Mate in 3: 7 minutes for SF 17

[chrisw]

Eh? Football didn't break when the EFL (English Football League) changed a win from 2 points to 3 points in 1981, so why would chess break?

Chess is not football. Chess is a game with complete information to both players, subject only to logic. Tons of chess combinations that involve a draw of any sort would be massively devalued if two draws didn't equal one win.

I don't understand why people who don't like chess anymore can't just quit chess and start a game that they like better.

I don't understand why engineers that can't solve algorithmic chess problems resort to denial. Isn't failing to attempt to work out fortresses/null moves just a form of quitting?

[towforce]

I don't understand why people who don't like chess anymore can't just quit chess and start a game that they like better.

Apologies - need a quick trip off-topic here.

Have you missed what's happened over the last few decades? The "computer chess experience" we've been treated to has been PHENOMENAL!

* It was amazing that small computers could play chess - and they were fun to watch (like monkeys or children)

* The book "Chess Skill In Man And Machine" showed us exactly what was needed to become a GM (expert knowledge of 50,000 chess patterns), and gave a great insight into the game of chess. I read that book in 1984.

* The great programmers of the 8-bit CPU era managed to make them impressively strong. Good chess players, while not seeing them as serious opponents, started using them in a small way for occasional analysis or sparring

* Chessbase, a chess database, was created. I quickly understood what a great concept this was - for all areas of life. In terms of AI, it spurred my interest in case-based reasoning - a great way of making small but very useful AI systems. At the time, I thought this was the key to a super strong chess computer (have a database of chess positions - each with its own evaluation function (EF), find the closest match to the current position, and use that position's EF to assess the candidate moves).

* In 1996, the world watched Kasparov v Deep Blue. Kasparov won easily, and I remember very clearly what multiple people said of chess computers beating the world champion in a match: "Not in my lifetime."

* In 1997, Deeper Blue won the rematch. People argued about some of the games (especially game 2) for years to come

* Autotuning of chess engines became common, increasing their strength

* Computers just kept on getting faster. At one point, it turned out that computer and human working together ("3-hirn") was better than either alone. It became folklore that combining the strengths of human and machine was the way forward (this is where I see the chatbot world right now)

* That period was short lived: soon afterwards, computers working alone were better than computers and humans working together

* We got to a point where cheap computers provided enough power to be able to beat the strongest humans at tournament time controls

* Free programs became more than good enough for most people's chess needs (obviously this was bad news for the professionals in this game)

* Surprise surprise! NNs provided great new EFs! They were a lot slower than hand-coded evals (HCE), but it was worth the sacrifice in speed because they boosted engines by hundreds of ELO points. Turns out that the HCE developers had missed some important things about chess

* The first of the "final moments" in chess. Thanks to these strong computers, top level correspondence chess has suffered draw death. This is a good indicator that top chess engines on powerful hardware are rapidly becoming unbeatable: we are witnessing the end days of computer chess right now!

It has been an intellectual ride like few others. Those of us who have watched carefully, have learned a MASSIVE amount about the true nature of chess, and hence about the nature of human intelligence and machine intelligence. Armed with this experience, the chatbot revolution is feeling a lot like the second movie in the series to me!

[chessica]

[d]8/8/5k2/p6p/P1r2n1K/8/8/7R b - - 0 1

Igel 3.5.5 64 POPCNT:

