This is not A.I. by the way. People tend to use that term without knowing what it means. So far, A.I. is a pipe dream. This thing just constructed a database of positions.
Stockfish dethroned
Sort:
Has anyone tried to estimate the playing strength of Stockfish used in the match? 64-36 may look like a huge margin but it only translates to a difference of 100 elo. If Stockfish is playing at 3000, then AlphaZero would be 3100. 3100 is not that impressive for an engine.
Also, just spotted this from Wikipedia:
In a series of twelve 100-game matches starting from the 12 most popular human openings, AlphaZero won 290, drew 886 and lost 24.
So AlphaZero can still lose once it is forced into specific human openings (which may be faulty and/or give Stockfish better chances than no book). Even if AlphaZero is not playing perfect chess, it has led to a new generation of chess engines that can finally play the opening without crutches.
Elroch wrote:
Ummm...you are aware that the elo rating system is a relative rating system and only measures how someone in a pool will fare against someone else in that pool, right?
Yes, I know that, but it is designed to be consistent, and is empirically roughly so.. i.e., on average if A, B and C have three ratings, all three of the differences give a good estimation of expected results.
Elo ratings are not tied to actual chess skill, nor to chess at all. If Elo ratings are based on chess games, they quantify RELATIVE chess skill. It is irrelevant that they can also be used for other games.
You can make a rating pool for almost any competitive endeavor using it. If only 5 year olds played chess, someone would still have Carlsen's rating eventually.
It is unlikely that the spread of standards of play of 5 year olds would be as large as the spread of standard of play of all chess players. So you would probably get a bell-shaped curve of lesser width, and no-one who was 1600 points above the average.
You are proving my point. When the first engines were 2800, it was impossible to have a 3600 rating...
It was not impossible, it had just not happened yet. If you took a 3600 rated engine and played it against the engines that existed at that time, it could eventually achieve 3600 (by winning virtually every game it played).
it took years and years to work it's way slowly up a few ratings points at a time...exactly the way I said that 5000 is impossible right now while the best engines are 3400.
You don't use the word impossible in the correct way. You misuse it to to mean what happens to have been done at a particular time. This stops it being the absolute that it is when used correctly.
At 400 ratings points difference the rate of increase slows to single ratings points...so....going from 3600 to 5000 at that rate would take how long? Every draw sends you tumbling back, of course. I understand the system just fine. Move along.
Yes, it would take a lot of games. Are you now saying it is "impossible" to play a lot of games?
If the Elo system works, a sufficient difference in ratings makes draws unlikely enough to allow the correct rating difference to be achieved after enough games.
It is impossible to get a 5000 elo rating in the chess world as it stands right now. Since the post I was answering was speculating on AlphaZero's current rating level, the word impossible clearly works just fine 
...
A 5000 rating is not currently possible. Neither is Wesley So's (I think ity was So, anyway) guess of 4000 rating. This is something known as "impossible". It may never be possible, given that at some point ratings rules may change to disallow engines from step-laddering up their ratings by playing each other millions of games at high speed gaining a single point each time.
It doesn't matter what AlphaZero's actual playing strength is. The elo rating system is designed for relative comparison, not discrete measurement.
It's like saying the size of the expanding big bang is infinite. It is not. It is a finite entity expanding into a theoretically infinite expanse. But right now, today, it is finite. Tomorrow, it will be finite, and when the heat death of the universe comes, it will still be finite.
Years or decades from now if the pool expands (attenuates, really) to include 5000 ratings, and then somebody writes a post stating that the potential rating of XYZ engine should 10000, I will tell them that is impossible, and I will be right...again.
Dec 9, 2017
0
#105
Optimissed wrote:
This is not A.I. by the way. People tend to use that term without knowing what it means. So far, A.I. is a pipe dream. This thing just constructed a database of positions.
