Alphazero news

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.
jhellis3
Posts: 382
Joined: Fri Aug 16, 2013 10:36 pm

Re: Alphazero news

Post by jhellis3 » Sun Dec 09, 2018 8:21 pm

Hand-crafting an eval function, what you really get form a project like Stockfish is the understanding of how to write code to make sense of all the aspects putting together in a way that is statistically sound in practice..
I agree in principle, that if one wants to learn how a human might attempt to code a quality chess engine SF is a wonderful example. Unfortunately, from an objective strength point of view the effort is a rather futile one. Or rather if one genuinely cares about the best result possible, one would not bother coding an eval in the first place.

But if one is satisfied with inferior human coded evals, certainly anyone is free to use them as they see fit.
Last edited by jhellis3 on Sun Dec 09, 2018 8:28 pm, edited 1 time in total.

noobpwnftw
Posts: 333
Joined: Sun Nov 08, 2015 10:10 pm

Re: Alphazero news

Post by noobpwnftw » Sun Dec 09, 2018 8:25 pm

jhellis3 wrote:
Sun Dec 09, 2018 8:21 pm
By the end of the day, what you get out of it is a complete and mathematically sound solution to the problem, once and for all.
Well, I don't know about the once and for all part or SF would never significantly change (and it has). But I agree in principle, that if one wants to learn how a human might attempt to code a quality chess engine SF is a wonderful example. Unfortunately, from an objective strength point of view the effort is a rather futile one. Or rather if one genuinely cares about the best result possible, one would not bother coding an eval in the first place.

But if one is satisfied with inferior human coded evals, certainly anyone is free to use them as they see fit.
By that I mean the tablebases, SF is in another paragraph.
Last edited by noobpwnftw on Sun Dec 09, 2018 8:26 pm, edited 1 time in total.

Albert Silver
Posts: 2829
Joined: Wed Mar 08, 2006 8:57 pm
Location: Rio de Janeiro, Brazil

Re: Alphazero news

Post by Albert Silver » Sun Dec 09, 2018 8:29 pm

noobpwnftw wrote:
Sun Dec 09, 2018 8:20 pm
Albert Silver wrote:
Sun Dec 09, 2018 8:14 pm
it has developed on a concept of winning chances
Concept of winning chances is just as lame as concept of centipawns.

Given a mate-in-X position if it says anything else than 100% then it is wrong, if it says 100% then it will screw up your search.

It have its roots in image recognition NNs, which I also find hilarious, given a picture of a man, it might say "oh that looks 60% like a man, 30% like a dog, 5% like a car and 5% like a tree", very informative, isn't it? :D
You are missing the point. The development on a basis of winning chances has led to a very different basis for making decisions with far-reaching consequences.
"Tactics are the bricks and sticks that make up a game, but positional play is the architectural blueprint."

noobpwnftw
Posts: 333
Joined: Sun Nov 08, 2015 10:10 pm

Re: Alphazero news

Post by noobpwnftw » Sun Dec 09, 2018 8:33 pm

Albert Silver wrote:
Sun Dec 09, 2018 8:29 pm
You are missing the point. The development on a basis of winning chances has led to a very different basis for making decisions with far-reaching consequences.
If I understand your concept of "very different basis for making decisions with far-reaching consequences" correctly, are you referring to the most widely used method by opening books and chess databases?

jhellis3
Posts: 382
Joined: Fri Aug 16, 2013 10:36 pm

Re: Alphazero news

Post by jhellis3 » Sun Dec 09, 2018 8:33 pm

Approaches like A0, given a black-box of 28 million floating numbers, which is neither perfect nor give much information on how it would come up with such evaluations, is more of a "yet another closed-source chess engine" scenario in practice, where you have to "understand" things by looking at it's output, seems to me neither efficient nor intelligent.
Well, I hate to break it to you, but that is precisely the use case for most strong players: load the engine, get its opinion on some moves, and explore from there. While I am sure there are a handful of GMs who bother to look at the source of SF, they are in the vast minority, and there really is not anything for them to learn there that they do not already know (SF eval is quite basic). The knowledge real world users are seeking comes from the very deep (brute force) search that SF and other engines perform on a position.

noobpwnftw
Posts: 333
Joined: Sun Nov 08, 2015 10:10 pm

Re: Alphazero news

Post by noobpwnftw » Sun Dec 09, 2018 8:38 pm

jhellis3 wrote:
Sun Dec 09, 2018 8:33 pm
Approaches like A0, given a black-box of 28 million floating numbers, which is neither perfect nor give much information on how it would come up with such evaluations, is more of a "yet another closed-source chess engine" scenario in practice, where you have to "understand" things by looking at it's output, seems to me neither efficient nor intelligent.
Well, I hate to break it to you, but that is precisely the use case for most strong players: load the engine, get its opinion on some moves, and explore from there. While I am sure there are a handful of GMs who bother to look at the source of SF, they are in the vast minority, and there really is not anything for them to learn there that they do not already know (SF eval is quite basic). The knowledge real world users are seeking comes from the very deep (brute force) search that SF and other engines perform.
That is true, my point is that you can secretly switch their SF with "yet another chess program" and they may never notice it. Any 3000+ engine provides more than enough information for them already. We shall seek for higher objectives. It would be a scientific disaster if papers from the future look like lists of ingredients to conjure NNs.

