship because of its strong participant line-up and long time-control matches on high-end hardware, giving rise to very high-class chess.
The TCEC competition is divided into seasons, where each season happens over a course of a few months, with matches played round-the-clock and broadcast live over the internet. Each season is divided into several qualifying stages and one "superfinal", where the top two chess engines play 100 games to win the title of "TCEC Grand Champion". In the superfinal, each engine plays 50 openings, once as each side. Beginning in Season 11 in 2018, a division system was introduced; the top 2 engines in each division are promoted, and the bottom 2 are relegated. Currently, there are 5 divisions (a Premier division, and divisions 1-4); newcomers generally start in division 4.
PGN OF THE GAME:
1. d4 d5 2. c4 c6 3. Nf3 Nf6 4. Nc3 dxc4 5. a4 Na6 6. e4 Bg4 7. Bxc4 e6 8. Be3 Nb4 9. Rc1 Bh5 10. h3 Bg6 11. Nd2 h6 12. f4 Be7 13. O-O O-O 14. g4 a6 15. Kh1 b5 16. Be2 Bh7 17. f5 Re8 18. Qe1 Bf8 19. Qf2 Rc8 20. Rg1 exf5 21. gxf5 Kh8 22. Qf3 Rc7 23. axb5 axb5 24. Rg2 g6 25. Bf2 gxf5 26. Qf4 Ng8 27. d5 fxe4 28. Bd4+ f6 29. Rxg8+ Kxg8 30. Bxf6 Qb8 31. dxc6 Nd3 32. Bxd3 exd3 33. Nd5 Rf7 34. Rg1+ Bg7 35. Qxb8 Rxb8 36. Bxg7 Rxg7 37. Ne7+ Kf8 38. Rxg7 Kxg7 39. c7 Re8 40. Kg2 Kf6 41. c8=Q Rxc8 42. Nxc8 Bf5 43. Nd6 Bxh3+ 1-0

[@willyh.r.1216]

I really enjoy your deep strategy and technical comments. The game is great, so you are. Thank you.

[@Jozarovschesschannel]

Thank you for such a beautiful comment

[@ayonmaity6150]

Was waiting for Jozarov Sir to cover these games. Stockfish is The Winner of TCEC SEASON 19 in Brute force+Positional Style. Looking forward to the remaining games of the season and obviously waiting for TCEC SEASON 20.Keep up the good work Sir. God's grace be upon you and your family. Amen ✝️

[@Jozarovschesschannel]

Thank you very nuch

[@ayonmaity6150]

You're most welcome Sir ❤️❤️

[@michaelfortunato1860]

Excellent analysis. Thank you.

[@user-iz7yq1pk4h]

11:35 B f2 the best move of the game 115 watched ❤

[@saarike]

I have nothing more to say, but simply great video and game! Thank you.

[@Jozarovschesschannel]

💪

[@Stockfish1511]

I think showing some games where sf held dubious openings would be nice. Showed some crazy ways to hold bad positions

[@Jozarovschesschannel]

I could really make one video on that. Thanks for the suggestion

[@methylbenzodiazepine]

this is really great chess, thanks for covering it.

[@Jozarovschesschannel]

My pleasure!

[@robinmorritt7493]

Another 20 moves and I beat Leela C Zero/Lichess Stockfish. 🥳🥂

[@edgarramos1499]

I think the GM StockFish is cheating, I evaluated the game and it seems too perfect to be true. I suspect he had phone or a mic and someone was giving him engine moves!

[@Jozarovschesschannel]

🤪

[@depriveable]

Sarcasm on the top

[@milankotevski1663]

Stockfish it the God of chess engines.

[@pierrestober3423]

Stockfish is the king and Leela is the queen would be more like it.

[@ernest6728]

Alphazero eats stockfish all day

[@pierrestober3423]

@@ernest6728 Not at all. That was 3 years ago, there has been a lot of development since.

[@akshaygulabrao372]

I really like your video titles.

