So Alpha Zero was a hoax?

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Harvey Williamson, bob

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
David Xu
Posts: 47
Joined: Mon Oct 31, 2016 8:45 pm

Re: So Alpha Zero was a hoax?

Post by David Xu » Sat Mar 17, 2018 6:59 pm

noobpwnftw wrote:
David Xu wrote:I refer you to the title of this thread. "So Alpha Zero was a hoax?" sounds pretty conspiracy theory-esque to me. If this is not your view, then I apologize for lumping you in with everyone else, but frankly, if you don't want the label "conspiracy theorist", you're going to need to explain why you called AlphaZero an "attention-seeking attempt" in the LCZero thread earlier.
You are contradicting yourself, I had my opinion(as one of the general public and not a part of the almighty shareholders) but what's wrong if it does not favor the A0 side?
If you don't understand my point then I will try to rephrase it all together here instead of you quoting lines here and there without context.
The A0 introduces an approach to use generalized algorithms to play many different games(truth), particularly in chess they released some information that the training worked(truth) and claim that it performed better than SF(inconclusive, doubtful and unscientific) in a 100-game match(truth).

David Xu wrote:I fail to see what is unscientific about the preprint. They described the architecture of the network itself, they described the training procedure, and they described the results against a specific setup for Stockfish--results which are in principle replicable by a third party, given the information they provided. That's what it takes for a result to be "scientific". More varied experiments would have been nice, certainly, but that fact hardly invalidates the results of the experiment they did perform.
Stating those facts does not equal to being scientific.
I can state the fact based on the 10 games that was published and conditions given, it is obvious and convincing that A0 would have a 100% win rate over SF, does it sound scientific to you?
There is a reason why we run those tournament matches instead of playing one on each side and decide the winner, why would this sound irrational to you when it comes to DM should've played more games especially on a short time control and have more variety to be scientific?
David Xu wrote:what is so unbelievable to you about this result that you would sooner postulate deception on the part of Google DeepMind than take them at their word?
If it is so true then why the easy part(as I described above) is not done so to avoid these doubts is unknown to me. I would think of them as if they'd built a boat but sailing in a pool.
David Xu wrote:Monte-Carlo tree search is certainly a well-known technique that has existed for years, but the technique of using it as a policy improvement operator is, as far as I'm aware, a novel one. This technique, of course, the critical aspect of DeepMind's approach, since it's what allowed them to generate such high-quality training data.
What is so different from people are constantly improving alpha-beta search by implementing things like PVS, MTF-d and refining search bounds in many different ways? Does your term "decades-old" include all these too?
David Xu wrote: What I am saying is that the issue of overfitting is well-known, so bringing it up as a specific argument against neural networks in chess is not a particularly strong objection.
This aspect is particularly more important for playing games like Chess as opposed to playing Go, where you are more likely to make one blunder move that would cost you the entire game. There is less tolerance of a wrong "likehood" estimation that a NN would normally give, so this results in less numbers of samples we can gather and correlate and more issues with overfitting.
I have my reasons to support my opinion than calling something I don't agree "disingenuous".
Very disappointing. I was hoping, after some of the previous posts you'd made, that you might have an actual point, but it seems my initial assessment was correct after all: you don't have an overall point, and you're just floundering around looking for excuses to deny that AlphaZero defeated Stockfish +28 =72 -0 in a 100-game match.

Let us be clear: at no point did DeepMind claim that AlphaZero "performed better than Stockfish". That is a vague, meaningless claim that you hallucinated them making, presumably because in your mind any claim of a positive result against Stockfish automatically translates into a claim of superiority over Stockfish. What DeepMind claimed is that AlphaZero achieved a positive score--specifically a score of 64%--against a very specific Stockfish version with a very specific setup. That is the extent of their claims, and it suffices to demonstrate what they sought to demonstrate: that a reinforcement learning-based approach is capable of competing with the traditional alpha-beta approach. Anything else is mere interpretation on your part.

Your claim that chess is somehow different from Go, likewise, would have been much more convincing had it been made before the AlphaZero results were published--though even then it would have reeked of special pleading. (Did you know that Go also has long tactical sequences, of which a single inaccuracy can cost a player the game? Chess is not unique in this regard.) As it is, however, you are claiming that a specific task is unlikely to be possible after that task has already been accomplished--which makes your claim nothing short of ridiculous.