15/27 00:01,688 1.192k 698k +0,20 Sf4-g2+ Kh4xh5 Sg2-f4+ Kh5-h6 Tc4-c8 Kh6-h7 Kf6-e6 Th1-h2 Ke6-f5 Th2-f2 Kf5-g4 Tf2xf4+ Kg4xf4 Kh7-g7 Kf4-g3 Kg7-f7 Tc8-c7+ Kf7-f8 Tc7-a7 Kf8-e8 Kg3-g2 Ke8-d8
16/33 00:06,632 4.635k 706k +0,20 Sf4-g2+ Kh4xh5 Sg2-f4+ Kh5-h6 Tc4-c8 Kh6-h7 Tc8-c7+ Kh7-h6 Sf4-e6 Th1-h3 Kf6-f7 Th3-h2 Tc7-c1 Th2-f2+ Kf7-e7 Tf2-h2 Tc1-a1 Kh6-h5 Ta1xa4 Kh5-h6 Ta4-a3 Kh6-h5 a5-a4 Kh5-g4
17/33 00:12,703 8.900k 699k +0,20 Sf4-g2+ Kh4xh5 Sg2-f4+ Kh5-h6 Tc4-c8 Kh6-h7 Tc8-c7+ Kh7-h6 Sf4-e6 Th1-h3 Kf6-f7 Th3-h2 Tc7-c1 Th2-f2+ Kf7-e7 Tf2-h2 Tc1-a1 Kh6-h5 Ta1xa4 Kh5-h6 Ta4-a3 Kh6-h5 a5-a4 Kh5-g4
18/33 00:12,976 9.093k 701k +0,20 Sf4-g2+ Kh4xh5 Sg2-f4+ Kh5-h6 Tc4-c8 Kh6-h7 Tc8-c7+ Kh7-h6 Sf4-e6 Th1-h3 Kf6-f7 Th3-h2 Tc7-c1 Th2-f2+ Kf7-e7 Tf2-h2 Tc1-a1 Kh6-h5 Ta1xa4 Kh5-h6 Ta4-a3 Kh6-h5 a5-a4 Kh5-g4
19/33 00:14,985 10.475k 701k +0,20 Sf4-g2+ Kh4xh5 Sg2-f4+ Kh5-h6 Tc4-c8 Kh6-h7 Tc8-c7+ Kh7-h6 Sf4-e6 Th1-h3 Kf6-f7 Th3-h2 Tc7-c1 Th2-f2+ Kf7-e7 Tf2-h2 Tc1-b1 Kh6-h5 Tb1-b4 Th2-h4 Tb4xh4+ Kh5xh4 Se6-c5 Kh4-g3 Sc5xa4 Kg3-h2
20/35 00:38,289 26.582k 699k +0,20 Tc4xa4 Kh4-g3 Kf6-f5 Th1-f1 Ta4-a3+ Kg3-h2 Ta3-a2+ Kh2-g1 a5-a4 Tf1-f2 Ta2xf2 Kg1xf2 Kf5-e4 Kf2-e1 a4-a3 Ke1-d2 a3-a2 Kd2-c2 a2-a1D Kc2-d2 Ke4-d4 Kd2-c2
21/35 00:38,583 26.785k 694k +0,20 Tc4xa4 Kh4-g3 Kf6-f5 Th1-f1 Ta4-a3+ Kg3-h2 Ta3-a2+ Kh2-g1 a5-a4 Tf1-f2 Ta2xf2 Kg1xf2 Kf5-e4 Kf2-e1 a4-a3 Ke1-d2 a3-a2 Kd2-c2 a2-a1D Kc2-d2 Ke4-d4 Kd2-c2 Da1-b2+ Kc2xb2