[@casahilchoudhary]

Well done, bro. Another Engine game😌

[@Jozarovschesschannel]

I know you dont Like them so much, but at least try

[@casahilchoudhary]

@@Jozarovschesschannel 😅😁. Yeah sure. Why not darling.

[@Zawnpuia1640]

I like engine games a bit

[@utki17]

The author of the video has been a consistent follower of A0 / LC0 and SF development and LC0 and A0 are quite similar after all. Alpha zero was a project by deepmind (now google). They never released A0 to the outside world but just published the results on how it fared against stockfish-8(SF). The more you train such NN engines, the better it becomes but that requires immense amount of GPU as you might imagine, things we can't really afford. So after their initial results Google pretty much kept the project in-house and didn't release the software for any competition like TCEC. However Google did release the theory behind A0 out of which Leela was born. Leela trains herself using volunteer driven CPU and does the same thing as A0 but using 1000s of much smaller machines over the world. All you need to do to become a volunteer is just install their software and your machine will help train the next version of Leela. So after a few years of training leela became quite good and beat stockfish in the TCEC SUFI(Super final) quite easily in S-16 and S-17. This wasn't taken lightly by SF Dev Team as you might imagine :-). They worked hard and provided tons of patches and to everyone's surprise SF bounced back. For ex. last season Leela was literally annihilated by SF. Though some controversy revolves around how Leela scaled but SF-11 was really a beast (for ex. it beat other NN engines quite easily as well). As things stands now LC0 is literally 100s of elo above what A0 was. Also, at the point where A0 released their results (over SF-8), SF11 was already available. So, some questions have been raised about how good A0 NN was compared to latest SF and in general how good NNs would be for chess. Coming to your original question, is A0 dead ? -- NO. Google is actually working with Kramnik to explore different variations of chess with different rules. They are using A0 to checkout how the games look in an attempt to reinvigorate enthusiasm in chess and to make the sport overall more exciting which i think is amazing ! So when the author says A0 is history, he is saying mostly in the lines of the initial wave around NN engines poised to outshine anything and everything in chess :-). Believe me, SF has literally wiped out every NN in this and the last SUFI. In this SUFI especially stockfish has become a ginormous monster and looking at some other games its just crazy how good the engine is. Who knows what the future is but as long as there is SF, there is no point in saying NNs have transformed the chess etc. Having said that, NNs have played 2 key roles in SF development 1.) SF Patches were driven by the defeats from Leela -- good for SF team 2.) SF now uses a very small NN itself (called as NNUE) which has given it strength in closed positions like French / KID. Stockfish-11_nn (introduced in this SUFI ) is using both hand crafted evaluations and its own much smaller NN(trained from its own games) roughly in 50/50 capacity. But when it comes to endgames and tactics, and depth of calculation, the Hand crafted evaluation is just unbeatable (probably no NN can succeed no matter the amount of training -- chess IS that complex). In this exact game, which I followed till the end, you might be surprised to know that NNUE was totally out of ideas. Near the end-game right before the knight moves on 7th rank, SF's NNUE was reporting an eval of -1.28 and SF's Hand crafted engine was giving + 8.0 or something. In fact, you could see visually see the +8 for white and definitely no advantage for black. Leela was also just around +1.0 which means NNS can't understand some critical positions in spite of so much training. But not so much of a problem for SF. These results kind of prove that NNs are not the future after all since chess is pretty tactical as you might imagine.

[@Jozarovschesschannel]

Great comment, really enjoyed. What do you think about the game?

[@Jozarovschesschannel]

Did you Like it?

[@utki17]

@@Jozarovschesschannel yeah.. I loved it... I especially love your coverage since its so spot on & your thoughts are completely in line with almost every chess developer ! Thanks for that ! Meanwhile, I think you MUST cover game 60. Jeroen from TCEC (who sets the openings) called the move 150 as the best ever move played in TCEC history :-) I partially agree, but I think that game shows the real menace that Stockfish is ! Its just inch-perfectly tuned to win in literally any condition ! Would love to see what you think about that !

[@ivanzonder5228]

i checked the score was 53.5 - 46.5 in favour of Stockfish . how was Leela annihilated ?

