AlphaZero vs Stockfish

[Milos]

You don't seem to actually understand anything about that position.
It's weird that ppl are not capable even to do descent analysis of the game.
The sacrifice was not sound and Alpha0 won the game just because SF made couple of blunders in a row, first 21...Bf5 instead of 21...Kg7 and then even more crucial 22. ...Qa3 instead of 22...Qe6 and then Qf7.

22. .... Qe6 could be answered by Bxg5+ with

Analysis by asmFishW_2017-11-19_popcnt:

+- (21.36 ++) Depth: 34/62 00:02:17 3697MN, tb=160567


You probably mixed up some lines.

I meant optimal continuation of the actual game. So the line is:
20. h4 f6 21. Be3 Kg7
22. Rad1 Qe6 23. hxg5 Qf7
[d]rnb2r2/p3bqkp/1pp2p2/6P1/Q7/2P1B1P1/P4PB1/3RR1K1 w - - 1 24
That is a clear draw, and white can do nothing about it.

[Milos]

The sacrifice was not sound and Alpha0 won the game just because SF made couple of blunders in a row, first 21...Bf5 instead of 21...Kg7 and then even more crucial 22. ...Qa3 instead of 22...Qe6 and then Qf7.

WOW ! Now I've seen everything.
Milos has discovered that both AlphaZero and Stockfish made mistakes !!
I think AlphaZero played against the wrong opponent; it should have been Milos.
TOOOO Good !

This is the last time I answer to your retarded posts.
Stop trolling. You never ever have anything useful to say, and recently not only you are trolling but also going around and insulting ppl.
I am really clueless why mods allow it, but anyway in addition to ignoring that written diarrhoea, I will also report every post of yours that contains insults, just as a test to see when will moderation start doing something about it.

[IQ]

I meant optimal continuation of the actual game. So the line is:
20. h4 f6 21. Be3 Kg7
22. Rad1 Qe6 23. hxg5 Qf7
[d]rnb2r2/p3bqkp/1pp2p2/6P1/Q7/2P1B1P1/P4PB1/3RR1K1 w - - 1 24
That is a clear draw, and white can do nothing about it.

Are you sure? Did you look at 24 Qf4!? in that line.... the position looks far from clear - many many interesting lines. Also 23. Bd4 comes to mind.

[Werewolf]

Milos I have a question for you.

I've only just realised how great the hardware difference was. I knew it was big, but I didn't realise quite how big.

If Stockfish was:

- upgraded to the latest asmfish
- given 64GB Hash
- A well prepared opening book
- tablebases
- a faster dual xeon machine with HT off
- match played under a more flexible time control (say 90 min + 30 sec)

How much of the 130 elo difference could be clawed back?

[Jouni]

I think SF was handicapped at least by 100 ELO making whole experiment almost useless. How can very intelligent people at DeepMind do that kind of error? I admit they made impressive things with learning, but this test vs SF don't show that clearly. Even Komodo or Houdini will beat SF with these conditions!

[Milos]

Are you sure? Did you look at 24 Qf4!? in that line.... the position looks far from clear - many many interesting lines. Also 23. Bd4 comes to mind.

Can you show the winning line?
All I see is draw, even after depth 45 analysis of SF.

[Milos]

Milos I have a question for you.

I've only just realised how great the hardware difference was. I knew it was big, but I didn't realise quite how big.

If Stockfish was:

- upgraded to the latest asmfish
- given 64GB Hash
- A well prepared opening book
- tablebases
- a faster dual xeon machine with HT off
- match played under a more flexible time control (say 90 min + 30 sec)

How much of the 130 elo difference could be clawed back?

It depends mostly on the book used by SF, i.e. the difference could be reduced for almost 100 Elo (which was the initial advantage of A0 over SF after the match of 100 games).
With Cerebelum book, and conditions you suggested I believe they would be on par or almost on par.

[IQ]

Are you sure? Did you look at 24 Qf4!? in that line.... the position looks far from clear - many many interesting lines. Also 23. Bd4 comes to mind.

Can you show the winning line?
All I see is draw, even after depth 45 analysis of SF.

The problem with this position is that SF sees lots of draw variations but there is a lot of play left. This is the first position i cannot resolve either way, whenever I think I analysed it to a draw i find another resource which keeps the game going. And who knows how many of those lines contain a subtle Zugzwang that SF can't see. For example:

24. Qf4!? (my idea, was not among the first 4 choices by SF)
24. ... fxg5
25. Bd4+ Kh6
26. Qe5 Bf6
27. Qd6 Bg4 (27... Bf5 28. f3 g4 Bg6 29. f4 +-)
28. f4
(f3 Bf5 g4 is also interesting, and SF gets fooled here for a while with a white advantage, but i think it's draw)
28. ... Nd7 (28... Rd8!? with possible queen sac by alpha fxg5+ Kh5 etc.)
29. fxg5+
suddenly a lot of possibilities for white, SF scores everything as draw, but if you follow the lines you see that nothing really gets resolved - even with reduced material white keeps pressuring blacks king for example:
30. .... Kh5
31. Be3

or:

29. Bxc6


And there is also the whole Bd4+ complex of positions where white leaves his queen on the queenside....

[Werewolf]

Milos I have a question for you.

I've only just realised how great the hardware difference was. I knew it was big, but I didn't realise quite how big.

If Stockfish was:

- upgraded to the latest asmfish
- given 64GB Hash
- A well prepared opening book
- tablebases
- a faster dual xeon machine with HT off
- match played under a more flexible time control (say 90 min + 30 sec)

How much of the 130 elo difference could be clawed back?

It depends mostly on the book used by SF, i.e. the difference could be reduced for almost 100 Elo (which was the initial advantage of A0 over SF after the match of 100 games).
With Cerebelum book, and conditions you suggested I believe they would be on par or almost on par.

So with a fraction of the processing power Stockfish is almost as good as Alpha Zero. Interesting...

[hgm]

It is rather funny to see how the notion had Stockfish would be better if it was using a book seems to prevale. Imagine how strong a Chess player I would be, if I could make Kasparov do the moves for me...