stockfish 10 vs. Mephisto III S Glasgow

[mclane]

As I said before, you maybe don’t get it, so I repeat:
Why stockfish needs millions or hundred thousands of NPS ?

Why ?

Because it has no clue about chess.

It is the the materialistic bean counting method.
When the programmer has no clue about chess, he uses “smart algorithms” for search instead of concentrating on the chess part of the problem .


We had this several times before in history of computerchess.
Or do you think this is the first attempt to solve chess via heavy tree search ?
It isn’t.
Frans morsch came years over years with pre processing null move monsters.



But this brute force method is artificial stupidity.
You create a machine that misuses resources over resources, huge amount of hardware, huge amount of ram, rom and years and years of development and millions of automatically played games to mainly hide the fact that the programmers are not interested in chess content,
They could in the same way program stockfish checkers or stockfish go or backgammon,

For them the thing is only a tree search problem.

And they solve it in the materialistic way using zillions of NPS and zillions of autoplayer games to find out which change was “right”.

There is a door, that is locked, and they come with Dynamite to open it.

I prefer using the key.
Therefore I look into the 5-10 keys I have in my hand and put the one into the lock that I know will open the door,


Of course stockfish is a strong clever and smart search monster.
But do we need this for chess ?

The programs mainly fight against each other’s,
Houdini versus Komodo versus stockfish,
And in earlier years rybka was in and in earlier years shredder, hiarcs and mchess and Zappa were in,


Today a small group of maybe 3-4 engines reach the top elo ranking.
But these programs are very very similar and are attacked by NN engines.
These new neural net engines use similar waste of resources, the only difference is, there is no programmer.

But both concepts use huge hardware, big ram and rom , waste millions of NPS.

Do we really need this ??

Why ?


Most amateur engines that come out today use the open sources and jump in with 2700-2800 elo (e.g. Xiphos).

So they mainly copy the big ones.

This is computerchess today,

It’s all about hardware and massive usage of ram and rom and not about chess anymore.

[chrisw]

As I said before, you maybe don’t get it, so I repeat:
Why stockfish needs millions or hundred thousands of NPS ?

Why ?

Because it has no clue about chess.

not true. the evaluation function of SF knows a good deal about chess. The development paradigm is all about crowd sourced input of ideas and code and improvements, and much of that, some of that, enough of that has been with chessy knowledge ideas.

It is the the materialistic bean counting method.

Well, the paradigm has remained based on “material”, but the evaluation algorithm contains way more chess concepts that your old programs of the past.

When the programmer has no clue about chess, he uses “smart algorithms” for search instead of concentrating on the chess part of the problem .

This is true, and very few of the chess programmers can claim to be much above patzer level. Bruce Moreland once write that his motivation was that went to his local chess club, lost heavily, realised he was unlikely to get good himself, so decided to write a program to get good instead. I would guess that is a general motivation in this field.

We had this several times before in history of computerchess.
Or do you think this is the first attempt to solve chess via heavy tree search ?
It isn’t.
Frans morsch came years over years with pre processing null move monsters.

SFis an alliance between contributors who can hardly play chess but can program and contributors who can play chess and make suggestion (or code even). Seems like a strong alliance.

But this brute force method is artificial stupidity.

as is writing the same thing over and over again and never lostening to what anyone says.

You create a machine that misuses resources over resources, huge amount of hardware, huge amount of ram, rom and years and years of development and millions of automatically played games to mainly hide the fact that the programmers are not interested in chess content,
They could in the same way program stockfish checkers or stockfish go or backgammon,

Well, yes, they are like retailers who sell anything in a box, doesn’t matter what, as long as it sells.
But, the open source group projects are not like this. They also contain specialists who contribute knowledge. SF team more so than the NN team(s). The NN programs really don’t care at all what is the underlying (Deepmind made a big deal that the learning network technology was NOT field specific) so, naturally, people who “understand” the field will congregate around SF, because they are useful, but not around LC0 because they are not very useful there.

For them the thing is only a tree search problem.

And they solve it in the materialistic way using zillions of NPS and zillions of autoplayer games to find out which change was “right”.