[@blundergrandmaster3130]

Interesting comment, but it seems in my opinion premature to state that finally neural networks are not the future based on recent Stockfish domination. Neural networks guys will come up with new ideas and hardware will improve. We didn't see the end of it yet.

[@unsunglt7014]

Super Mighty Stockfish NNUE is the King and LCO is the Queen.

[@Jozarovschesschannel]

And Magnus is the prince

[@unsunglt7014]

NOpe its Bobby Fischer.Sorry to say but today's chess is not really pure talent because of these engines.

[@yaasdpalala4492]

Their pricne is Bobby StockFischer

[@deepmindofx8148]

@@unsunglt7014 These engines aren’t the greenbalt computer Fischer beat in the 70s in fact Humans will need extract every once of brain power to even draw against these engines

[@deepmindofx8148]

KZbin: Stockfish NNUE is the most powerful chess entity in history Alpha Zero: Is he though?

[@Abeni28]

Yes, by far

[@josephcrespo7822]

@@Abeni28 alpha zero literally only practiced for 9 hours, give him a day and he'll wreck stockfish NNUE

[@Abeni28]

@@josephcrespo7822 that's not the way it works lol, as your Elo grows, so does the complexity of the game, which means it won't keep growing like it did in the beginning. Even back then when it "beat" Stockfish 8, AZ ran on a supercomputer while SF8 was on way weaker hardware, was stripped of its book, and was forced to make a move after a minute which isn't optimal for Stockfish.

[@HomelanderOneShotsTheVerse]

Yes, stockfish 13 with a good software will aboustely crush alphazero. No question about it.

[@sumanthmuppalla6597]

Stockfish NNUE now fully ready to face Alphazero

[@unsunglt7014]

actually they are no match anymore SFnn will destroys A0 easily.

[@josephcrespo7822]

@@unsunglt7014 no way, give alpha zero a few more hours and it'll absolutely destroy stockfish NNUE, there's no doubt about it, it's not even fair, the hardware on deep mind is miles above what they use on stockfish NNUE XD

[@unsunglt7014]

@@josephcrespo7822 stockfish 10 or 11 is more than enough for A0.this is the reality bro.

[@josephcrespo7822]

@@unsunglt7014 we need a rematch

[@unsunglt7014]

@@josephcrespo7822 agree but against sF 11 only!

[@casahilchoudhary]

Another Hoodie with a Story? 😁

[@ivanzonder5228]

what is the current score ? anyway i'm more interested in hybrid of Stockfish NNUE , i wonder if DroidFish will use it eventually

[@ak-nb2ds]

droidfish already has it in the latest update

[@LuaChels]

Current score excluding draws is 15 - 8 in favor of sf

[@russianbotfarm3036]

Maybe if we jeer at and mock, Alpha Zero, Google will warm it up and train it further, and show us new levels of chess! :)

[@milankotevski1663]

Alpha Zero would be destroyed. Leela is Alpha Zero on steroids, and it's still weaker than Stockfish.

[@jtmv8915]

Standard (non-shogi) chess has shown to be too tactical a game for the kind of NN architecture used by alphazero, which is an extremely deep (I think like 20 consecutive residual layers) architecture meant for computer vision problems (e.g. image recognition). Apparently this works great for go, but it's just too expensive for chess. It needs to search more nodes to find tactics. However, I've never considered tcec as a rigorous judge of engine strength. It only judges which engine is better at exploiting deep human lines. It's basically for entertainment. The only way to know which engine is stronger is to play 1000 games with no forced opening lines. Leela or Alphazero can't be forced to play moves it wouldn't play, and evaluate positions it wouldn't put itself in. Chess.com's CCC is a much more accurate judge of which engine is stronger. Also we will never get a StockfishNN vs AlphaZero match, because Google Deepmind would only do something if it guaranteed good PR. Even if Google did a match they would probably handicap Stockfish some way such as removing its opening book, to ensure it loses without a fair contest.