22/35 00:39,337 27.315k 694k +0,20 Tc4xa4 Kh4-g3 Kf6-f5 Th1-f1 Ta4-a3+ Kg3-h2 Ta3-a2+ Kh2-g1 a5-a4 Tf1-f2 Ta2xf2 Kg1xf2 Kf5-e4 Kf2-e1 a4-a3 Ke1-d2 a3-a2 Kd2-c2 a2-a1D Kc2-d2 Ke4-d4 Kd2-c2 Da1-a6 Kc2-d2
23/35 00:42,967 29.856k 694k +0,20 Tc4xa4 Kh4-g3 Kf6-f5 Th1-f1 Ta4-a3+ Kg3-h2 Ta3-a2+ Kh2-g1 a5-a4 Tf1-f2 Ta2xf2 Kg1xf2 Kf5-e4 Kf2-e1 a4-a3 Ke1-d2 a3-a2 Kd2-c2 a2-a1S+ Kc2-d2 h5-h4 Kd2-e1 h4-h3 Ke1-d1 h3-h2 Kd1-d2
24/35 00:47,082 32.712k 695k +0,20 Tc4xa4 Kh4-g3 Kf6-f5 Th1-f1 Ta4-a3+ Kg3-h2 Ta3-a2+ Kh2-g1 a5-a4 Tf1-f2 Ta2xf2 Kg1xf2 Kf5-e4 Kf2-e1 a4-a3 Ke1-d2 a3-a2 Kd2-c2 a2-a1S+ Kc2-d2 h5-h4 Kd2-e1 h4-h3 Ke1-d1 h3-h2 Kd1-d2 h2-h1D Kd2-c3
25/35 00:49,240 34.246k 695k +0,20 Tc4xa4 Kh4-g3 Kf6-f5 Th1-f1 Ta4-a3+ Kg3-h2 Ta3-a2+ Kh2-g1 a5-a4 Kg1-h1 a4-a3 Kh1-g1 Kf5-g4 Tf1-e1 Sf4-e2+ Kg1-f2 Kg4-h4 Te1xe2 Ta2xe2+ Kf2xe2 Kh4-h3 Ke2-d1 a3-a2 Kd1-d2 a2-a1D Kd2-c2 Kh3-g2 Kc2-d2 Da1-b2+ Kd2-e1
26/38 01:14,568 51.874k 695k +0,20 Tc4xa4 Kh4-g3 Kf6-g6 Th1xh5 Kg6xh5 Kg3-f3 Ta4-a3+ Kf3xf4 a5-a4 Kf4-e5 Ta3-d3 Ke5-e6 Kh5-g6 Ke6-e7 Td3-d2 Ke7-e6 a4-a3 Ke6-e5 a3-a2 Ke5-f4 Td2-d7 Kf4-e4 Td7-a7 Ke4-e3 a2-a1T Ke3-d2 Ta7-a8
27/41 01:54,226 79.157k 696k +0,20 Tc4xa4 Kh4-g3 Kf6-f5 Kg3-f3 Ta4-a3+ Kf3-f2 a5-a4 Kf2-e1 h5-h4 Ke1-d2 h4-h3 Th1-b1 h3-h2 Tb1-b5+ Kf5-g4 Tb5-b1 Kg4-f3 Kd2-c2 Ta3-a2+ Tb1-b2 Ta2xb2+ Kc2xb2 h2-h1D Kb2-a3 Dh1-a1+ Ka3-b4 a4-a3 Kb4-c4 a3-a2 Kc4-c5 Da1-b2 Kc5-d6 a2-a1D Kd6-e7
28/44 03:23,304 141.571k 693k +M3 Tc4-c3 Th1-h2 Kf6-g6 Th2-a2 Tc3-h3+

