Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / AlphaZero beats Stockfish 8 by 64-36
1 2 3 4 5 6 Previous Next  
Parent - - By Venator (Silver) [nl] Date 2017-12-10 07:00
By saying that "correspondence chess is solved" you say that in a VLTC (very long time control) match a next generation NN program will not be able to beat current Stockfish. This is simply hogwash :-).

If next SF is 50 elo stronger than the one you currently use, of course you will take the stronger one to analyse. This is the same as admitting that current SF might be out-analysed by the stronger one.
Parent - - By Kreuzfahrtschiff (***) [de] Date 2017-12-10 10:36
there will be no program able to win against current, YES

50 elo is in blitz games, but that is not relevant
Parent - - By Venator (Silver) [nl] Date 2017-12-10 13:10
A next generation NN program will not only beat SF by a large margin in VLCT matches, but also beat you in a lot of correspondence games as soon as your opponents start to use it and you won't ;-).
Parent - By Chris Whittington (**) [fr] Date 2017-12-10 13:35 Upvotes 2
Dunning-Kruger, peak Mount Stupid. Why bother?
Parent - By Chris Whittington (**) [fr] Date 2017-12-08 17:45 Upvotes 2
Well, the smart ones have already seen everything has changed, and why and how. You not, it appears. Everybody at his own speed. If/when you catch up, maybe you'll be worth a conversation, but for now, you are wrong in too many ways.
Parent - - By Venator (Silver) [nl] Date 2017-12-09 08:13 Upvotes 1
Hi Chris,

Nice to see you back!

100% draw rate because, probably, most of your opponents are using Stockfish or equivalent.

Agreed, everyone uses the same software, thus everyone has the same blind spots. F.e.:

1. The top engines often wrongly evaluate pawn down endings, thinking they have a clear advantage. This could negatively impact the calculation of the main line, where they discard a possible drawing ending in favour of a worse middlegame position. I.e. they wrongly think a +0.35 middlegame position is better than a +0.50 pawn down ending, which is in fact drawn.

2. You often see a lot of 0.00 scores attached to middlegame positions with a lot of pieces on the board. If 0.001% of these 0.00 scores are wrong, this could correspond with a lot of losses nobody is currently aware of.

Of course a next generation NN program will flawlessly take advantage of both situations.
Parent - - By Chris Whittington (**) [fr] Date 2017-12-09 11:09
Hi Jeroen,

Interesting times!

Yes, same blind spots, same drawish sub tree result, all searching similar subset of all known chess positions and missing out the forest of Mr Tal.

I think you are a bit wrong concentrating on these numbers +0.5 and so on. The implication being that more evaluation function tinkering can fix the problem. Evaluation tinkering and adding all the terms IS the problem. It's the old precision/accuracy problem. We can be very precise, but when we add it up, we're not accurate.

In old paradigm terms, thinking material plus positional, one can only get at the winning games in the wilder parts of the forest by overruling the material factor with the positional one (positional > material). Is not doable in old paradigm, because the downside risk is too high, they'll just lose way too many games. Positional is too difficult to work out. Fundamentally, it's the  -(necessarily overriding) "material" concept that is responsible for the fail. Both in evaluation and search decision process.

NN is measuring something different, I am hunting around for the word picture of an integrated evaluation, but I think it might be "Energy".
Parent - By Venator (Silver) [nl] Date 2017-12-09 12:20
I only use the numbers to make it more visible and to compare :-).

I am very interested in what next generation NN programs will learn about chess what we people still do not know. Perhaps they have completely different (and changing) piece values depending on certain patterns. That kind of stuff. Serious chess history is maybe 120-140 years old. I am 100% sure there are lots of things we humans haven't discovered yet.

After the Go match between Alpha Go and Ke Jie I have been following some YouTube articles where a 9 dan professional is analysing Alpha Go Zero games. Very interesting stuff, the program comes up with surprising ideas all the time. The Go world has accepted the new situation and tries to learn from Alpha Go.
Parent - - By Venator (Silver) [nl] Date 2017-12-10 07:08
I also like to see a NN program that gives me the "winning probability" of each move. Same as Alpha Go is doing.