[@pierrestober3423]

The problem is that when engines aren't given an opening, they tend to play the same lines every time which is quite boring. I remember an occasion where two engines played exactly the same game from start to finish. Anyway, I don't think that deep learning is dead yet. Alpha zero was only three years ago, who can tell what will happen in three more years ?

[@pierrestober3423]

Actually, it's been less than three years because alpha zero came out in December 2017. So yeah things are evolving very quickly.

[@dimkilago2958]

Without forced openings is always draw.There is nothing there to look,nothing new.With the variety of openings you can really see which engine is better,new ideas.

[@milankotevski1663]

You know nothing about chess engines.

[@jtmv8915]

@@milankotevski1663 Are you a developer for one these chess engines? Feel free to point out any issues with what I said. I have a background in neural networks, but have never worked on a chess engine specifically. Though I am quite familiar with the technical details of the AlphaZero algorithm. I did look up later that the network is actually 40 layers deep.

[@rexcavalier]

What if LCZ will also be enhanced with a semi brute force calculation like SF but also use her neural evaluation like SF?

[@Jozarovschesschannel]

That would be sick

[@russianbotfarm3036]

Use next-gen GPUs, and maybe more of them, if LC0 is built to allow it to scale that way.

[@kingkura]

This is an ABC ...um LCZ beat down, there is a KZbinr in love with LCZ...I wonder why that person is not posting any LCZ games🤔🥊👰👙😭

[@Jozarovschesschannel]

Who is that 😄

[@LuaChels]

Lmao

[@redrinesnow9143]

why they dont let them play without pre-arrange openings? i would like to see what they choose to play

[@supernukey419]

1. Most engines suck in openings 2. Games would be nostly deterministic 3. Pretty much all games would be draws

[@milankotevski1663]

You'd be seeing same openings over and over again.

[@twoking10]

What does the NNUE stand for? Is that an acronym for something?

[@cutiecat2754]

Efficiently Updateable Neural network

[@Tail_Yellow]

only 6+? pshhhh leela: hold my beer

[@TrumanBurbonk]

Hi, can you explain why Alpha Zero is dead if the opponent was Leela?

[@utki17]

The author of the video has been a consistent follower of A0 / LC0 and SF development and LC0 and A0 are quite similar after all. Alpha zero was a project by deepmind (now google). They never released A0 to the outside world but just published the results on how it fared against stockfish-8(SF). The more you train such NN engines, the better it becomes but that requires immense amount of GPU as you might imagine, things we can't really afford. So after their initial results Google pretty much kept the project in-house and didn't release the software for any competition like TCEC. However Google did release the theory behind A0 out of which Leela was born. Leela trains herself using volunteer driven CPU and does the same thing as A0 but using 1000s of much smaller machines over the world. All you need to do to become a volunteer is just install their software and your machine will help train the next version of Leela. So after a few *years* of training leela became quite good and beat stockfish in the TCEC SUFI(Super final) quite easily in S-16 and S-17. SF. This wasn't taken lightly by SF Dev Team as you might imagine :-). They worked hard and provided tons of patches in SF and the bounced back. For ex. last season Leela was literally annihilated by SF. Though some controversy revolves around how Leela scaled but SF-11 was really a beast (for ex. it beat other engines quite easily as well). As things stands now LC0 is literally 100s of elo above what A0 was. Also, at the point where A0 released their results (over SF-8), SF11 was already available. So, some questions have been raised about how good A0 was compared to latest SF. So coming to your original question, is A0 dead ? -- NO. Google is actually working with Kramnik to explore different variations of chess with different rules. They are using A0 to checkout how the games look in an attempt to reinvigorate enthusiasm in chess and to make the sport overall more exciting. So when the author says A0 is history, he is saying mostly in the lines of the initial wave around NN driven theory engines poised to outshine everything :-) SF has literally *wiped out* NNs in this and the last SUFI. In this SUFI especially, stockfish is a gynormous monster and looking at some other games, its truly a thing of beauty. Who knows what the future is but generally speaking SF would seem to stay on the top for a very long time so there is no point in blindly saying NNs have transformed the chess etc. But having said that, NNs have played a key role in 2 SF development 1.) SF Patches were driven by the defeats from Leela -- good for SF team 2.) SF now uses a very small NN itself (called as NNUE) which has given strength to SF in closed positions like French / KID (more like made it flawless). Stockfish-11_nn (introduced in this SUFI ) is using both hand crafted evaluations and its own much smaller NN(trained from its own games) roughly in 50/50 capacity. But when it comes to endgames and tactics, and depth of calculation, the Hand crafted evaluation is just unbeatable (probably no NN can beat ever, no matter the training). In this exact game, which i followed till the end, you might be surprised to know that NNUE was totally out of ideas. Near the end-game right before the knight moves on 7th rank, SF's NNUE was reporting an eval of -1.28 and SF's Hand crafted engine was giving + 8.0 or something. In fact, you could see visually see the +8 for white and definitely no advantage for black. Leela was also just around +1.0 which means NNS can't understand some critical positions. But not so much of a problem for SF. These results kind of prove that NNs are not the future after all since chess is pretty tactical as you might imagine.

