recent article on alphazero ... 12/11/2017 ...

[Uri Blass]

No, the point is, they are not actually capable of beating SF in fair and optimal conditions for SF,

'Fair and optimal conditions' meaning that SF should not have been forced to play the moves by itself, but some other entity (namely a book) should have been allowed to play the moves instead...

which would mean they wouldn't have generated nearly as much publicity as they did in case of dominant victory, therefore they used totally immoral approach of crippling SF in any way possible that is not immediately obvious and using absolutely unfair comparison to obtain that marketing goal.

Yeah, sure. It is very crippling when you have to find your own moves, just as the opponent does. Or when you have to play at fixed time per move, just as your opponent does. In fact anything that doesn't rig the odds massively in your favor would be highly unfair. After all, Stockfish is the TCEC champion. How dare they subject it to the same conditions as the opponent!

Note that stockfish is not optimized to play at fixed time per move and I believe that it is possible to change the program to play better at fixed time per move.

For example in case of a fail low that the program see that it probably has no time to solve maybe it is better to do a second search at reduced depth to come with a different move(this idea at least need testing in games).

In normal time control usually it does not happen and stockfish can finish the iteration so changing it may not pass SPRT in the framework but in fixed time per move things are different.

I am also not sure if stockfish take the correct decision in case of fail high that it did not solve and part of the fail high are wrong fail high so the question is if to trust the fail high and play a new move or not to do it.


A third idea is practically to decide about the move that you are going to play earlier than 1 minute per move if you finish iteration close enough to 1 minute and for example if you finish iteration in 50 seconds start calculating the next move based on the move that you expect in the next 10 seconds.

Maybe the stockfish team could test these ideas in the framework if the google team could tell them that they are going to test stockfish at fixed time per move some months before the match.

[hgm]

Obviously fixed time per move is a less friendly TC than classical or incremental. And most likely specialized tricks to perform better at it would be possible. If you care about that performance.

But AlphaZero also did not use such tricks. It just searched a fixed amount of time of the current position, not caring whether the move it had to play was already determined beyond any doubt is 10% of that time, or whether the position was so complex that it would have liked to search 3 times longer to get sufficient accuracy.

[Milos]

That is how one converts capital investment to cost of use. The M$50 equipment cost was your number.

This is how one like you, who has absolutely no clue about finance (btw. as demonstrated at various occasions in CTF) make some calculations.
Ppl who understand even the basics of finance don't write such ridiculous comparisons.

Ummm... The way I make estimations in my field uses more conventional arithmetic, where 5'000TPUs x (9h+12h+34h) x 1$/h = 275k$. Rather than 275000k$. And note that I was just addressing the 9hr Chess part.

Wow, you found a typo, how clever mister Harm.
It doesn't matter what you were addressing, when it is clearly wrong (in addition to your ridiculous interest rate calculation) coz Shogi and Go training was not done by aliens but also by Google.

[Milos]

But AlphaZero also did not use such tricks. It just searched a fixed amount of time of the current position, not caring whether the move it had to play was already determined beyond any doubt is 10% of that time, or whether the position was so complex that it would have liked to search 3 times longer to get sufficient accuracy.

Alpha0 didn't use such tricks because ppl from Google either didn't know how to make better time manager for it, or they simply knew that no matter how they write time manager the amount of Elo gained is far smaller than what SF loses due to idiotic TC.
Alpha0 almost certainly didn't have time manager. SF does have a good one.
Using time control that doesn't require any time manager making sure you are not handicapped but in the same time handicapping your opponent is low and pretty ugly. Those ppl from DeepMind are not idiots and they most certainly understood what the choice of TC meant for A0 and SF.
So those excuses for kids under 3 years old, you should leave for your grand kids (if you have them), ppl at this forum are not that stupid as you think of them Mr. Harm.

[hgm]

Or they they simply considered time managing not challenging enough to waste their time on...

Note that fixed time per move is effectively the predominant TC used in Shogi and Go games, where initially all time you use is counted against the initial time on your clock, after which you enter byoyomi phase, which is fixed time per move. The initial time divided by the typical number of moves in a game is usually much lower than the byoyomi time, so there is very little advantage in having it at all. It will be used up by the time the game gets interesting.

Why should they bother to make something special just for Chess, which they more or less did as a side dish to the more interesting stuff?