Of course I expect a next generation NN program giving correct winning probabilities f.e. in endings a man down and in blocked positions, two situation the current top programs do not grasp yet, even when searching very deeply.
Parent - By Chris Whittington (**) [fr] Date 2017-12-10 11:13
just as long as you realise the "correct winning probability" is a) subjective and b) nor correct. AlphaZero is just less wrong.
Parent - - By Venator (Silver) [nl] Date 2017-12-08 15:22
I disagree. There are still areas where the top programs can be greatly improved. To name just a few:

* Blocked positions
* Fluid pawn structures keeping pawn levers
* Correct evaluation of pawn down endings

The TCEC 10 superfinal has shown that these three issues are still there. And I wonder how many good defensive moves are rejected because an engine incorrectly evaluates the resulting ending.
Parent - - By Kreuzfahrtschiff (***) [de] Date 2017-12-08 15:43 Edited 2017-12-08 15:46
u may disagree, but its just true, corr chess is another level

i try the positions u describe, but my opps know how to respond properly, so no wins in those areas too, its just highest level
Parent - - By Venator (Silver) [nl] Date 2017-12-08 16:38 Edited 2017-12-08 16:40
i try the positions u describe, but my opps know how to respond properly

Perhaps because the humans interfere and prevent the engine(s) from taking a bad decision?

Alpha Zero has shown that it can produce moves that the current top engines fail to find in a reasonable time. And Alpha Zero is just the first step in this direction. I am 100% sure a further developed Alpha Zero will continue to find ideas and moves that will be impossible to find with the current top engines. Just as it is happening with Alpha Go Zero in the game of Go. There are billions of brilliant moves hidden in the search trees that are now rejected, but which a further developed Alpha Zero will find and play.
Parent - - By Kreuzfahrtschiff (***) [de] Date 2017-12-08 16:55 Edited 2017-12-08 17:10
"reasonable time" is no argument if i have all the time i need

go is much more complicated as chess, what is practically solved, so not much to gain by AZ
Parent - - By Venator (Silver) [nl] Date 2017-12-08 17:30
I disagree again. With Shredder and Fritz people already thought that the limits of chess had been reached. They were completely wrong.

"reasonable time" is no argument if i have all the time i need

It *is* a good argument: you might think a 5 day 24/7 analysis with SF is good enough, but it could turn out that you needed more than 10 days to see through a brilliant new Alpha Zero concept.

Furthermore: humans made the rules and tricks in the current chess engines and those human ideas might just be completely wrong on many occasions.....
Parent - - By Kreuzfahrtschiff (***) [de] Date 2017-12-08 17:45 Edited 2017-12-08 17:49
u overestimate that thing a lot  ;) 

but AZ cannot make a win out of a draw position ( i mean the start position)  :)

and human ideas i dont use anyway since 4-5 years  ;)
Parent - - By Venator (Silver) [nl] Date 2017-12-08 18:11 Edited 2017-12-08 18:14
u overestimate that thing a lot  ;)

On the contrary, you fail to notice that there is a new way of playing chess in the making :-).

I am sure in 10-20 years the NN programs will discover a lot of mistakes in your correspondence games ;-).

and human ideas i dont use anyway since 4-5 years  ;)

SF uses human ideas, so you are playing more "human-like" than you can imagine ;-).
Parent - - By Kreuzfahrtschiff (***) [de] Date 2017-12-08 18:13 Edited 2017-12-08 18:20
not in the last 2 years, except some fun games and one mouse slip  ;)

btw, i analysed some of that AZ games and stockfish played poorly, my machines found it at once
(15 minutes per move at most and i have only 16 cores)
Parent - - By Venator (Silver) [nl] Date 2017-12-09 07:47 Edited 2017-12-09 07:58
Elsewhere you describe that your way of playing correspondence chess is to deeply analyse one line that has a positive score and discard the 0.00 lines. This obviously has two flaws:

1. SF dismisses a lot of pawn down endings that are actually drawn, thinking they are clearly worse than f.e. +0.35 middlegame positions => thus it might "think" it has an advantage while in fact there is none;
2. SF says a lot of positions are 0.00, even very crowded middle game positions with a lot of pieces on the board => many of these 0.00 evaluations can be wrong.