[@TrumanBurbonk]

@@utki17 Thanks for the detailed explanation ( although I don't understand these acronyms ). So, can we assume that Alpha Zero is stronger than Leela as it uses a much stronger CPU?

[@utki17]

@@TrumanBurbonk I'll clarify :- NN is neural network A0 is Alpha Zero LC0 is Leela chess Zero SF-8 is StockFish 8 Last year it was StockFish 11 and so on. NNUE is Efficiently updatable Neural Network (written backwards -- that's the best humor software engineers have.. lol), which is a different kind of NN (but still an NN). Let me explain how a neural network based chess engine might work in a very simple way so that you can appreciate this more. I'm taking a stab at explaining NNs in a way that needs no math or computer knowledge. Here it goes :- Step - 1) Make a problem computer can understand Let's take chess position. The first question we need to ask is how can we meaningfully represent a chess position numbers. Why do we do this? Because computers can understand only numbers so if we are working with computers this is necessary. Its actually quite simple to do :- Let's assign 1 = where W.king sits, 2 = where W.queen sits, -1= where B.king sits 0 = empty square etc. Now your chessboard can simply become just a list of 64 numbers which can be easily fed to a computer. For ex. the starting position in chess (starting from bottom left corner) :- [W.Rook, W.Knight, W.Bishop, W.Queen, W.King, W.B, W.Kn, W.R, W.Pawn, P , P , P , P , P ,P , P , 0,0,0,0,0..., B.Knight, B.Rook] becomes [5,4,3,2,1,3,4,5,6,6,6,6,6,6,6,6,0,0,0,0..., -4, -5] and so on. To save some space, I'm calling this as [64 inputs]. It doesn't matter what these numbers are as this is just our representation. So we can say some chess position like [0,0,0,0,0,0,0,0,0,0,0,0,0,0,,0,0,0,0,0,5,-4,1,0,0,-6...] is probably some end game ? Since all the A file is empty etc. ? Our next goal is to find the best moves given a [64 input]. Again, we need a numerical representation of "moves" so that we can pick the "best move". Now for a given [64 inputs], let's say you have 143 valid moves. For the sake of simplicity let's say you always have 143 moves in chess given any position. I know this is a weird approximation, but let's focus on the NN aspects and forget about the other trivial things. So, let's say our 143 moves are :- 1. White Queen to F4, 2. Black Bishop to A5, 3. White Knight to H3 .... 55. White castle Queen side ... 142. Black Bishop - C7 143. Pawn H8=Queen and so on ! Just pick something randomly, doesn't matter. So, given any [64 inputs] our software should produce 143 number between 0 & 1 corresponding to all 143 moves which we defined above. closer to 1 means great move and vice versa. For ex. if :- [0,0,0, ... -2,4,-5,5...] produces something like [0.93, 0,44, 0.24, 0.12, ... ] so we know MOVE 1 is best (let's say 0.93 is highest in that list). So we pick move 1 is Queen - F4 or, [something else] produces something like [0.03, 0,94, 0.04, 0.12, ... 0.99, 0.13] so we know MOVE 142 is best(0.99 is highest). and move 142 is Bishop - C7 So we need to do this at every move in a chess game and as long as we get a best move, we have a software capable of playing chess. Now let's see Step 2