You claim not to be a "conspiracy theorist", but the only possible way to make sense of your statements as a whole is if we assume that you believe DeepMind outright lied about their achievement against Stockfish. Only then could you get away with claiming that neural networks cannot perform in chess, and only then could you make the assertion that DeepMind's behavior was "unscientific". This, however, would mark you as a conspiracy theorist of the deepest kind, regardless of whether you accept that label. The alternative is that you have simply been saying contradictory nonsense.

Conspiracy theories or illogical nonsense--neither of these interest me as a potential source of discussion, unfortunately, so I'm afraid I'll have to withdraw for now.

User avatar
Eelco de Groot
Posts: 4103
Joined: Sun Mar 12, 2006 1:40 am
Location: Groningen

Re: So Alpha Zero was a hoax?

Post by Eelco de Groot » Sat Mar 17, 2018 7:30 pm

syzygy wrote:
CheckersGuy wrote:
Vizvezdenec wrote:Anyone who is working in anything remotedly close to science knows that you will never use smth like "it took ONLY 3 hours to..." and statements like that in any scientific-related paper unless you want to get fired.
So seeing statement like this used in pre-print constantly, using old fish, providing cherrypicked games, etc. makes any reasonable person who is not a bandwagoner to think about A0 paper how you really should - good PR project.
Sure, NNs are incredible, but will they make a revolution in chess engine development? My guess is no, at least with current level of NNs. And by the time NNs will need reasonable time to train on pure community enthusiasm, IMHO, alpha-beta engines will anyway play perfect chess.
So if it really did take 3 hours, they should have lied about it ? :D
I guess they then should have left out the word "only", which in fact they did not include :-)
If it really was more like 1700 (machine-)years, somebody calculated on the back of a cigarbox here, I don't know if that is in the right ballpark, but I see on Fishcooking that Fishtest to some measure now has 1000 machineyears... That would translate to much the same sort of progress, and Stockfish may actually be very close if not now, then in a few years time.
Debugging is twice as hard as writing the code in the first
place. Therefore, if you write the code as cleverly as possible, you
are, by definition, not smart enough to debug it.
-- Brian W. Kernighan

syzygy
Posts: 4426
Joined: Tue Feb 28, 2012 10:56 pm

Re: So Alpha Zero was a hoax?

Post by syzygy » Sat Mar 17, 2018 7:56 pm

Eelco de Groot wrote:
syzygy wrote:
CheckersGuy wrote:
Vizvezdenec wrote:Anyone who is working in anything remotedly close to science knows that you will never use smth like "it took ONLY 3 hours to..." and statements like that in any scientific-related paper unless you want to get fired.
So seeing statement like this used in pre-print constantly, using old fish, providing cherrypicked games, etc. makes any reasonable person who is not a bandwagoner to think about A0 paper how you really should - good PR project.
Sure, NNs are incredible, but will they make a revolution in chess engine development? My guess is no, at least with current level of NNs. And by the time NNs will need reasonable time to train on pure community enthusiasm, IMHO, alpha-beta engines will anyway play perfect chess.
So if it really did take 3 hours, they should have lied about it ? :D
I guess they then should have left out the word "only", which in fact they did not include :-)
If it really was more like 1700 (machine-)years, somebody calculated on the back of a cigarbox here, I don't know if that is in the right ballpark, but I see on Fishcooking that Fishtest to some measure now has 1000 machineyears... That would translate to much the same sort of progress, and Stockfish may actually be very close if not now, then in a few years time.
Translated to "Commodore 64"-years the number is even more impressive.

I think the paper does not say 3 hours but 4 hours to reach the point where AlphaZero surpassed Stockfish in strength. Those were really 4 hours and really not 1700 years, because we would all have been long dead if the latter number had been correct.

But my point was just to make fun of how people read stuff into the paper that is not there ("ONLY" 3/4 hours) and then criticise exactly that stuff that is NOT in the paper. Biased people are biased, what else is there to say.

Vizvezdenec
Posts: 52
Joined: Fri Jan 12, 2018 12:30 am

Re: So Alpha Zero was a hoax?

