3

Answering this question, @SmallChess offered this answer:

Before Google's chess journey, the chess engine community believed: AI chess will not beat classical programming because neural networks (or other models) run much slower.

  1. I was looking for a reference to this quote. Did someone make this prediction in the past?

  2. I have another, perhaps more fundamental question related to another part of @SmallChess' answer:

AI machine learning is not new to chess. There were serious attempts before Google established. Unfortunately, nobody had the determination, skills and resources to do a good job.

Is it possible to pin down precisely what they have done better than others? Could it be simply that they use a lot more of computing ressources than before?

  1. Subsidiary question: For a layman, who doesn't know how to compare TPU
    and CPU, instead of 4 hours, how long would Alphazero's training have taken with Stockfish's configuration? (44 threads on 44 cores - two 2.2GHz Intel Xeon Broadwell CPUs with 22 cores instead of a single machine with 4 first-generation TPUs. See Supplementary Materials for their Science paper.)
  • 3
    You might be interested in the Leela Chess Zero project, which is attempting to duplicate AlphaZero using the information given in the papers. They've done a really good job, and Leela is almost as good as Stockfish now. You can read about the hardware they used as well on their website. lczero.org – Allure Jan 13 at 23:47
  • Just a note-AI didn't beat classical engine-in the match classical engine had many features cut of running on poor hardware and made mistakes obvious even to same engine on a descent phone... That was marketing action to attract money; there are still 5-15 years until AI for real will be able to compete with classical chess engine; but the time will come-history repeats-same was Human Vs PC-top GM's for money or because of bad conditions(PC don't get tired, Kasparov played with no rest days) lost to PCs that were not able to beat human, but since ~2015 PC for real became stronger than Human. – Drako Jan 14 at 7:13
  • @Drako have you read the Science paper and the supplementary material? They seemed to have improved a lot on the previous critics. – Kortchnoi Jan 14 at 8:55
  • 1
    I second the Leela Chess Zero project. Specifically to answer 2 though - Yes, they have a lot better hardware and resources at their disposal but because of the good hardware they were able to test the concept of "Zero". And it turns out that the AI in the longer run, learns to play better when it learns everything from scratch rather than learning from concepts that humans have learned from playing those games. They also created an architecture of Neural Network which is more general than just being good at 1 specific task, rather the architecture of the NN is generalized enough to learn.. – stackErr Jan 14 at 17:04
  • ..game of Go or chess. And possibly more. – stackErr Jan 14 at 17:04
1
  1. Pretty much everybody. The reason being that it is STILL slower. The evaluation function for alphazero has WAY more variables so despite more computing power, alphazero can only calculate a fraction of number of positions than stockfish. However the evaluation function for alphazero is much much better due to self-learning and also takes forever to calculate.

https://arxiv.org/pdf/1712.01815.pdf

Consider this. Google used 5000 super computers to train the AI to play before the game. Now consider what happens if you give stockfish access to 5000 super computers to think for 4 hours and if it memorized all those lines.

  1. The issue comes down to computing power.

  2. Not in your lifetime.

It is a bit difficult to find Ghz to tflop conversions but consider this.

https://www.pugetsystems.com/labs/hpc/Intel-Core-i9-7900X-and-7980XE-Skylake-X-Linux-Linpack-Performance-1059/

28 cores on 2.6Ghz gives about 1.123Tflops. So for 88 cores running at 2.2Ghz each is around 3Tflops. So your system gives around 3Tflops.

Now I can't find any information on the first generation TPUS but the second generation TPUs have 180 Tflops each.

https://en.wikipedia.org/wiki/Tensor_processing_unit

So 1 TPU = 60 of your systems. So 5000*60*4 = 1200000 hours = 136 years.

Also should clearly show how unfair the alphazero vs stockfish match is.

Edit: I made a mistake, you said 2 22 core systems. My calculations had 88 cores. So you would need to double the previous total to 272 years.

  • Thanks @Matthew Liu. 5000 super computers, are you sure? In the Supplementary Materials, they mention that "Each MCTS was executed on a single machine with 4 first-generation TPUs". So, based on your computation this would mean roughly 960 hours instead of 1200000! – Kortchnoi Jan 14 at 20:24
  • @Kortchnoi It clearly states in the paper they used 5000 first gen TPUs and 64 second gen TPUs. No idea about how good their first gen is but if its even half as good as their second gen it's still going to take somebodys lifetime to train on a normal machine. Also in the match the hardware alphazero was running on was literally over 100x faster than the one stockfish was on. Deepmind is trying to cover up that fact by saying alphazero evaluates way 1/1000 positions per second than stockfish, which is moot if alphazero's evaluation of the position is a million times better. – Matthew Liu Jan 14 at 20:40
  • thanks for the comment. Right but this is stated in the working paper version (Dec 2017) not in the published and refereed one (Dec 2018). I don't know if they used a new protocol or corrected a potential mistake. – Kortchnoi Jan 14 at 20:57
  • 1
    @Kortchnoi There were no serious changes made. Also I think you are mistaken about the hardware used for training and for the actual match vs stockfish. For training they used 5000 first gen tpus for training games and 64 second gen tpus. For the match they used 4 tpus against stockfish, which is literally over 100x more powerful than what stockfish was running on in that match. Yes they used 5000 super computers to train alphazero, not an impressive task when you can just throw money as a super rich company. – Matthew Liu Jan 14 at 21:40
  • 1
    I don't see why you consider the alphazero vs stockfish match so unfair because of what went into training alphazero. Training it is one thing, actually playing is another, and I don't see why anyone should pull their hardware punches when it comes to training. Why should DeepMind prepare an alphazero that won't be effective for 136 years? – Allure Jan 15 at 2:34
1

(1) Let's go back to 2016 when Google had just beaten the best Go player...

http://talkchess.com/forum3/viewtopic.php?f=2&t=59072&hilit=machine+learning+chess+slow+machine+learning&start=50

Milos who I believe is a PhD claimed machine learning had been tried "N" times.

enter image description here

Dr Bob, the legendary engine programmer who wrote Crafty believed NN was too slow.

enter image description here

(2)

Is it possible to pin down precisely what they have done better than others? Could it be simply that they use a lot more of computing ressources than before?

  • Google indeed had more computing resources but that's not all ...
  • Google had a superior deep learning model
  • Google adopted MCTS instead of pure alpha-beta
  • Google developed the model from self-playing, not like learning from GM games

This was the network used in NeuroChess. It's a shallow network in our modern standard. It's NOT a deep NN.

enter image description here

(3). I suspect it'd take weeks if not months. Not sure.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.