[@utki17]

STEP-2) So now you have to produce a machine (software) like this :- [64 numbers] ---> something ---> [143 numbers] Now let's expand the *something* based on what we do know. We do know that we must have some relation between the 64 input numbers because they depend on one another. What does that mean ? Well, For ex. let's consider this position:- [..., 5, -4, -1, 5, ...] One can say that because 5 and -4 are next to one another *something* needs to consider that. Also it needs to see where the king is, is it castled and what not. Essentially all numbers of [64 inputs] should be related to one another somehow. This is something we need to capture in simple math next. [64 numbers]---> {some math across 64 inputs} --> [some intermediate results]---> {some math across intermediate results & 64 inputs} ---> [143 outputs] Let's assume that [some intermediate results] has 256 numbers in it and each of the 256 numbers are *computed* from 64 inputs in variety of different ways ! How can we do that? This can be some very simple equations like this one :- first number of [some intermediate results] = (input-1) * 0.56 + (input-3) * 0.45 - (input-7) * 0.88 second number of [some intermediate results] = (input-4) * 0.19 - (input-1) * 0.80 - (input-19) * 0.34 + (input-34 * 0.11) ... . . . . . . 256th number of [some intermediate results] = input-64 * 0.23 - input-13 * 0.66 + input-19 * 0.08 There is absolutely no pattern here. These numbers 0.56, 0.45 are all completely random. And the relation is also random. For ex. second number of [some intermediate results] depends on more than 3 values in the chessboard unlike other entities, But this model is just fine ! Even the number 256 is random. Some models are better than others etc. and yes, keep in mind computational complexity ! But this model works too. At some point we are going to find a good values for 0.56, 0.45 etc, and hopefully that's our chess engine Let's shorten this :- [64 numbers]---> {some math across 64 inputs} --> [some intermediate results]---> {some math across intermediate results & 64 inputs} ---> [143 outputs] to [64 numbers]---> layer1 ---> [143 outputs] And maybe with 256 intermediate results, now you have introduced 3000 RANDOM decimal numbers in Layer 1(like 0.56, 0.45 etc). For us, only the RANDOM decimals are the ONLY unknowns which we need to find at some point. You give some [64 numbers] -- a chess position to this thing, it'll surely produce some random 143 numbers in the output. And yes, one of the 143 numbers will be the highest too. Even though the output is random, we have a model that can simulate chess in computer. Maybe 1 LAYER is not enough, so we can even do this :- [64 numbers]---> 500 layers ----> [143 outputs] Each layer can be different with completely different math around it and In fact we can even have infinite layers in some models (don't ask me how). Now with 500 layers you are allowing TRILLIONS and TRILLIONS of math operations with let's say MILLION decimal unknowns in all the 500 layers. Now we need "real" GPU computing powers. Overall the chess problem has become a problem of finding these MILLION numbers..and yes.. its a problem modern computers can attempt to solve ! So let's see how they do it.

[@utki17]