Post by Vizvezdenec » Sat Mar 17, 2018 8:33 pm

syzygy wrote:
Eelco de Groot wrote:
syzygy wrote:
CheckersGuy wrote:
Vizvezdenec wrote:Anyone who is working in anything remotedly close to science knows that you will never use smth like "it took ONLY 3 hours to..." and statements like that in any scientific-related paper unless you want to get fired.
So seeing statement like this used in pre-print constantly, using old fish, providing cherrypicked games, etc. makes any reasonable person who is not a bandwagoner to think about A0 paper how you really should - good PR project.
Sure, NNs are incredible, but will they make a revolution in chess engine development? My guess is no, at least with current level of NNs. And by the time NNs will need reasonable time to train on pure community enthusiasm, IMHO, alpha-beta engines will anyway play perfect chess.
So if it really did take 3 hours, they should have lied about it ? :D
I guess they then should have left out the word "only", which in fact they did not include :-)
If it really was more like 1700 (machine-)years, somebody calculated on the back of a cigarbox here, I don't know if that is in the right ballpark, but I see on Fishcooking that Fishtest to some measure now has 1000 machineyears... That would translate to much the same sort of progress, and Stockfish may actually be very close if not now, then in a few years time.
Translated to "Commodore 64"-years the number is even more impressive.

I think the paper does not say 3 hours but 4 hours to reach the point where AlphaZero surpassed Stockfish in strength. Those were really 4 hours and really not 1700 years, because we would all have been long dead if the latter number had been correct.

But my point was just to make fun of how people read stuff into the paper that is not there ("ONLY" 3/4 hours) and then criticise exactly that stuff that is NOT in the paper. Biased people are biased, what else is there to say.
"In chess, AlphaZero outperformed Stockfish after just 4 hours (300k steps)"
Okay, it was "just", but it's direct citation from that preprint.
As I already said, no person that are even close to working near science will ever use phrases like "it took just 4 hours to", you will always write "after 4 hours of work such a result was achieved".

Uri Blass
Posts: 8456
Joined: Wed Mar 08, 2006 11:37 pm
Location: Tel-Aviv Israel

Re: So Alpha Zero was a hoax?

Post by Uri Blass » Sat Mar 17, 2018 8:53 pm

Vizvezdenec wrote:
syzygy wrote:
Eelco de Groot wrote:
syzygy wrote:
CheckersGuy wrote:
Vizvezdenec wrote:Anyone who is working in anything remotedly close to science knows that you will never use smth like "it took ONLY 3 hours to..." and statements like that in any scientific-related paper unless you want to get fired.
So seeing statement like this used in pre-print constantly, using old fish, providing cherrypicked games, etc. makes any reasonable person who is not a bandwagoner to think about A0 paper how you really should - good PR project.
Sure, NNs are incredible, but will they make a revolution in chess engine development? My guess is no, at least with current level of NNs. And by the time NNs will need reasonable time to train on pure community enthusiasm, IMHO, alpha-beta engines will anyway play perfect chess.
So if it really did take 3 hours, they should have lied about it ? :D
I guess they then should have left out the word "only", which in fact they did not include :-)
If it really was more like 1700 (machine-)years, somebody calculated on the back of a cigarbox here, I don't know if that is in the right ballpark, but I see on Fishcooking that Fishtest to some measure now has 1000 machineyears... That would translate to much the same sort of progress, and Stockfish may actually be very close if not now, then in a few years time.
Translated to "Commodore 64"-years the number is even more impressive.

I think the paper does not say 3 hours but 4 hours to reach the point where AlphaZero surpassed Stockfish in strength. Those were really 4 hours and really not 1700 years, because we would all have been long dead if the latter number had been correct.

But my point was just to make fun of how people read stuff into the paper that is not there ("ONLY" 3/4 hours) and then criticise exactly that stuff that is NOT in the paper. Biased people are biased, what else is there to say.
"In chess, AlphaZero outperformed Stockfish after just 4 hours (300k steps)"
Okay, it was "just", but it's direct citation from that preprint.
As I already said, no person that are even close to working near science will ever use phrases like "it took just 4 hours to", you will always write "after 4 hours of work such a result was achieved".
I think that alphazero never outperformed stockfish(or at least it is not fair to say it) because they did not use a fair time control.

Remember that stockfish is optimized to be the best bullet engine(patches that pass are tested at 60+0.6 time control and usually only with one cpu).

Maybe stockfish could do better in long time control in case of using 600+6 time control in the framework and we do not know and maybe some ideas that failed in 60+0.6 could pass at 600+6.

Milos
Posts: 3348
Joined: Wed Nov 25, 2009 12:47 am

Re: So Alpha Zero was a hoax?

Post by Milos » Sat Mar 17, 2018 9:43 pm

syzygy wrote:Those were really 4 hours and really not 1700 years, because we would all have been long dead if the latter number had been correct.
Ofc not. Those were 20'000 TPU hours or 130'000 USD.
5'000 TPUs is extremely high number. Probably close to total number of TPUs that Google had one year ago. Anyone who thinks Google actually took all those 5'000 TPUs away from their datacenters even for 9 hours (the total training time) is, kindly speaking, naive.
The actual total training time was most probably much, much higher than those 4 (or 9) hours, but they simply scaled it in the paper to the number they'd get if they used all 5'000 TPUs at once. And then they even had guts to use construction such as "It took just 4 hours...".
Nothing but PR. Ppl who think that's the way to write a scientific paper simply haven't ever written one.

syzygy
Posts: 4426
Joined: Tue Feb 28, 2012 10:56 pm

Re: So Alpha Zero was a hoax?

Post by syzygy » Sat Mar 17, 2018 9:55 pm

Milos wrote:
syzygy wrote:Those were really 4 hours and really not 1700 years, because we would all have been long dead if the latter number had been correct.
Ofc not. Those were 20'000 TPU hours or 130'000 USD.
5'000 TPUs is extremely high number.
Sure, but the point is they did it in just 4 hours. Coming up with an approach that allows for massive parallelisation is a feat in itself.
The actual total training time was most probably much, much higher than those 4 (or 9) hours, but they simply scaled it in the paper to the number they'd get if they used all 5'000 TPUs at once.
So you are saying they are lying. Boring.

Milos
Posts: 3348
Joined: Wed Nov 25, 2009 12:47 am

Re: So Alpha Zero was a hoax?

Post by Milos » Sat Mar 17, 2018 10:03 pm

syzygy wrote:
Milos wrote:
syzygy wrote:Those were really 4 hours and really not 1700 years, because we would all have been long dead if the latter number had been correct.
Ofc not. Those were 20'000 TPU hours or 130'000 USD.
5'000 TPUs is extremely high number.
Sure, but the point is they did it in just 4 hours. Coming up with an approach that allows for massive parallelisation is a feat in itself.
They claim they did it in 4 hours. They could claim they did it in 2 minutes using 600'000 TPUs it would be equally meaningless claim.
Massive parallelisation of self-playing games is a trivial task. The fact that you see some feat in it speaks volumes.

syzygy
Posts: 4426
Joined: Tue Feb 28, 2012 10:56 pm

Re: So Alpha Zero was a hoax?

Post by syzygy » Sat Mar 17, 2018 10:15 pm

Vizvezdenec wrote:As I already said, no person that are even close to working near science will ever use phrases like "it took just 4 hours to", you will always write "after 4 hours of work such a result was achieved".
You have never seen a scientific paper where the authors write that their approach clearly outperform the known approaches?

syzygy
Posts: 4426
Joined: Tue Feb 28, 2012 10:56 pm

Re: So Alpha Zero was a hoax?

Post by syzygy » Sat Mar 17, 2018 10:25 pm

Milos wrote:
syzygy wrote:
Milos wrote:
syzygy wrote:Those were really 4 hours and really not 1700 years, because we would all have been long dead if the latter number had been correct.
Ofc not. Those were 20'000 TPU hours or 130'000 USD.
5'000 TPUs is extremely high number.
Sure, but the point is they did it in just 4 hours. Coming up with an approach that allows for massive parallelisation is a feat in itself.
They claim they did it in 4 hours. They could claim they did it in 2 minutes using 600'000 TPUs it would be equally meaningless claim.
Massive parallelisation of self-playing games is a trivial task. The fact that you see some feat in it speaks volumes.
You are spouting baseless claims but what is new. I'm not sure why your are not simply stating that they faked all their games, which would be easy enough to do.

It is not easy at all to throw massively parallel hardware at a task. Alpha-beta does not seem to scale at all beyond 64 threads, to give just one example.

The AlphaZero approach does not scale endlessly either, as there will be limits on parallelising the adjustment of the weights based on the training games they produced. But it scales well enough that they were able to surpass Stockfish in just four hours.

Post Reply