[Milos]

Or they they simply considered time managing not challenging enough to waste their time on...

Note that fixed time per move is effectively the predominant TC used in Shogi and Go games, where initially all time you use is counted against the initial time on your clock, after which you enter byoyomi phase, which is fixed time per move. The initial time divided by the typical number of moves in a game is usually much lower than the byoyomi time, so there is very little advantage in having it at all. It will be used up by the time the game gets interesting.

Why should they bother to make something special just for Chess, which they more or less did as a side dish to the more interesting stuff?

Excuses, excuses all so "convenient", but all pointless.
Let me repeat again:

The amount of Elo gained (by TM for A0) is far smaller than what SF loses due to idiotic TC, and

Those ppl from DeepMind are not idiots and they most certainly understood what the choice of TC meant for A0 and SF.

So talking about random convenience when it was in fact a deliberate choice shows nothing but weak moral principles.

And this comment about chess as a side dish, you really have a sense of humor Mr. Harm, despite being a physicist .
No one would even bother to read that paper if there was no chess in it (the paper that was so rushed and awfully written just to make it published during London Chess Classic where it would generate the most publicity), the same as actually no one in this forum bothered to read AGZ paper.
Sometimes I wonder, are you really that naive, or you just think that other ppl are nothing but idiots assuming that they would believe your "arguments"???

[abulmo2]

Alpha0 didn't use such tricks because ppl from Google either didn't know how to make better time manager for it, or they simply knew that no matter how they write time manager the amount of Elo gained is far smaller than what SF loses due to idiotic TC.

Or it could be the opposite. The output of the NN and MCTS obviously allows a far more accurate time manager than what can be done in an alphabeta searcher, and they did not want to pollute their results with another source of superiority.
And the same can be said about the opening book. Given the calculus power of Google, they probably can compute an opening book far superior than that of Cerebellum/BrainFish. And using it will also just pollute their results.
They just wanted to demonstrate that mcts + NN on appropriate hardware is better than alphabeta + traditionnal eval with the latest release (yes SF8 IS the latest release, the nightly build of a development version is not a release) of one the strongest chess engine on a strong current hardware.

[hgm]

Sometimes I wonder, are you really that naive, or you just think that other ppl are nothing but idiots assuming that they would believe your "arguments"???

More likely those that so desperately keep denying them.

[Milos]

Or it could be the opposite. The output of the NN and MCTS obviously allows a far more accurate time manager than what can be done in an alphabeta searcher, and they did not want to pollute their results with another source of superiority.
And the same can be said about the opening book. Given the calculus power of Google, they probably can compute an opening book far superior than that of Cerebellum/BrainFish. And using it will also just pollute their results.
They just wanted to demonstrate that mcts + NN on appropriate hardware is better than alphabeta + traditionnal eval with the latest release (yes SF8 IS the latest release, the nightly build of a development version is not a release) of one the strongest chess engine on a strong current hardware.

What DeepMind ppl wanted or not wanted with A0 has absolutely nothing to do with why they crippled SF.
1min/move is idiotic TC in chess, period. DeepMind ppl also know that very well. No alpha-beta chess engine would work very well with that TC. They also know that. Still they have deliberately chosen to play at that TC. That in scientific world is called deliberate bias in data measurement or even cherry picking.
The same issue is with hash. Opening book is a part of the engine, even more for Cerebellum/SF. The EGTBs too. The fact that one doesn't incorporate them into one single executable doesn't mean they are not part of the package called engine today.
There are standards how engines are tested. These standards are well established and quite known.
The way DeepMind tested A0 vs SF deliberately ignores all those standards. So result of A0 by itself is important but if the goal was, as you said, to "demonstrate that mcts + NN on appropriate hardware is better than alphabeta + traditionnal eval", that goal hasn't been met because available facts don't support it.
What is much more probable is that publishing the result of that particular match is nothing but a Google PR for their cloud ML services.
And ppl who believe that Google is doing something for the sake of science and general human well-being and not for satisfying their shareholders have serious disability in understanding of real world around them.

[hgm]

Conspiracies, conspiracies! No matter how many totally unfounded conjectures we have to make, if they support a conspiracy theory the must be true!

The claim that TCEC is ot really an 'engine competition', becasue the engies cannot use the books that are an integral part of them, is also new from you.