Step 3.) Start playing. All we are trying to find is the best combination of MILLION numbers (with our fixed equations and relations) so that this entity can start producing ouputs which correspond to a good move in chess ! So we let the computer play a game with itself ! -- this is what Alpha Zero did in its very first game ! GAME 1 :- We give an input board (starting position - [5,4,3,2,1,3,4,5,6,6,6,6,6,6,6,6,0,0,0,0...]) to 2 engines with 2 sets of random MILLION numbers. Clearly Both Black and white will Predict "some" best move and play along but we all know that these moves will be completely garbage to say the least. 1.) White starts with F4, -- ugly move 2.) Black play H4 next -- worse ! .. .. ... After 23 moves. There is a mate in 1 and black finds it. Even though mate was visible to humans in move 10 but the engines go their own way. We let these engines play one another 10000 times, and let's say BLACK wins 7000 - 3000. So we proclaim that the engine playing with BLACK chess pieces is the CURRENT-WINNER ! Now somehow we want to memorize something from this game and surely the answer lies in those MILLION numbers of the engine playing with BLACK chess pieces. This is same as when you first played chess, you got checkmated on F7 (who hasn't) by the bishop, you remember something from that game. How do we do that with computers ? On the very first go, there is just one winner (and no former winner) so the CURRENT-WINNER plays against itself. Let's say its CURRENT-WINNER vs CHALLENGER. In the CHALLENGER, we *twist some of the knobs* of these MILLION numbers so that the CHALLENGER plays a slightly different chess. Which direction we twist the knobs, depends on the previous results but we ensure that CHALLENGER is slightly different from CURRENT-WINNER so that we continue to explore different lines but not completely randomly. Then, another 100000 games. Let's say this time, CURRENT-WINNER wins then we trash the CHALLENGER and attempt some other combination of MILLION numbers (derived from the CURRENT-WINNER) until the CHALLENGER wins. Then we average out the MILLION numbers, from the CURRENT-WINNER and the CHALLENGER and promote the CHALLENGER to CURRENT-WINNER. This goes on forever. In fact Leela is doing this as we speak. This method ensures that the CURRENT-WINNER is always better than the PREVIOUS-WINNER which means, the engine is "learning". Slowly you begin to see this :- first 1000 games --- horrible games ! next 10000 games -- some beginner level chess games 40 million'th game ---- Alpha Zero that can beat Stockfish and literally any human. With such mechanism the engine has eventually tuned the MILLION numbers in such a way that the complex equations it uses are able to solve many chess positions. It has learnt obvious tactics because it lost so many times to it. It learned many opening ideas on its own which were known to us from ages. Also, in some positions it plays completely new moves (which humans' don't know) because event hough it explored them quite randomly in its training, since the CHALLENGER kept winning those ideas are now rewarded. This is not exactly how A0 works (its much more complex), but roughly this is the idea behind neural networks based chess.

[@luwangth999]

Can this StockFish challenge Alpha?! Alpha seems too strong..not even to join this competition..

[@dimkilago2958]

Stockfish 12 is 200+ elo points stronger.Stockfish 11 was also stronger.

[@unsunglt7014]

Alpha zero is the strongest they say but only against stockfish 8 classic.

[@Jozarovschesschannel]

They should let Alpha run again to face this Nnue

[@blijebij]

@@Jozarovschesschannel I tried a game of alpha vs stockfish 8 and put sf 12nnue on alpha's place...to see if he would do a lot of diff moves then alpha. Result..95% of the moves vs sf 8 was exactly the same as alpha. The other moves he chose slightly better moves if sf12nnue is right here. So I think they would be more or less equallly matched.

[@dimkilago2958]

@@blijebij This is not so serious criterion.Maybe are just not so accurate moves from sf 8 and not the other way(deep moves from stockfish 12 and Alpha).And it has to do also with the opening.

[@DanielDurham121]

Enjoyed the content, but I think you should try to change your style up from just doing exactly what Agadmator does.

[@lautarohoyos6074]

Finish her!

[@Jozarovschesschannel]

FatAlity

[@dimkilago2958]

I hear that Leela is better positionally but Stockfish won some crazy games in close structures and Leela didn't.

[@Stockfish1511]

Not anymore. New stockfish is better than leela in absolutely every way. Leela might be better in couple of openings. But right now stockfish is way to strong

[@ary_disini]

Stockfish is just like lynux, everyone can upgrade it

[@l00d3r]

Stockfish 12 is a hybrid engine that also uses a neural net, which allows it to play close positions better

[@theobserver314]

*Deep Blue entered the chat.

[@Mohammed_Ali_4941]

this channel is better than agadmator.

[@mlungisimanzini8143]

Lol

[@maadbashir9447]

First veiw

[@Jozarovschesschannel]

Finallly