stockfish 10 vs. Mephisto III S Glasgow

[mclane]

who is interested in a chess program that needs to "EVALUATE" millions of NPS ????

its clear that such a chess program is an idiot that gets a fast hardware.


IMO all these "matches" only show that computerchess has made no progress other then the hardware beeing used has made big progress and the weak software suddenly comes into ply 40.

and you call this CHESS ???
thats not chess.

[Uri Blass]

who is interested in a chess program that needs to "EVALUATE" millions of NPS ????

its clear that such a chess program is an idiot that gets a fast hardware.


IMO all these "matches" only show that computerchess has made no progress other then the hardware beeing used has made big progress and the weak software suddenly comes into ply 40.

and you call this CHESS ???
thats not chess.

It is not correct.
The software is designed for the new hardware and not for small number of nodes per second.

I believe that the problem is also not the evaluation function or at least not only evaluation function that is designed for more than small number of nodes.
If you disable some pruning stockfish with the same number of nodes is going to do singinificantly better with the same small number of nodes.

[JohnW]

who is interested in a chess program that needs to "EVALUATE" millions of NPS ????

its clear that such a chess program is an idiot that gets a fast hardware.


IMO all these "matches" only show that computerchess has made no progress other then the hardware beeing used has made big progress and the weak software suddenly comes into ply 40.

and you call this CHESS ???
thats not chess.

So how is it that Mephisto III Glasgow running as a UCI engine on the same hardware as Stockfish 10 loses?
Personally I wouldn't be able to discern between playing chess program with knowledge and a chess program that evaluates.

[Raphexon]

Comparing nodes 1:1 between two engines is also false.

As it stands, SFdev is the strongest entity for a Pentium II (Or any CPU with MMX or better), and that's a CPU from 1997.
SF-HC would be the strongest on the older 32 bit CPUs. (going back all the way to 1985)

On an i386 SF11 should still be able to get at least 500 nps, if you want to have an actually fair matchup. (Hint: SF will smash)

To say software has made no (or marginal) progress in 30-40 years is downright idiotic.

[mclane]

who is interested in a chess program that needs to "EVALUATE" millions of NPS ????

its clear that such a chess program is an idiot that gets a fast hardware.


IMO all these "matches" only show that computerchess has made no progress other then the hardware beeing used has made big progress and the weak software suddenly comes into ply 40.

and you call this CHESS ???
thats not chess.

So how is it that Mephisto III Glasgow running as a UCI engine on the same hardware as Stockfish 10 loses?
Personally I wouldn't be able to discern between playing chess program with knowledge and a chess program that evaluates.

We would have to find out how big the factor is.

How much is a typical pc of today faster then the 68000 cpu with 12 mhz and wait states (slow ram).

Todays chess engines rely on search depth. Why. Because it is there,

Why not use it ?

But is this playing chess or playing chess with the hardware instead of the software ?!

I give you another example. Today many cars have more then 100 PS.
I remember days when a car had 9-29 PS !!

Now the main question, is the >100 PS car so much better then the 9-29 PS car ??

Of course todays software has made much progress in all areas.
Of course todays cars have made progress in all areas.

But people who drive a car with only 9-29 PS are suddenly very surprised how good those old cars drove.

[Uri Blass]

Comparing nodes 1:1 between two engines is also false.

As it stands, SFdev is the strongest entity for a Pentium II (Or any CPU with MMX or better), and that's a CPU from 1997.
SF-HC would be the strongest on the older 32 bit CPUs. (going back all the way to 1985)

On an i386 SF11 should still be able to get at least 500 nps, if you want to have an actually fair matchup. (Hint: SF will smash)

To say software has made no (or marginal) progress in 30-40 years is downright idiotic.

Of course using the same number of nodes is not the way to test if software made progress in playing strength but simply using the same hardware but if you test by distorted metric of 1000 nodes per move then I see no reason to take stockfish as the representive of software of today.

Better is to make a tournament of chess engines with 1000 nodes per move and test the winner against mephisto.
You already wrote that Herman2.8 is better than Stockfish in these conditions.
Maybe there is some software that is better than Herman2.8 in the same conditions.

[cpeters]

who is interested in a chess program that needs to "EVALUATE" millions of NPS ????

its clear that such a chess program is an idiot that gets a fast hardware.


IMO all these "matches" only show that computerchess has made no progress other then the hardware beeing used has made big progress and the weak software suddenly comes into ply 40.

and you call this CHESS ???
thats not chess.

I'm testing this since days.

Stockfish on a (mid nineties/motorola 68k) emulated hardware runs properly/adequately throttled in the middlegame sub 400 Nodes/second - nowhere near millions and still it smashes the Mephisto Amsterdam.
Will it it get smashed by other dedicated comps? For sure/possible.
It's interesting...


Greetings!

meanwhile: if you're interested I'll send you the pgn*; with the deterministic behavior of the emulation you can cook games then even on real wood. But there would be no fun in that, just wasted time.

