Passed pawns - a secret downfall of brute force chess engines : Stockfish 9 vs Leela 395 : Veresov

[@kingscrusher]

Replayable game with nicely indented variations: www.chessworld.net/chessclubs/ltpgnviewer32/ltpgnboard.asp?GameID=5023533&v=acXg3gZ-ufY

[@robertoh.20]

thank you kingscrusher, i enjoy your videos very much.

[@teslathejolteon8007]

Guys! When Leela was at ID 189 I believe, I held a 100 game match between stockfish 9 on my phone and Leela. Of course, stockfish won most of the games, but they drew a lot of games and Leela won some games! Anyone interested in the games Leela won?

[@infinitysalinity7981]

189 is ancient history

[@teslathejolteon8007]

So? She still won, as far as I know, it was unknown that Leela was capable of doing this at ID 189

[@deliveryman]

What phone?

[@teslathejolteon8007]

Andrew Atkins Samsung Galaxy Core Prime

[@deliveryman]

Jesse Smeets lol no wonder. don’t even bother sharing.

[@klimtkiller]

"black's advantage is mammoth". that's a new one. thanks ffor the vid

[@plasmaastronaut]

you're missing that neural nets are happy to mess about when they have winning positions. only when their potential to win is threatened do they play strong moves. A quirk of neural nets is they get very good at knowing when they can mess around and when they have to play the best moves. So once the passed pawns were in a winning formation, a lot of those late game moves from leela were idle.

[@bcfblack]

Another very exciting game! I'm curious about what the hardware data is for this game? Stockfish 9 doesn't typically choose the Veresov on move 2 of double queen pawn, and my own home version of Stockfish detects that Qe3 is a mistake after about 10 seconds. I love this series completely and I believe in your optimism for the future of Leela, but I must admit that I am a little suspicious of moves like Qe3 coming from top engines.

[@InfiniteUniverse88]

Decoy attack. The e5 threat was the decoy.

[@msabagh100]

Well, this game looks a little bit suspicious to me! Leela is not even able to make a draw against Stockfish 8 on my computer. She’s also having hard time playing against engines like Critter, Komodo9,.... How come she wins a game against Stockfish 9 in such a brutal way?! I don’t know. Maybe something is wrong with my hardware.

[@robertleeluben]

Nxb2 was just screaming to be played.

[@leslassiter6378]

Exactly, first move I thought of.

[@electricmaster23]

What were the hardware specs for both engines? :)

[@THEBOYDALEK]

Hmm I can't quite understand why in this game stockfish 9 would make this dubious move unless time controls are short. I have downloaded the pgn and using a simple 2011 Mac-air went straight to the critical position, and yes for a few seconds it analyses Qe3 as good but then after a short time it removes it from the list (6 lines) and then it disappears totally, and evaluates -0.49! If I play it manually immediately it replies with Nxb2. So somewhere along the line I suspect there is something awry! Much as I love some of the fantastic sacs/moves/strategies etc that Leela/Alpha Zero has shown it is capable of, it is almost as if the desire to illustrate the brilliance of AI, is allowing an unfair competition.Great analysis from KC though, so thanks.

[@Christodoulosk]

Amazing game!

[@modolief]

Can you please put the hardware specs in the video description? For example, does the CPU for Stockfish consume the same amount of power as the GPU for Leela Zero? And can you also please put the time control?

[@Norpan506]

Stockfish is basically making useless moves.

[@josh995]

What is the strongest leela id? Also is there anything that can beat stockfish 9 except for alphazero?

[@smashu2]

something is wrong with this game SF9 does not play 2 Nc3 against d5 ever so explain me was it a forced opening ?? like in tcec ??

[@devvanbutler2758]

can john d set leela up as black vs stockfish (white) 1 d4 c4 2 d5 f5?

[@tobiaszb]

10:10 Why black bishop doesn't take the rook imediately on E6?

[@batistalift]

Bxe6+ wins the quality back

[@martinbelmont]

What rating does stockfish 9 have?

[@deliveryman]

Close to 3600 on elite hardware. This was not obviously SF9 as full strength, maybe 1 core? My Stockfish at home on 4 cores thought Qe3 was the best move for about 10 seconds and then switched to something else.

[@drob9673]

Qe3 is a blunder. Or a way engines emulate weakness. Basically, engine to match opponent strength declared in the setting will try to match it by making a mistake. ELO strength is often measured in how many steps behind can opponent be and still win or how many strength points behind, can opponent be and still win. In this case, Nxb2 gives a clear advantage of +1 (to black), despite lost Knight. So, for example, a player with ELO 2000 playing against an engine ELO 3200 should not be able to exploit that kind of advantage. This game is interesting because Stockfish underestimated the overall position as Kingcrusher said. Stockfish could simply "decide" to blunder at this point because it didn't have an opportunity to give +1 advantage to the opponent in previous positions.

[@nabildanial00]

Latest Stockfish realized that Qe3 is a blunder almost immediately. It seems that SF 9 has much inferior evaluation on similar positions.

[@kingscrusher]

Stockfish 9 is the latest Stockfish as far as I am aware - 6th July 2018. I think there is a difference between time management in a practical game with a time limit for all moves, and "infinite analysis" on a particular position where it might get to more easily know a bigger extent of how bad Qe3 is.

[@nabildanial00]

Nope, SF 9 is the latest "official" release but it is not the strongest. Latest dev version are always stronger than the official release. With the latest asmFish (assembler version of SF), it is *at least* 60 Elo stronger than SF 9 and it doesn't take even a sec to know that Qe3 is a blunder. It also finds Nxb2 in less than 1 sec on my machine.

[@drob9673]

Yes. Qe3 looks like engine emulated weakness.

[@kingscrusher]

Nabil that is interesting - is it still essentially using brute force approaches or is it also trying to leverage other AI techniques ?

[@dalsanto4]

What was the hardware?

[@kingscrusher]

See thread: groups.google.com/forum/#!topic/lczero/7Ffntxej6gc for details

[@dalsanto4]

kingscrusher thanks kc loving the leela games :)

[@Norpan506]

great games!

[@azem9338]

leelaZero ist still weaker than Stockfish 9

[@Glider324]

Thanks for sharing this wonderful game. I think we need to drop the whole bubble hype train of "brute force vs non brute force". Leela is brute force. Stockfish plows through millions of trees of variations. Leela plows through millions of networks of memory configurations.

[@kingscrusher]

The thing is I quote " A real intelligence doesn't break when you change the problem " - I would recommend this article: www.technologyreview.com/s/608911/is-ai-riding-a-one-trick-pony/ - it is quite an interesting read I stumbled on myself just earlier today. Also these interviews are great and insightful: kzbin.info/www/bejne/iYvTeoGXa7qpprM kzbin.info/www/bejne/fZDCfaWNnb6io7c kzbin.info/www/bejne/qJPWopSArNWobKM

[@boboswae_11]

Either isn't really true. Stockfish is a highly optimized tree search algorithm that uses human based heuristics to cut its search on a large search space; while Leela chess uses a neural network to create a learned evaluation function and policy, aka intuition. The main difference between the two is that the neural network, because it has intuition, searches through a lot less positions (million of less positions) but still yields very good moves. There is a tradeoff though, the learned evaluation is very computationally expensive, thats why they need graphics cards to be practical.

[@jmm00702]

u dont know what youre talking about, bud

[@sallymalati4364]

◥◥ اا many thanks ❤ ✔ Like ◤◤

[@klirsnichy]

Amazing game