Stockfish dethroned

macer75
hairhorn wrote:
This is pretty wild. It's far worse at playing Black, though, only 3 wins out of 50 games, the rest of them draws. As White it manages 25/25 win/draw.

Or maybe it's just because there's an inherent disadvantage to playing as black?

hairhorn
There may be, but it can't be nearly that large.
Nilmp

Remarkable stuff. Some things to note:

- the underlying search algorithm is different to existing engines, so it’s not meaningful to compare search depth. Alpha Zero seems based on Monte Carlo tree search which is probabilistic in nature and more like a shoot-out concept

- ratings are hard to assess until a collection of programs arise which can beat Alpha Zero some of the time. It would be like trying to rate StockFish by only playing humans. 

- we can’t yet assess the relevance of time controls, or extra hash size, latest SF etc. Chess players are keen to find all these things out but remember this is an AI project first and foremost so we’ll have to wait a little bit to find out

- I built an engine and was always bored by how specific it had to be to get good results. For example, if the depth is greater than 7 and this is the 9th move to search and blah blah then reduce the depth by 2, unless it’s being attacked by a pawn on the 6th tank blah blah. I also hate the quiescent search concept, just feels wrong. This new approach is exciting because it feels open ended and more “pure”.

- is the Queens Indian busted now? happy.png

wrimle

The engine is learning to play by playing against itself. As white has the first move advantage it is natural that it evolves a style where it tries to win with white, and tries to avoid losing with black. This may be the reason why it wins less with black. If it ways trained playing against a weaker white opponent it might have evolved a more aggressive style with black as well.

sammy_boi
hairhorn wrote:
There may be, but it can't be nearly that large.

Argument from personal incredulity heh.

Yes, white has a small advantage, and advantages are more pronounced the higher the level of play.

nighteyes1234
chiefonion wrote:

Yea, what were the time controls? It looks like Stockfish made some silly trades.

 Stockfish didnt want anything to do with a lot of the losing moves. Either they were flat out blunders it was forced to play or in the drawn positions it was forced to try for a win or something. There was one where Alpha Zero was seemingly forced to go for a win, but then Stockfish continued to make those forced blunders.

 

Still, there was the French defense game where Alpha Zero as white won clearly. Stockfish cant see the Rc1-Qe3-Qc3 move.




 

 

hairhorn
sammy_boi wrote:
hairhorn wrote:
There may be, but it can't be nearly that large.

Argument from personal incredulity heh.

Yes, white has a small advantage, and advantages are more pronounced the higher the level of play.

Argument by error. In the limit, the disadvantage of playing Black should be reflected precisely by the game statistics, not "amplified" by them.

If White had anywhere near a 25-3 advantage over Black at the highest levels of play, that would have been noticed decades, if not centuries, ago.

Colin20G

This game was supposed to keep mankind busy for a couple more thousands years. Instead, it might be dried up in a matter of decades. Why is it so great? I don't know.

sammy_boi
Colin20G wrote:

This game was supposed to keep mankind busy for a couple more thousands years. Instead, it might be dried up in a matter of decades. Why is it so great? I don't know.

Spoken like a beginner who thinks memorizing 10 moves in just the right opening will guarantee victory.

This changes nothing for human chess.

hairhorn

Indeed, if billion-dollar companies are spending their time on chess engines, it makes chess look more relevant, not less.   

Debistro
Colin20G wrote:

This game was supposed to keep mankind busy for a couple more thousands years. Instead, it might be dried up in a matter of decades. Why is it so great? I don't know.

Uhh, not decades. This AI will solve chess by the looks of it.

sammy_boi
Debistro wrote:
Colin20G wrote:

This game was supposed to keep mankind busy for a couple more thousands years. Instead, it might be dried up in a matter of decades. Why is it so great? I don't know.

Uhh, not decades. This AI will solve chess by the looks of it.

Spoken like someone who doesn't know how big 10^40 is.

You'd need a storage device larger than the earth itself. Good luck with that.

Debistro
sammy_boi wrote:
Debistro wrote:
Colin20G wrote:

This game was supposed to keep mankind busy for a couple more thousands years. Instead, it might be dried up in a matter of decades. Why is it so great? I don't know.

Uhh, not decades. This AI will solve chess by the looks of it.

Spoken like someone who doesn't know how big 10^40 is.

You'd need a storage device larger than the earth itself. Good luck with that.

And you are still thinking old school. Ever heard of nano technology? You are still thinking based on the old way of approaching chess AI...

But when I say solve chess, it means refuting openings with certainty.

Depends on many openings it can refute.

Debistro
chiefonion wrote:

So, how do we apply this technology to real life? If we take one of those Japanese robots that seems eerily real, add AlphaZero to it, we can have a partner that learns to treat us better. We go to work, come home, and it has calculated how to be a better wife or husband. I like this idea.

 

 

 

 

You get the idea....Tongue Out

Lyudmil_Tsvetkov

I am reposting my message to Debistro from the 'Stockfish is blind' thread, as this is all SCAM.

 

And you are awfully wrong.

Alpha will not get from the ground. It is somewhere a 2850 engine, so 97th best among chess engines on a single core.

If anything was a scam, it is this, because the hardware advantage of Alpha over Stockfish was 16/1, 4 TPUs make around 1024 standard cores, while Stockfish has been using only 64.

How fair is this?

On equal hardware, Stockfish is 300-400 elo stronger.

Besides, the real gardware advantage might have been even bigger, 30/1 or so, as TPUs are extremely efficient for chess calculations, while with many cores there are diminishing returns in SMP implementation.

It was all hype on the part of Google to sell their products after that.

And many people bought into it.

As I have investigated computer chess and computer chess evaluation extremely closely, I know 100% that the approach of Alpha simply will not work, maybe to a certain point, but not after that.

Go and chess have almost nothing in common in terms of evaluation, with chess being maybe 1000 times more complex. So Alpha will never solve this puzzle.

 

Currently, they are tuning at the level of 2850, where improvements are very easy to find; later, they will have tremendous difficulties with their approach.

In order to get to much stronger level than current top engines in chess, one must first get rid of all inherent redundancies and come up with new more sophisticated terms, which they have not done.

 

So that, basically, it was all due to the 30/1 hardware advantage(plus the simulated opening book only Alpha had, with most games I replayed being decided in the early opening).

Lyudmil_Tsvetkov

Again, Alpha is just a 2850 engine, and only the 30/1 hardware advantage decided in its favour.

Still 400 elos below Stockfish.

me_roma
chiefonion wrote:

So, how do we apply this technology to real life? If we take one of those Japanese robots that seems eerily real, add AlphaZero to it, we can have a partner that learns to treat us better. We go to work, come home, and it has calculated how to be a better wife or husband. I like this idea.

 

 

 

 

She also goes to work, comes home and since she is way too smarter than you, finds out a way to make you do the chores and dishes XD #PerfectAlphaZeroWife

Lyudmil_Tsvetkov

This project will never succeed as is essentially flawed theoretically.

They will never quite get to the level of Stockfish on single core.

Chess will not be solved 100 years from now.

There are no magical solutions to chess, only very hard work can help find scientific discoveries and scientific discoveries in turn can improve the level of play.

Alpha, on the other hand, wnated to piut all the burden on the hardware, that simply does not work, what kind of an achievement is that?

And yes, Alpha had simulated book, that is why it was able to fianchetto its bishop king side in many games, while Stockfish never did that.

Where hardware does not work, you see the level of their evaluation achievement: 1.d4.

Elroch
Lyudmil_Tsvetkov wrote:

Again, Alpha is just a 2850 engine, and only the 30/1 hardware advantage decided in its favour.

Still 400 elos below Stockfish.

A TPU is indeed about 30 times faster than a CPU for neural network code. However, the strongest chess player is not necessarily the strongest one on specific limited hardware. 

The graphs of Elo against thinking time suggest Stockfish would not ever gain 100 more points with CPU speed. I would have expected it to theoretically approach perfection, so this is a big surprise to me. It is just about possible that the hash table being limited to 1 Gbyte limits the upside. What else could it be?

Lyudmil_Tsvetkov

What graphs?

The match was played at 1 minute fixed time per move.

Stockfish approaching perfection? SF is currently at around 3200 or so, and the perfect player would be at least 5000 elo, maybe much more, this has been discussed frequently on Talkchess.