Google's AlphaGo team has been working on chess

[jorose]

Very cool! I am especially surprised they still relied on a MCTS approach in chess. I don't think anybody can actually reproduce these results at the moment with hardware at home but this certainly marks a significant shift in how computer chess will develop.

I am curious what kind of performance their program would be able to achieve on sub 2k off the shelf commercial hardware. Considering the power of their TPUs I imagine the penalty would be pretty huge. Regardless, commercial hardware is a question of when, and not if. Perhaps someone will improve their approach specifically for chess in some way?

I am curious if the same amount of people will work on the tinkering form of chess programming.

[Rémi Coulom]

Money would be a better measure.

The AlphaZero training system costed $ 4 millions of hardware. (figures given for alpha go zero, don't have source under hand)

The paper says they use 5,000 first-generation TPUs, and 64 second-generation TPUs. Such hardware is not available for sale, but might be similar to a V100 in terms of computing power. A single PCI V100 costs about 10,000 Euros in Europe. But if you buy 5,000, you can certainly get a much cheaper price. Of course you also need the computers that host them, and the power supply (250W*5,000 = 1.25 MW).

This being said, I would not be surprised if their trained network could still beat Stockfish on ordinary hardware. And I expect deep-learning hardware will become much cheaper and commonplace in the future. Even cell-phones are starting to have deep-learning hardware now.

A distributed open-source effort might be enough to produce a super-strong network in a few months. This is what Gian-Carlo has started with Leela in Go. Maybe he'll do it for chess and shogi, too.

[mcostalba]

I have read the paper: result is impressive!

Honestly I didn't think it was possible because my understanding was that chess is more "computer friendly" than Go....I was wrong.

It is true, SF is not meant to play at its best without a book and especially 1 fixed minute per move cuts out the whole time management, it would be more natural to play with tournament conditions, but nevertheless I think these are secondary aspects, what has been accomplished is huge.

[Michel]

Money would be a better measure.

The AlphaZero training system costed $ 4 millions of hardware. (figures given for alpha go zero, don't have source under hand)

The paper says they use 5,000 first-generation TPUs, and 64 second-generation TPUs. Such hardware is not available for sale, but might be similar to a V100 in terms of computing power. A single PCI V100 costs about 10,000 Euros in Europe. But if you buy 5,000, you can certainly get a much cheaper price. Of course you also need the computers that host them, and the power supply (250W*5,000 = 1.25 MW).

This being said, I would not be surprised if their trained network could still beat Stockfish on ordinary hardware. And I expect deep-learning hardware will become much cheaper and commonplace in the future. Even cell-phones are starting to have deep-learning hardware now.

A distributed open-source effort might be enough to produce a super-strong network in a few months. This is what Gian-Carlo has started with Leela in Go. Maybe he'll do it for chess and shogi, too.

I have a question that perhaps you can answer right away.

Almost a 1000 CPU years went into tuning SF until today....

Would you say that the training of AlphaGo required less or more resources than this?

[Rémi Coulom]

I have a question that perhaps you can answer right away.

Almost a 1000 CPU years went into tuning SF until today....

Would you say that the training of AlphaGo required less or more resources than this?

According to the paper, they trained for 9 hours, over 5000 TPUs.

5000 * 9 / 24 = 1875 TPU-days

A TPU is a bit like a super-powerful GPU. A very rough estimate may be that 10 GTX 1080 ti may have the power of a TPU. So if you get 100 people volunteering their GPU full time, that would take about 6 months. That looks doable.

[jdart]

Stockfish does such heavy pruning that it is throwing away most of the nodes in its search trees. But the ones it does search, it searches very deeply. I see a lot of high-level computer games won by tactics or by endgame play that requires deep search. Shannon Type II (selective search) has never worked well in any of the past 5-6 decades. But maybe this effort is showing that eval is more important than has been thought, and search less important.

--Jon

[Henk]

What I understand is that the neural network predicts the winning probabilities for each valid move in a position.

Don't understand that these predictions will be good if it doesn't do a search but only simulation.

So how is it possible that monte carlo simulation is better than an alpha beta search.

[clumma]

What I understand is that the neural network predicts the winning probabilities for each valid move in a position.

Don't understand that these predictions will be good if it doesn't do a search but only simulation.

So how is it possible that monte carlo simulation is better than an alpha beta search.

The trick is to stop thinking in terms of tactics and search, and start thinking in terms of learning a really complex evaluation function. As the paper explains, alpha-beta can amplify any error in the evaluation function, whereas MCTS (plus a little noise) averages it out. So tuning alpha-beta is, in a sense, harder.

-Carl

[clumma]

Time is misleading in DeepMind's papers, as they use thousands of "computers" (not even commercially available). Money would be a better measure.

Neural nets scale well on cheap (per FLOP) SIMD hardware, so you'll definitely pay more for the traditional engine at the same level of performance.

-Carl

[Uri Blass]

Most of us here suspected that this could happen once Giraffe showed it can beat Stockfish's eval.

Just the fact that the new approch to chess programming worked incredibly well is fantastic even if it didn't beat the best.

Daniel

How do you decide that giraffe's evaluation is better than stockfish?
If the definition is by using fixed number of nodes with the same search function then having a better evaluation than stockfish is easy if you do not care about time.

Evaluation is basically a function that take a position and return a number.

In this case I can define the evaluation of the position to be the result of the search of stockfish when it searches 20 plies forward.

I am sure this evaluation is better than stockfish's evaluation and can beat stockfish's evaluation when you search the same number of nodes(of course you do not count the nodes that you search to calculate the evaluation because they are defined to be part of the evaluation).

Uri