Of course in both situations a next generation Alpha Zero will flawlessly point out these mistakes and take advantage of them. In fact, if only 0.001% of the 0.00 SF evaluations are wrong, this can correspond with a lot of losses that currently nobody is aware of. Everybody uses the same program to analyse positions, thus everybody has exactly the same blind spots.
Parent - - By Kreuzfahrtschiff (***) [de] Date 2017-12-09 11:27
if u run at depth 62 u dont miss such a thing  :)
Parent - By Venator (Silver) [nl] Date 2017-12-09 11:56
You don't get the point :-).
Parent - - By Venator (Silver) [nl] Date 2017-12-10 07:09
Let me ask you a simple question:

If in 10 years there is a NN program that is 300 elo stronger than the Stockfish you currently use in your correspondence chess games, will you use it?
Parent - - By Kreuzfahrtschiff (***) [de] Date 2017-12-10 10:32
there will be no program 300 points better ever, but i will use the strongest program, because its saving time
Parent - - By Venator (Silver) [nl] Date 2017-12-10 13:07
there will be no program 300 points better ever

Of course there will be. Even the traditional programs add 40-50 elo per year.

i will use the strongest program, because its saving time

You will use it as you would otherwise risk a lot of losses :-).
Parent - By Kreuzfahrtschiff (***) [de] Date 2017-12-10 13:47
there is a limit in elo progress and its already there in corr chess, so, ...

ask everyone who plays for the world championship in corr chess
Parent - - By oudheusa (*****) [nl] Date 2017-12-09 10:40
No strong corrr player would have found the Nh6 sacrifice.
Parent - - By Chris Whittington (**) [fr] Date 2017-12-09 11:12
there are no "strong correspondence players", the correct term is Stockfish operator with computer
Parent - - By Thomas A. Anderson (*) Date 2017-12-17 01:58
I agree in almost everything you are posting here. Thanks for your educational stamina. While you might get upset by some post that are related to correspondance chess, don't be fooled into saying such generalizations taring all CC people with the same brush.
Parent - By Chris Whittington (**) [fr] Date 2017-12-17 10:44 Upvotes 1
yeah, I agree my evidence is a bit anecdotal ;-)
it's just that the people who claim to be correspondence champions and so on, on these forums, do also make other posts and generally appear (to my sensitive neurology) to fairly stupid and unable to think at a good level. So I doubt their chess is more than operator plus engine. Obviously this is anecdotal and not a general rule.
Parent - - By Kreuzfahrtschiff (***) [de] Date 2017-12-09 11:26 Edited 2017-12-09 11:57
sure we would have seen it, but its only a draw

i am no romantic, i am scientific
Parent - - By Kappatoo (*****) [de] Date 2017-12-09 11:29
Could you give some analysis? I'm curious about that game.
Parent - - By Kappatoo (*****) [de] Date 2017-12-09 12:32
Yes, this one. What was the draw you were talking about?
Parent - - By Kreuzfahrtschiff (***) [de] Date 2017-12-09 12:48 Edited 2017-12-09 13:05
better as 27... Bg6? was

27... Bxe4 28. Qxe4 Kg8 29. Qe6+ Rf7 30. Qg4 Nd7 31. Rxd7 Qxc3 32. Rc1 Qf6 33. Bd4 Qg6 34. Rc7 Re8 35. R1xc6 Qf5 and draw