There is a door, that is locked, and they come with Dynamite to open it.

I prefer using the key.
Therefore I look into the 5-10 keys I have in my hand and put the one into the lock that I know will open the door,

But you don’t have a key.

Of course stockfish is a strong clever and smart search monster.
But do we need this for chess ?

The programs mainly fight against each other’s,
Houdini versus Komodo versus stockfish,
And in earlier years rybka was in and in earlier years shredder, hiarcs and mchess and Zappa were in,


Today a small group of maybe 3-4 engines reach the top elo ranking.
But these programs are very very similar and are attacked by NN engines.
These new neural net engines use similar waste of resources, the only difference is, there is no programmer.

But both concepts use huge hardware, big ram and rom , waste millions of NPS.

Do we really need this ??

Why ?


Most amateur engines that come out today use the open sources and jump in with 2700-2800 elo (e.g. Xiphos).

So they mainly copy the big ones.

This is computerchess today,

It’s all about hardware and massive usage of ram and rom and not about chess anymore.

It will be good for you to talk your way out of this obsessive computer chess hobby.

[mclane]

So what is the future?
Making Stockfish 11 that is 40 elo stronger?
Followed by Stockfish 12 that is again 30 elo better.
Or waiting until the NN beat Stockfish with another graphic card ?



The future is about elo. And it is about search. It is about hardware.

[jp]

There is a door, that is locked, and they come with Dynamite to open it.

I prefer using the key.
Therefore I look into the 5-10 keys I have in my hand and put the one into the lock that I know will open the door,

But you don’t have a key.

Yep. And we don't even know a key exists.

[jp]

The NN programs really don’t care at all what is the underlying (Deepmind made a big deal that the learning network technology was NOT field specific) so, naturally, people who “understand” the field will congregate around SF, because they are useful, but not around LC0 because they are not very useful there.

Though it really IS field specific in many ways, whatever DeepMind wants to claim...

[smatovic]

Nobody forces you to participate in fruit fly races, or fruit flies on
steroids. I am sure there are plenty of computer chess fans out there playing
older engines, or dwelling in computer chess history, or curious about new kind
of NN fruit flies.

I am an amateur computer chess programmer, and in no Elo rush, still I enjoy
this hobby and the community around it.

--
Srdja

[mclane]

I am only disappointed that the variance reduced and no alternative or new approaches happen.

I know Stockfish is a clever program. But it is search based.
And i think this is not the way to go.
With search Stockfish outcomputed any opponent to an elo level that is far away from the best human chess player.
But with this strategy in cannot change the quality of chess.
It can reach a higher elo.
But it is no quality step forward.
Because it is about smart algorithms and not about chess itself.

[smatovic]

I agree, who knows how many alternative approaches lay in a drawer with dust
on it cos they were simply not competitive to the Elo tuned engines.
But hey, that is currently the way the competition goes...and regarding new
approaches - we got the first wave of competitive NN engines, w00t!

--
Srdja

[Frank Brenner]

So what is the future?
Making Stockfish 11 that is 40 elo stronger?
Followed by Stockfish 12 that is again 30 elo better.
Or waiting until the NN beat Stockfish with another graphic card ?



The future is about elo. And it is about search. It is about hardware.

In the past as well as today and in the future chess programming is about the optimization of the code and the addition of very intelligent search techniques and the integration of knowledge, which is meanwhile generated by self-learning neural networks.
These neural networks have to be created and optimized by the programmers first.
For this purpose a lot of intelligence is put into the development of the software and into the architecture of neural networks.

This leads to the fact that nowadays the strongest chess programs achieve 3500 Elo while the old Mephisto 3 program - even if it were rewritten on a graphics card or on the fastest CPU - would only achieve a maximum of 1700-1900 ELO.

The reason for this is the very poor code of Mephisto 3, which still operates with ancient and trivial algorithms that are outdated today.

[mclane]

That is your interpretation.
I do not subscribe to it.
Better programming artificial intelligence then artificial stupidity.
Most of the progress in computerchess comes from hardware progress.