[Ras]

You conflate the limited and one-dimensional ML engineer/student with creative human thought.

As evidenced by engines being much stronger than humans, creative human thought is not only unnecessary for playing chess, it's inferior.

Spare us please the cyclic arguments about statistical winning/strength.

You failed to provide any actual definition of playing strength and think "winning more games" is cyclical? That's ridiculous.

[Ras]

Have you missed what's happened over the last few decades?

No, you have missed the point of my post.

[Viz]

All machine learning engineers and students and no chess has had the effect to optimise Stockfish into a bean counter.

Pretty chess doesn't win the games. Winning games does (d'uh). And winning games is how we determine the stronger player.

Reminds me of the same kind of thought when computers became stronger players than humans. Big disappointment because while that had been anticipated, the expectation had been that this would only happen when and because computers would understand chess in terms of human intelligence. That's why chess had been the poster child of the early AI research. Only that it turned out that human intelligence and understanding not only isn't required for chess, it's actually weaker than crunching numbers.

You conflate the limited and one-dimensional ML engineer/student with creative human thought. AB with material pruning is a horrible algorithm whose drawbacks are easily seen with these null move and fortress examples. If the engineers could figure out a way to "materialise" or generalise-compute the manifestations of these examples then they'ld have a better algorithm, only they can't and instead deny the problem. Spare us please the cyclic arguments about statistical winning/strength.

So basically if some patch tomorrow will magically solve this position in 2 seconds (which is a pretty usual occurence for this type of things - and as we can see this was the case for all previous SF versions) this "wouldn't count" because it will still use "horrible algo with easily seen drawbacks".
Also pls come up with a smarter algo that doesn't have this drawbacks, should stomp current AB engines easily.

[Viz]

And yes, all of talkchess dinosaurs always wanted for sf to be more human like.
Well there you go. In this position most of human players wouldn't even try to calculate some sort of fancy zugzwang mate because it's an easy technical win.

[chrisw]

All machine learning engineers and students and no chess has had the effect to optimise Stockfish into a bean counter.

Pretty chess doesn't win the games. Winning games does (d'uh). And winning games is how we determine the stronger player.

Code: Select all

Which is why 5-0-5 is “stronger” than 0-10-0, the former won five games

Reminds me of the same kind of thought when computers became stronger players than humans. Big disappointment because while that had been anticipated, the expectation had been that this would only happen when and because computers would understand chess in terms of human intelligence. That's why chess had been the poster child of the early AI research. Only that it turned out that human intelligence and understanding not only isn't required for chess, it's actually weaker than crunching numbers.

Only then in turned out that Alpha Zero destroyed SF8 with massive positional chess understanding and everybody dumped the false notion that chess was all about search depth and nps. Hyatt and Co were proved completely wrong.

[chrisw]

You conflate the limited and one-dimensional ML engineer/student with creative human thought.

As evidenced by engines being much stronger than humans, creative human thought is not only unnecessary for playing chess, it's inferior.

Spare us please the cyclic arguments about statistical winning/strength.

You failed to provide any actual definition of playing strength and think "winning more games" is cyclical? That's ridiculous.

It’s your argument that’s cyclical.

Well, I wrote that “strong”, used correctly, is about what’s needed to break something. SF breaks on fortresses, null move and others. In that sense it is quite weak and you’re content to leave it so because you can’t fix it (to date).

[chrisw]

All machine learning engineers and students and no chess has had the effect to optimise Stockfish into a bean counter.

Pretty chess doesn't win the games. Winning games does (d'uh). And winning games is how we determine the stronger player.

Reminds me of the same kind of thought when computers became stronger players than humans. Big disappointment because while that had been anticipated, the expectation had been that this would only happen when and because computers would understand chess in terms of human intelligence. That's why chess had been the poster child of the early AI research. Only that it turned out that human intelligence and understanding not only isn't required for chess, it's actually weaker than crunching numbers.

You conflate the limited and one-dimensional ML engineer/student with creative human thought. AB with material pruning is a horrible algorithm whose drawbacks are easily seen with these null move and fortress examples. If the engineers could figure out a way to "materialise" or generalise-compute the manifestations of these examples then they'ld have a better algorithm, only they can't and instead deny the problem. Spare us please the cyclic arguments about statistical winning/strength.

So basically if some patch tomorrow will magically solve this position in 2 seconds (which is a pretty usual occurence for this type of things - and as we can see this was the case for all previous SF versions) this "wouldn't count" because it will still use "horrible algo with easily seen drawbacks".
Also pls come up with a smarter algo that doesn't have this drawbacks, should stomp current AB engines easily.

Patches and bean counters go together it seems.

AZ doesn’t have “patches”, it appears to have fortress understanding and MCTS isn’t subject to nullmove last time I looked. It may be that the inclusion of eight blocks of past game history is the knowledge needed for fortress detection. These are not patches, they’re fundamental to the design. Whereas, fundamental to the design of AB as per SF, is either the use of badly guessed material adjustments for dealing with pruning, or else failure to estimate mating attacks at all. This getting worse recently as SF has become more and more bean counter, presumably due to the “patches”.