Most people who are expert on AI would disagree with the first three sentences, but you could argue this is a matter of semantics. The way this AI learnt chess is currently a hot topic to those who study how the brain works, which is a good indication of why it deserves the term.
The last sentence is simply wrong and could hardly be further from the truth. AlphaZero only stores positions to the minimum degree necessary to play chess: it doesn't even have an opening book (or any tablebase). All of the functionality is incorporated in the parameters of the neural networks, which encapsulate the concepts it learns about chess and how they related to each other. Very, very different to positions (which are like what it observes and "imagines" as it analyses lines that look appealing to its networks, somewhere about half way between the way a conventional computer does this and a human does it).
It should probably be labelled as "Computational Intelligence", although the definition of this term is rather disputed.
Dec 9, 2017
0
#107
btickler wrote:
It is impossible to get a 5000 elo rating in the chess world as it stands right now.
Not strictly speaking. 5000 Elo corresponds to 99.983% score against a 3500 rated engine. So, it would involve a dull succession of thousands of wins against AlphaZero. Against Carlsen, it would take considerably more. This is wildly unlikely, but the unlikeliness of this does not make it impossible: it makes it (very) IMPROBABLE. We do know that AlphaZero can lose from reasonable-looking positions: it did so against Stockfish when forced to play standard openings it wasn't very keen on.
It may be that even a perfect chess player would not have a true rating as high as 5000: no-one knows the answer to that question for sure, but many would guess so. However, guesses (or even good guesses) are not enough to prove something impossible.
Elroch wrote:
Optimissed wrote:
This is not A.I. by the way. People tend to use that term without knowing what it means. So far, A.I. is a pipe dream. This thing just constructed a database of positions.
Most people who are expert on AI would disagree with the first three sentences, but you could argue this is a matter of semantics. The way this AI learnt chess is currently a hot topic to those who study how the brain works, which is a good indication of why it deserves the term.
The last sentence is simply wrong and could hardly be further from the truth. AlphaZero only stores positions to the minimum degree necessary to play chess: it doesn't even have an opening book (or any tablebase). All of the functionality is incorporated in the parameters of the neural networks, which encapsulate the concepts it learns about chess and how they related to each other. Very, very different to positions (which are like what it observes and "imagines" as it analyses lines that look appealing to its networks, somewhere about half way between the way a conventional computer does this and a human does it).
Agreed, third sentence is garbage and shows no understanding of what AlphaZero actually does. Google uses the term "machine learning" for a reason. That reason is to distinguish their current endeavors from actual AI. Machine learning may be a key linchpin in achieving AI...but it is not AI.
Dec 9, 2017
0
#109
As my first point said, here it comes down to semantics. Most specialists call this a limited form of AI and all acknowledge it is not the strongest form of AI - an entity as adaptable as a human or more so.
Elroch wrote:
It may be that even a perfect chess player would not have a true rating as high as 5000: no-one knows the answer to that question for sure, but many would guess so. However, guesses (or even good guesses) are not enough to make something impossible.
Again, actual playing strength is irrelevant. If a race of omnipotent/omniscient beings were being rated and they started in a pool at 1200, they would remain at 1200 even though every one of them is a "perfect" chessplayer.
Your argument is akin to Spinal Tap..."but this amplifier goes to 11...". No, it doesn't...the scale is currently limited to 10, by definition. The ratings bubble right now as it exists traverses 0-3600-ish. It will slowly expand...but not this week . So...speculation of 4000-5000 rating is absurd and impossible because that's not even on the viable scale right now.
Dec 9, 2017
0
#111
btickler wrote:
Elroch wrote:
It may be that even a perfect chess player would not have a true rating as high as 5000: no-one knows the answer to that question for sure, but many would guess so. However, guesses (or even good guesses) are not enough to make something impossible.
Again, actual playing strength is irrelevant. If a race of omnipotent/omniscient beings were being rated and they started in a pool at 1200, they would remain at 1200 even though every one of them is a "perfect" chessplayer.
Correct.