*or a debug file from cutechess with nps ond so on

[mhull]

who is interested in a chess program that needs to "EVALUATE" millions of NPS ????

its clear that such a chess program is an idiot that gets a fast hardware.


IMO all these "matches" only show that computerchess has made no progress other then the hardware beeing used has made big progress and the weak software suddenly comes into ply 40.

and you call this CHESS ???
thats not chess.

So how is it that Mephisto III Glasgow running as a UCI engine on the same hardware as Stockfish 10 loses?
Personally I wouldn't be able to discern between playing chess program with knowledge and a chess program that evaluates.

We would have to find out how big the factor is.

How much is a typical pc of today faster then the 68000 cpu with 12 mhz and wait states (slow ram).

Todays chess engines rely on search depth. Why. Because it is there,

Why not use it ?

But is this playing chess or playing chess with the hardware instead of the software ?!

I give you another example. Today many cars have more then 100 PS.
I remember days when a car had 9-29 PS !!

Now the main question, is the >100 PS car so much better then the 9-29 PS car ??

Of course todays software has made much progress in all areas.
Of course todays cars have made progress in all areas.

But people who drive a car with only 9-29 PS are suddenly very surprised how good those old cars drove.

We can simplify the problem by thinking in terms of "path length to Elo". In other words, how many machine instructions are required to be executed in order to play at game of classical chess at an Elo of FIDE 2600 (or 2700 or 2800...)?

Obviously software has something to do with the answer since not all chess software yield the same strength on the same hardware. But somewhat in the spirit of the uniform-platform competition, a software choice could be made for a "uniform software" competition where the variable was the instruction-rate, which is to say, the speed of the hardware. If the instruction rate for a "chess workload" could be calculated (this was done for decades on mainframes that specialized in diverse workloads like Fortran, CICS, IMS, DB2, Batch, TSO, etc.) then something similar could be calculated for a particular program at a resulting Elo performance.

Where I work the path-length of high-speed transaction types in their real-time database system has been studied over 50 years. The path-length of each transaction is known and is continually updated when there is a software change. NOTE: This became more difficult with the advent of superscalar processor designs but path-length is still estimated today for this decades-old database system.

At some point I began to think about computer chess strength in the same way. How many machine instructions are required to be executed to play a game at a particular Elo strength? Can we actually measure this? We could probably estimate it. How long ago was it when a human could even maintain an even score with a computer program? If we know about when that was and the hardware/software that is in play at the time, we could postulate a path-length-to-grandmaster value.

Computer play at >=2600 probably came and went a couple decades ago, at least.

Although he didn't express it in these terms, I think Steven Edwards was exploring the possibility that path-length-to-Elo could be shortened using more intelligent pattern recognition techniques in the search. At the height of his enthusiasm, he hoped his program, Symbolic, would be able to play grandmaster chess running on a 700 Mhz, single core Apple PPC laptop.

[mclane]

Of course Mephisto III has no chance if you put them both on 68000 hardware and even when stockfish makes 400 NPS it is still stockfish.

But in my experiment stockfish has to use as many NPS mephisto III S is using.
You can say this is unfair because it was never constructed to play under these circumstances . But it was only an experiment.

Todays chess engines do heavily use the hardware resources.
Even without evaluation they have due to the hardware power enormous search depths.
My point is that the big success in the hardware area MADE the new engines so strong.
They rely on this hardware progress.

Without this hardware power (not only speed but also limited resources concerning ram and rom) the engines would not have these high elos.

There had been huge developments in algorithms, but most of these algorithms are in the search area.

There is not much progress in the chess area.

Instead of teaching the engines how to play chess modern chess programmers rely on hardware.

The hardware progress replaces the knowledge progress.

Thats why these programs until today do not plan chess.
They make a move on the base of their search and evaluation function but they do not play chess.
Its insane that they have elos of 3000 and more and have no real clue about chess,
that is because they rely on hardware power and search instead of understanding how chess works.
They are mainly move generators. And e.g. by chance they find a mate.
But they do not plan to mate.
They do not plan to move pieces into this or that area to do this or that.

They are still stupid fucking machines.

I would have preferred that computerchess develops into another direction.
AI.
But instead the engines play like robots. Mechanical chess.

Why do they need a search tree anyway to play chess ?!
They could analyse the position, know the rules and develop an idea and try it out without millions and millions of NPS beeing computed.
Instead trying out a few moves that work.

That was the idea of mephisto III S glasgow.

Analyse the position and use a few good looking moves and trying to build a tree on a the base of this better moves.

This gives the dedicated chess computer doing only 1-3 NPS (8 bit) and 4-10 on the motorola 68000 16 Bit hardware.

[mclane]

IMO it would have made sense to continue Mephisto III and come out with a version Mephisto IV (or orwell4 to handle misunderstandings with ed schroeders MM4) with hash tables and more knowledge.
For that the old source code of Mephi III S would be needed and programmers beeing able to program in CDL2, the programming language it was once designed.
https://www.schach-computer.info/wiki/i ... -S_Glasgow