stockfish 10 vs. Mephisto III S Glasgow

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Dann Corbit, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
Post Reply
jp
Posts: 1442
Joined: Mon Apr 23, 2018 5:54 am

Re: stockfish 10 vs. Mephisto III S Glasgow

Post by jp » Mon Apr 08, 2019 9:51 am

mclane wrote:
Mon Apr 08, 2019 8:17 am
Better programming artificial intelligence then artificial stupidity.
Most of the progress in computerchess comes from hardware progress.
There's been very little progress in AI (minus the hardware progress you don't like).
And the AI field split from chess many decades ago.

User avatar
mclane
Posts: 18388
Joined: Thu Mar 09, 2006 5:40 pm
Location: US of Europe, germany
Full name: Thorsten Czub
Contact:

Re: stockfish 10 vs. Mephisto III S Glasgow

Post by mclane » Tue Dec 03, 2019 7:39 pm

The failure of todays chess programs is that they heavily rely on search tree instead of knowledge or intuition.




Houdini 6.03:
27.Ld1 a3 28.La1 h5 29.Sf1 Sxc4 30.Txg6 Kxg6 31.Dxh5+ Kg7 32.Sg3 (…)
= (0.00) depth: 28/54 00:00:47 743MN, tb=10

Komodo 12.3:
27.Dg4 Tg8 28.h5 gxh5 29.Df5+ Kh8 30.Dxh5 Txg3 31.Dxh6+ Kg8 32.fxg3 (…)
= (0.06 ++) depth: 30 00:00:30 320MN, tb=3

Stockfish 10:
27.Ld1 h5 28.Sf1 De7 29.Se3 De8 30.Sf5 a3 31.Lc3 bxc3 32.Sd6 (…)
= (0.19 ++) depth: 31/50 00:00:31 302MN, tb=18

Ethereal 11.25:
27.Ld1 a3 28.La1 h5 29.De3 Dc5 30.Lxh5 Dxe3 31.Lxg6+ Kh6 32.fxe3 (…)
= (0.00) depth: 26/55 00:00:42 468MN, tb=1


These chess engines we all know search 26-28 searches deep, with extended search up to 55 plies.
But since they do not play chess or have a clue what chess is all about, they
Evaluate the position as drawish.

They do compute many many thousands of NPS in this time.
If each node they compute is grain of sand, how big would the pile of sand be these programs generate.

And all this sand is not helping to understand the main conclusion.




This is really not AI but instead artificial stupidity.

But, we have an engine that is different:

Leela Chess Zero 21.1 (41800):
27.h5 g5 28.Df5+ Kh8 29.f4 a3 30.La1 De7 31.Ld1 exf4 32.Tb3
+/- (1.41) depth: 15/45 00:00:31 375kN
-------------------------------------------------------------------------------------------


And we have another engine that is Artificial Intelligence made from 1983 (computerchess from 36 years ago) :

Mephisto III, 6,1 Mhz on small dedicated chess computer. It computes only 1-3 Nodes per second.


After 50 Sekunden it changes best move to 27.h5!
After 15 Minutes the Score is +0,90 , search depth is 4/14 .

Consider that mephisto III is not running on a pc with thousands of MHz, not using Nvidia card for 1000 dollar.

It runs on a 8 bit fossil 1802 CMOS cpu with only 32 KB ROM and 4 KB RAM.
https://en.wikipedia.org/wiki/RCA_1802

Relate this hardware to the PC you use with Houdini or komodo or Stockfish and the ram and HDD space you have on these machines.

This shows IMO that the way these top programs emulate chess is wrong.

All these huge amount of resources are senseless.
What seems like a fairy tale today may be reality tomorrow.
Here we have a fairy tale of the day after tomorrow....

Jouni
Posts: 2387
Joined: Wed Mar 08, 2006 7:15 pm

Re: stockfish 10 vs. Mephisto III S Glasgow

Post by Jouni » Tue Dec 03, 2019 9:03 pm

Most engines I try play here 27.h4 instantly and it seems to be draw :)

Analysis by Stockfish 211119 64 BMI2:

1.h5 g5 2.Qf5+ Kh8 3.f4 a3 4.Ba1 Qe7 5.Bd1 exf4 6.Rb3 Ra6 7.Qg6 Qf8 8.c5 Ne5 9.Bxe5 fxe5 10.c6 Bxc6 11.dxc6 Nc4 12.Nxc4 Raxc6 13.Nd6 Rxd6 14.Rxc8 Rxd1+ 15.Kh2 Qxc8 16.Qxh6+ Kg8 17.Qxg5+ Kf8 18.Rxb4 a2 19.Qf6+ Kg8 20.Qg6+ Kf8
= (0.00) Depth: 31/55 00:00:32 187mN, tb=964
Jouni

syzygy
Posts: 4975
Joined: Tue Feb 28, 2012 10:56 pm

Re: stockfish 10 vs. Mephisto III S Glasgow

Post by syzygy » Tue Dec 03, 2019 10:48 pm

mclane wrote:
Tue Dec 03, 2019 7:39 pm
This is really not AI but instead artificial stupidity.
The progress you've made in about 8 months in this thread disproves the existence of HI.

User avatar
Ovyron
Posts: 4410
Joined: Tue Jul 03, 2007 2:30 am

Re: stockfish 10 vs. Mephisto III S Glasgow

Post by Ovyron » Wed Dec 04, 2019 4:17 am

mclane wrote:
Tue Dec 03, 2019 7:39 pm
This is really not AI but instead artificial stupidity.

But, we have an engine that is different:

Leela Chess Zero 21.1 (41800):
27.h5 g5 28.Df5+ Kh8 29.f4 a3 30.La1 De7 31.Ld1 exf4 32.Tb3
+/- (1.41) depth: 15/45 00:00:31 375kN
-------------------------------------------------------------------------------------------
Yeah, Leela is different. But can Leela beat Stockfish from that position? If not, Leela is the stupid one.

User avatar
mclane
Posts: 18388
Joined: Thu Mar 09, 2006 5:40 pm
Location: US of Europe, germany
Full name: Thorsten Czub
Contact:

Re: stockfish 10 vs. Mephisto III S Glasgow

Post by mclane » Wed Dec 04, 2019 5:19 am

The question is do we really need all these hardware and software explosion ?

Mainly this is needed to beat the programs that are similar. To increase ELO in series of test games.
But is this chess ?!

Wouldn’t it make more sense to use the gigantic resources to bring chess on another level where the machine understands chess instead of finding tactical shots or outsearch the opponent ?
What seems like a fairy tale today may be reality tomorrow.
Here we have a fairy tale of the day after tomorrow....

User avatar
mclane
Posts: 18388
Joined: Thu Mar 09, 2006 5:40 pm
Location: US of Europe, germany
Full name: Thorsten Czub
Contact:

Re: stockfish 10 vs. Mephisto III S Glasgow

Post by mclane » Wed Dec 04, 2019 5:21 am

syzygy wrote:
Tue Dec 03, 2019 10:48 pm
mclane wrote:
Tue Dec 03, 2019 7:39 pm
This is really not AI but instead artificial stupidity.
The progress you've made in about 8 months in this thread disproves the existence of HI.
My part on this posting is mainly to quote the source:
https://glarean-magazin.ch/2019/04/18/k ... ch-report/
What seems like a fairy tale today may be reality tomorrow.
Here we have a fairy tale of the day after tomorrow....

User avatar
Ovyron
Posts: 4410
Joined: Tue Jul 03, 2007 2:30 am

Re: stockfish 10 vs. Mephisto III S Glasgow

Post by Ovyron » Wed Dec 04, 2019 10:27 am

mclane wrote:
Wed Dec 04, 2019 5:19 am
Wouldn’t it make more sense to use the gigantic resources to bring chess on another level where the machine understands chess instead of finding tactical shots or outsearch the opponent ?
To achieve what, exactly?

I keep talking about nails and hammers, but I don't get your point so lets talk about them. You are claiming that to be able to hit certain nails software from 35 years ago used some kinds of hammers. Let's say that they made hammers with gigantic heads that were able to hit big nails and that over time people have designed hammers with much smaller heads but longer handles that allow them to hit a big nail repeatedly to deliver the same effect.

You go and shorten the handle of the Stockfish hammer and then hit big nails with them to no effect while showing that the Mephisto hammer with a huge head from 35 years ago was able to hit that nail with a smaller handle.

Then you say that the future of this looks grim because people have forgotten about making hammer heads bigger and will continue to dedicate their time to longer and longer handles that allow users to hit bigger nails harder by using the hammer several times.

People have gone and shown that for nails big enough, the Mephisto hammer is ineffective, NO MATTER HOW LONG YOU MAKE ITS HANDLE. And then you say that the Mephisto hammer is good enough to hit big enough nails, so it'd be pointless to continue to make handles larger, and people would be much better focusing all their resources into creating the largest hammer head imaginable so that it can hit any kind of nail, no matter the size.

See the contradiction now?

-If in 1984 they made Mephisto III, a program that already understands chess, then doing that is a solved problem. You can't ask for people to use their resources into creating a program that understands chess because it already was done in 1984.
-If Mephisto III understands chess, then why giving it the same resources as today's programs makes it produce such poor moves? Where's its understanding of chess being shown?
-If in the time Mephisto III analyzes 10 positions another program can analyze 100, 1000 or 1000000 positions, then why compare the second program's 10 positions to it? Why punish the second program for being much faster?
-If improving Stockfish by another 40 elo is a waste of time, how would making a program that understands chess yet plays 40 elo weaker than Stockfish isn't a waste of time? And how do you measure its chess understanding? Shouldn't it be by its moves? And if its moves are better, why is it losing its games?

The same answers that lead to the conclusion that current engines' development should halt because there's no use for more elo points are the same that lead to the conclusion that we don't have to bother programming software that understands chess.

Spliffjiffer
Posts: 249
Joined: Thu Aug 02, 2012 5:48 pm
Location: Germany

Re: stockfish 10 vs. Mephisto III S Glasgow

Post by Spliffjiffer » Wed Dec 04, 2019 10:56 am

here this game is archived:
TCEC Cup 2, Finale, last game
http://legacy-tcec.chessdom.com/archive.php

just for the theoretical chess part that was mentioned above:
i honestly doubt that black can hold this after 27.h5 !?...look in the archive.php, Leela played excellently !
Wahrheiten sind Illusionen von denen wir aber vergessen haben dass sie welche sind.

User avatar
mclane
Posts: 18388
Joined: Thu Mar 09, 2006 5:40 pm
Location: US of Europe, germany
Full name: Thorsten Czub
Contact:

Re: stockfish 10 vs. Mephisto III S Glasgow

Post by mclane » Wed Dec 04, 2019 11:26 am

There is no sense in making these programs 40 Elo better.
They will not play better because they have no clue what they are doing. They are blind.
They mainly win because the opponent is at the time 40 Elo blinder. you only do competition. But not chess.

You could, if you like these wins and Elo increases, do the match without changing the software just by pressing the turbo button on machine A and reducing the speed on the opponent machine. You would get the same Elo increase. But overall it will not play chess on another level.

Lc0 is another paradigm. But only if it helps to make a better chess. You must increase the quality of chess. That will overall increase the strength.

As long as the programs do not plan to win but make moves , that win material or mate, they do not understand chess.
What seems like a fairy tale today may be reality tomorrow.
Here we have a fairy tale of the day after tomorrow....

Post Reply