Your argument is akin to Spinal Tap..."but this amplifier goes to 11...". No, it doesn't...the scale is currently limited to 10, by definition. The ratings bubble right now as it exists traverses 0-3600-ish. It will slowly expand...but not this week . So...speculation of 4000-5000 rating is absurd and impossible because that's not even on the viable scale right now.
The ratings scale defined by Professor Elo's mathematical model of the statistics of a two player zero sum game extends from minus infinity to infinity: It has an arbitrary offset, as you point out (and most people know). There is surely some limit in both directions: in the upper direction, a perfect player has some finite rating, in the lower direction there is a similar concept for a player who obeys the same rules as chess except they count a win as -1 and a loss as +1. It seems clear that any two players with this objective could quite easily achieve a draw against each other, but they could lose most of the time against abysmal players who were not actively trying to lose.
When I talk about absolute ratings rather than relative ratings, you can assume I mean relative to the scale which is used by FIDE for human players. The CCRL scale for computers is not very different to this, as the early ratings for computers were based on play against humans.
pfren wrote:
mcris έγραψε:
Stockfish 8 is not an old version, it is last, I just visited their website. Although I agree that 1 GB of RAM is lowering its capability.
Stockfish 8 is eleven months and a few days old, genius.
The latest Stockfish version changes almost every day, and can be found here: http://abrok.eu/stockfish/
To be more precise, 1 year, 1 month and couple of days older.
Elroch wrote:
The DeepMind paper refers to a 1 GB hash table, not to any RAM limitation.
Meanwhile, here are the reactions of some of the world's top chess players, plus a Stockfish developer:
https://www.chess.com/news/view/alphazero-reactions-from-top-gms-stockfish-author
Of whom, only Aronian, Karyakin and Nakamura have a clue.
Tord has been out of the business for a very long time, though he is still considered the main developer.
Elroch wrote:
As my first point said, here it comes down to semantics. Most specialists call this a limited form of AI and all acknowledge it is not the strongest form of AI - an entity as adaptable as a human or more so.
Intelligence is encountered only in sentient human beings, capable of adapting.
AI is some kind of gross misunderstanding.
Spoken like someone who's never seen the videos of crows making tools.
Dec 9, 2017
0
#117
Lyudmil_Tsvetkov wrote:
Elroch wrote:
As my first point said, here it comes down to semantics. Most specialists call this a limited form of AI and all acknowledge it is not the strongest form of AI - an entity as adaptable as a human or more so.
Intelligence is encountered only in sentient human beings, capable of adapting.
AI is some kind of gross misunderstanding.
Said the neanderthal.
Humans are the ones approximating intelligence. Just look at all the dumb stuff in the world.
An intelligence beyond humans is possible with AI, and one step closer with AlphaZero.
Dec 10, 2017
0
#118
pfren wrote:
mcris έγραψε:
Stockfish 8 is not an old version, it is last, I just visited their website. Although I agree that 1 GB of RAM is lowering its capability.
Stockfish 8 is eleven months and a few days old, genius.
The latest Stockfish version changes almost every day, and can be found here: http://abrok.eu/stockfish/
Sorry, but are those Stockfish 9? Until then...
Dec 10, 2017
0
#119
mcris wrote:
pfren wrote:
mcris έγραψε:
Stockfish 8 is not an old version, it is last, I just visited their website. Although I agree that 1 GB of RAM is lowering its capability.
Stockfish 8 is eleven months and a few days old, genius.
The latest Stockfish version changes almost every day, and can be found here: http://abrok.eu/stockfish/
Sorry, but are those Stockfish 9? Until then...
Another idiot thinks they made a meaningful comment.
Stockfish 8 is not an old version, it is last, I just visited their website. Although I agree that 1 GB of RAM is lowering its capability.
Stockfish 8 is eleven months and a few days old, genius.
The latest Stockfish version changes almost every day, and can be found here: http://abrok.eu/stockfish/