duncan
Posts: 9990
Joined: Mon Jul 07, 2008 8:50 pm

Re: Alphazero news

Post by duncan » Sun Dec 09, 2018 9:01 pm

Albert Silver wrote:
Sun Dec 09, 2018 8:14 pm
duncan wrote:
Sun Dec 09, 2018 7:10 pm
Albert Silver wrote:
Sun Dec 09, 2018 4:51 pm
duncan wrote:
Sun Dec 09, 2018 4:40 pm
Albert Silver wrote:
Sun Dec 09, 2018 4:34 pm


Yet, that is what AlphaZero has done, and we are even able to bring this to the home user's PC thanks to Deep Mind's generosity with their knowledge, as well as the fantatsic Leela Chess community efforts. In other words that is not limited to some absurdly exotic hardware no one could ever hope to obtain. I am not even commenting on the whole self-learning process, which is what has been the focus.

What is more, to achieve this, you are looking at an incredibly evolved eval function (not precisely, but it helps illustrate the point) that has roughly 28 million values compared to a few thousand at most for even the most sophisticated predecessors. In all the years I have seen discussions on the fight between smart searchers and fast searchers, I have never seen anyone come close to imagining that is how enormous a difference it would take, much less realize and prove it.

It is pure genius.
Playing devil's advocate. Yes it is genius, but the chess community is not looking for genius we are looking to revolutionise chess playing and alpha zero seems to have maxed out at a level not significantly higher than an old version of stockfish without a book.

A dazzling 'toy', but a 'toy' nevertheless.

The future may be different, but that remains to be seen.
Needless to say, I disagree with pretty much every point you made. We will have to agree to disagree.
Needless to say, you are far more likely to be right than I am, as I know nothing about computer chess, but please could you point out any factual errors in my post.

good article here.
https://en.chessbase.com/post/the-full- ... -long-last
When you say 'toy', it is a comment laced with prejudice, and also suggests this is not a serious thing. If a program can fight with the very best, I find that hard to call a 'toy'. Look at plain Leela right now. It is right up there and has been in development for less than a year. That is nothing short of incredible, and improvements have been coming along all throughout.

Then there is the plain reality that it is using a design no one had ever done before, able to fight with the best even when running a thousand times slower. More than that, it finds moves and ideas that are completely different from Stockfish and Co. It also misses moves that they find, so this is not more of the same, but something new you can complement your analysis and chess with. It can play longterm sacrifices naturally, with no need of artificial levers, as it has developed on a concept of winning chances, and has no real understanding in centipawns. The one Lc0 puts out is a weird translation, but you can see the true winning % if you activate the UCI option.

For me, all this is revolutionary.
Agreed I should not have used the word toy, on a serious top ranking engine.

Albert Silver
Posts: 2829
Joined: Wed Mar 08, 2006 8:57 pm
Location: Rio de Janeiro, Brazil

Re: Alphazero news

Post by Albert Silver » Sun Dec 09, 2018 9:03 pm

noobpwnftw wrote:
Sun Dec 09, 2018 8:33 pm
Albert Silver wrote:
Sun Dec 09, 2018 8:29 pm
You are missing the point. The development on a basis of winning chances has led to a very different basis for making decisions with far-reaching consequences.
If I understand your concept of "very different basis for making decisions with far-reaching consequences" correctly, are you referring to the most widely used method by opening books and chess databases?
I am referring to the engine, not the opening book or chess databases.
"Tactics are the bricks and sticks that make up a game, but positional play is the architectural blueprint."

noobpwnftw
Posts: 333
Joined: Sun Nov 08, 2015 10:10 pm

Re: Alphazero news

Post by noobpwnftw » Sun Dec 09, 2018 9:17 pm

Albert Silver wrote:
Sun Dec 09, 2018 9:03 pm
noobpwnftw wrote:
Sun Dec 09, 2018 8:33 pm
Albert Silver wrote:
Sun Dec 09, 2018 8:29 pm
You are missing the point. The development on a basis of winning chances has led to a very different basis for making decisions with far-reaching consequences.
If I understand your concept of "very different basis for making decisions with far-reaching consequences" correctly, are you referring to the most widely used method by opening books and chess databases?
I am referring to the engine, not the opening book or chess databases.
But the fact that using winning chance representation is not new and have been found functionally equivalent to centipawn representation. It is a compromise because the evaluation is not perfect. Switch from centipawn to "centiking" does not remove that compromise in any way.

People happen to use it because usually NNs have an output range of [-1, 1], and converting it linearly gets you the "winning chance".

Sesse
Posts: 180
Joined: Mon Apr 30, 2018 9:51 pm
Contact:

Re: Alphazero news

Post by Sesse » Sun Dec 09, 2018 9:52 pm

corres wrote:
Sun Dec 09, 2018 7:07 am
For reverse engineering it need to make working the studied object.
No, this is incorrect.

Post Reply