and if 29. Bd4 Rf7 30. Qg4 Nd7 31. Qxd7 Rd8
Parent - - By Kappatoo (*****) [de] Date 2017-12-09 13:05
Thanks!
Parent - - By Kreuzfahrtschiff (***) [de] Date 2017-12-09 13:10
and this i found only after looking at the game for 1 hour
so what chances has AZ against a 2 days per move corr player?
Parent - - By Kappatoo (*****) [de] Date 2017-12-09 13:23 Upvotes 2
Thanks again for the analysis.
I don't really understand your reasoning here, though.
First of all, this program was the result of just a few hours of training, starting from zero knowledge. I know that they used a lot of hardware in training the neural network, and I know that the conditions weren't ideal for Stockfish. But for the life of me I cannot understand how one can fail to be impressed that they surpassed, or at least levelled, decades of research in a few hours.
Secondly, what do you think they could achieve if this was more than just a side project? Again, this was the result of a few hours of training.
Thirdly, the programs played at a rate of 1 minute per move. It's not a big surprise that one can improve their play by taking an hour, or even two days for a move. The more relevant questions are: Can you match what Alpha does in a minute? And would you be able to hold yourself in a correspondence match, if DeepMind set all their resources to it?
Parent - - By Kreuzfahrtschiff (***) [de] Date 2017-12-09 13:58
as i said before, corr chess is solved, so AZ can not win
Parent - - By Kappatoo (*****) [de] Date 2017-12-09 14:02
You still play correspondence chess, right? Can I ask why?
Parent - By Kreuzfahrtschiff (***) [de] Date 2017-12-09 14:09
because i want the ficgs title

if thats done, i will only defend it  :)  with draws
Parent - - By zaarcis (***) [lv] Date 2017-12-09 14:24
And thats all? Did you even try other options?

I would be kind of ashamed to be satisfied with one or two lines while there's so much to explore. (What I'm doing now as it's interesting.)
Parent - - By Kreuzfahrtschiff (***) [de] Date 2017-12-09 14:44
if i see a draw variation i dont search for 10 of them, thats a waste of my time
Parent - - By zaarcis (***) [lv] Date 2017-12-09 14:50
Sure it is, it's not my game anyway. :)
Parent - By Kreuzfahrtschiff (***) [de] Date 2017-12-09 14:55
i prove a draw and will only discuss this furthermore if someone could refuse it, thats called effective analysis
Parent - By glennsamuel32 (*) Date 2017-12-17 04:44

>No strong corrr player would have found the Nh6 sacrifice.


Actually, the top 3 engines find it instantly but as the second PV, with scores from -0.10 to -0.50 after a minute...
This is an instance of far superior hardware searching very deep...
Parent - - By turbojuice1122 (Gold) [us] Date 2017-12-10 03:51

> its still nearly 100% draw rate if i play corr chess against top players


> no program can change that


I think one could quite safely say that this program would destroy the top corr players in the world.  Such a program would greatly benefit from the time scaling of long games, something from which modern computer programs benefit only fractionally.  You can't make a prediction based on the results of play against traditional computer opponents when these results show that the traditional computer programming techniques are dead.
Parent - - By Kreuzfahrtschiff (***) [de] Date 2017-12-10 10:37
i can and i am right, every serious corr player would agree
Parent - By Venator (Silver) [nl] Date 2017-12-10 13:13
No they won't :-). They will all use the much stronger NN program, as it will easily beat the old and traditional program.
Parent - - By rocket (***) [se] Date 2017-12-08 13:25
The very first self-play game must be played random since it has no past experience of other games. And you guys just said that the same position cropping up twice is unlikely, so how then is this data put to use with positions that at best only resemble other past positions?
Parent - - By Chris Whittington (**) [fr] Date 2017-12-08 13:55 Upvotes 1
Yes, correct, first game is random moves, most of the earlier games will seem very random. Do you remember learning to walk? You got there in the end right? With each blundering step, step by step, your motor neurons and sensory neurons gradually organize them selves.
One thing to note, as a network trains, the effect on the finished network of the early learning is diminished, it's the later learning cycles that dominate. So, it doesn't really matter that early training was on dumb data, all that matters is that the early learning generated some progress to gradually improve on. And, obviously in this case A0, it did.
Up Topic The Rybka Lounge / Computer Chess / AlphaZero beats Stockfish 8 by 64-36
1 2 3 4 5 6 Previous Next  

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill