Objectively Speaking, Is Magnus a Patzer Compared to StockFish and AlphaZero?

SmyslovFan,

What do you think this portends for the near future of competitive OTB chess by humans, and the interest in competitive OTB chess by humans?  

Increase, same, or decrease?

Obviously the rules of the game are inputted, and the objective (winning). Yes, it starts with that. But beyond that . . . ? The neural network figures out the rest.

And yes, obviously there's a lot of computing involved to create Alpha Zero. But it's also been directed to self-learn and create its own knowledge—which is the key breakthrough.

The purpose of AlphaGoZero is really to test the proficiency of self-learning AI. The future goal of DeepMind is to apply this technology to find advances in medicine that humans haven't been able to find on our own.

Chess is just a rules-based game that provides a nice opportunity to test the AI's ability to self-learn within a closed system.

Here's an interesting article about it: https://medium.com/intuitionmachine/the-strange-loop-in-alphago-zeros-self-play-6e3274fcdd9f

Elroch:  "The computer went from pathetic beginner to the best chess player in the world by self-learning without the slightest interference from humans."

BoggleMeBrains (from another thread):  "I think all the quibbling over what restrictions were placed on Stockfish is beside the point.  Forget Stockfish.  The real story about AlphaZero isn't that it's better than Stockfish, it's that it represents the point at which machines truly surpassed humans at chess.  Conventional chess engines need to have their entire evaluation function provided for them by humans.  They may play better than humans, but are really nothing more than human knowledge running on fast hardware.  AlphaZero surpassed human chess knowledge entirely by itself.  It designed its own evaluation function.  The achievement is already beyond impressive even if Stockfish running on a supercomputer with opening books and tablebases might have proved stronger.  People fixating on whether the games were fair are kind of missing the true marvel of technology we've witnessed here.  Something with no human-provided knowledge can out-Karpov Karpov."

What does this portend for the near future of competitive OTB chess by humans?

Deep Blue did not kill it.  But will AlphaZero?

IMO, tablebases would increase the strength of AlphaZero, but you would need a large database of computer chess games to make an opening book that would have any chance of competing with what AlphaZero has learnt to play. The problem is that a hundred games by weaker players could be refuted by innovations in AlphaZero's on the fly analysis (which itself are influenced by its own experience in a large number of self-play games, many at a very high standard.

It wasnt playing on a super computer lol. It was playing on what you can pick up at best buy looking at a measly 80k positions per second while SF was calculating 70+million .. Smh. People read nothing... 

The engines of today are bound by their human handlers and rely on brute force alone to find moves rather than breaking out of their constraints and actually learning chess. 

These are some of the dumbest comments i have seen on chess.com ever. 

It used 4 of googles tensor processing units (TPUs) which might be equivalent to about 1000 Intel cores (say 14 of the latest 72-core Xeons) for the purpose of neural network computations.

It's fair to say this is a super computer, although not on the scale of the really big ones used for big science.

The games were spectacular, but Smyslov would not quite have thought this way.

"...it would be possible to write an AI..."

Well, you said it yourself, SOMEONE has written/is writing the AI.

AI is a non-concept. Human beings can adapt, that is a big difference. Adapt, due to the good/bad surrounding conditions. There is no intelligence without pain and happiness. Computers don't have intelligence.

Those TPUs cost half a million bucks.

The number of positions calculated is completely irrelevant.

It is much slower, as it evaluates many more things on a single node, but that does not mean it does that in a great way, only that it does so. One can significantly decrease speed even by doing different types of unnecessary calculations. Most probably, it has been checking its Value[piece][from][to] array

throughout nodes.

That is rigth, this was a Supercomputer versus a PC.

So, the achievement in terms of AI was rather small, definitely, much lower than SF.

This is the crux of your argument, a logical fallacy:

Artificial Intelligence is the ability to adapt.

Computers cannot have artificial intelligence, therefore they cannot adapt. 

AlphaZero has adaptive learning. 

Therefore, AlphaZero must be a fraud. 

Q.E.D.

Facts don't matter in such a construct.

It is far away from being a super computer in all definitions. It is a single machine, that by the way is very energy efficient.

NN computations work very well on TPUs, better than on GPUs and much better than on CPUs. For the type of computation it is about 15 to 30 times faster (sayed by Google about TPUv2), so remove the last zero from your 1000 to be more realistic. Because of recursion and branching they are useless for conventional chess engines which run much better on CPUs. And to have 64 cores is not an insignificant amount either.

It would be possible to optimize both sites, the Stockfish setup or specialize the generic approach of AlphaZero. But to show that Stockfish could be beaten was not the point of the work. It's about a completely self learning NN that can reach top level—something that was doubted before because of the very tactical nature of chess.

It's nice to see that human opening theory was reconfirmed by an algorithm without any human input. Despite of that look at the games and you will see something special happened there. (Or to be more precise: DeepMind surprised the Go world one or two years before, and now they are doing it again in the chess and Shogi world.)

It would nice to see some of AlphaZero's Shogi games and some chess games played during different learning phases. We had saw before that DeepMind released a lot of data to the Go world, so it is not unlikely that we will see a little bit more of AlphaZero's chess games in the future. But it will be on the chess programmers to create something you can call StockfishZero or something like that. In Go there is already a project called LeelaZero: http://zero.sjeng.org/

Seeing some numbers like 5.000.000 games, three days of computation (Go) is one thing—but you get a feeling of it how much computation is necessary to create games and train a good NN if you try do it yourself. And even 4 hours of computation on 5064 TPUs is an awful lot, that's for sure.

Alphazero neither played in nor won a TCEC championship. Houdini overtook Komodo as championship with Stockfish coming third.

AlphaZero isn't a chess engine. It's a self-learning neural network.

You can't have AlphaZero compete fairly in the TCEC because it keeps rewriting itself, learning and becoming stronger with every game.

Chess engines don't have this ability. They're just stuck playing at their maximum strength, with no ability to improve beyond their programming.

And we ain't seen  nothin' yet.   Come back in 100 years.  Well...aside from the fact most of us will be  well into our dirtnaps...probably nothing to comeback to but radioactive ruin as far as the eye could see if there there was any eye.   Human  eye that is.   One of  AlphaZero  great-great-////-great grandkid's  eyes maybe.   

So what percentage of the time DO you think being up a rook in an otherwise normal position, in a game between > 1500 players, is a win?   That is just a quibble. 

OK, so AZ would do pretty well even if it was not allowed to do any further analysis.  That implies that AZ can evaluate any position, and it learned to do this solely by playing (initially) random games. 

I guess it may be that this is the kind of question that can't be answered in a blog post, but what I am trying to figure out is the form of that evaluation method and how it gets built.

The last slight is ridiculous for more than one reason. One of them is that Stockfish is not an AI, it is strong because its designers are experts in the design of fast, strong chess engines. The DeepMind team include no experts on chess or engines, because they didn't design one. AlphaZero did.

The computational demands of AlphaZero are entirely because the networks are large and deep. These involve sizeable matrix operations at every step of training and of application in a game. The programmers gave these networks no chess knowledge: rather this is where AlphaZero stores the knowledge about chess that it derives from its experience.

The nature of the evaluation method is quite simple. It has some sort of representation of the position as an array of numbers which are the inputs to the neural network  - the neural network doesn't know what they mean, it has to work this out from they way the relate to the results of games and to their values on other moves - and a large deep neural network with thousands (not sure how many thousands) of nodes in many layers which take the representation of the board and output a number, the expected score from the position. [I hope I haven't missed some published detail].

How the evaluation method gets built comprises two parts (if my understanding is correct - I am supplementing what is published with general ideas about deep reinforcement learning). The obvious one is when a game ends: the exact value of the position is available, and that can be used to adjust the neural network to improve its evaluations of earlier positions in the direction of the right result. The second one is that when it evaluates a position, if this evaluation is a surprise compared to the evaluation of previous positions, the network is tweaked to make the evaluations of previous positions a bit more in agreement with the later evaluation.

The first form of feedback is basically making the evaluation compatible with the absolute value of clear positions. The second form of feedback is basically making the evaluation compatible with the legal continuations in a position: the reason is that the perfect evaluation of a position is the same as that of a later position reached by perfect play.

[Technical points are (1) that this adjustment takes the form of nudging every parameter in the direction that would achieve the desired change proportional to how strongly it does so. This is called "gradient descent" (2) it actually takes a minibatch of 4096 positions and combines the desired adjustments to make one change to the entire neural network. the reason for this is that it allows more parallel computation to speed up the process, otherwise it would be just as good to do one at a time].

[I should correct one careless thing I have said sometimes: there is actually only one neural network which takes a board as input - this provides both the expected score and an array of probabilities that each of the legal moves is best. This is better than two networks, because some parts of the network can be used for both purposes].

Here you are factually wrong.

The team includes at least 3 chess programmers. Matthew Lai, the author of Giraffe and Talkchess member, is one of them. It is maybe for a reason that Giraffe, following the very same approach as Alpha, is rated only around 2400 on single core.

It is the huge hardware that made the difference and not the approach.

Self-learning, self-learning, what do you mean self-learning and AI.

It is just tuning itself on multiple levels, while still following the primary human code, that is all.

And the primary human code included at least that winning is good and that, in case of a win, certain terms/psqt should be bonised, given larger values.

What is working throughout the iterations is the human code, including by how much should those values be increased. Alpha is just executing the code and returning its findings.