Enough is Enough

https://rebel7775.wixsite.com/rebel/simex

Tested most of the top engines at depth=1, thus testing net similarities.

It's a never ending story, apparently the new boys in town are borrowing or voluntarily sharing nets.

Rephrase : after the long discussion at Talkchess I standby my conclusion in the link above. It's unclear what exactly is going on but the new trend is clear and I don't want to be a part of that.

Time to stop development.

Jammer!!! Helaas.

You should IMO try to continue the 8 bit rom machines. E.g. put endgame knowledge into them. Or try to rewrite them the way the elo rating increases to the level it was in 1989 or arround.

If software has made progress it should be possible to rewrite them the way they get increased ELO.

You have shown that you can make a state of the art chess engine that plays interesting and strong , even with neural net.

But my question is, with the knowledge of today, could somebody write an 8 bit MM4 or MM5 or Academy or Polgar or Milano or nigel short successor, that can jump in elo related to the old dedicated machines ?

Ok. Thats off topic here and would belong more into the topic here:

https://prodeo.actieforum.com/f12-dedicated-chess-computers

Mclane wrote:

...with the knowledge of today, could somebody write an 8 bit MM4 or MM5 or Academy or Polgar or Milano or nigel short successor, that can jump in elo related to the old dedicated machines ?

Yes - using automated engine parameter tuning, and other optimisation techniques.

If normal PC computerchess is senseless because the programs become to similar and the hardware produces more and more draws, going back to 8 bit could be fun again.

Ok these machines had only 32-64 kb rom.

That will be hard without an Apple 2E with large 5 inch floppy drives and without the 6502 source code on it.

Admin wrote:

That will be hard without an Apple 2E with large 5 inch floppy drives and without the 6502 source code on it.

Sounds like your thinking is 40 years behind where it could be.

I have no knowledge of the 6502 assembly language (apart from the fact that the instruction set was small, the 8088 had a medium sized instruction set, and the Z80 had a relatively large instruction set (so the program could be smaller, but it often took multiple clock cycles to execute an instruction)).

So I asked Gemini, "Write code to compare two 64 bit arrays in assembly language for a 6502 CPU." She replied:

Absolutely, let's craft that assembly code snippet for you.

Code:

Code snippet
; Assuming the two arrays are named array1 and array2,
; each with 64 elements (64 bytes)

        LDX #63        ; Initialize counter for 64 elements
        LDA #0         ; Clear accumulator for carry flag use

compare_loop:
        LDA array1,X    ; Load byte from array1
        CMP array2,X    ; Compare with byte from array2
        BNE not_equal   ; Branch if not equal

        DEX             ; Decrement counter
        BPL compare_loop; Branch if positive (more to compare)

        ; If we reach here, arrays are equal
        ; You can set a flag or jump to an 'equal' label

not_equal:
        ; Arrays are not equal
        ; You can set a different flag or jump to a 'not_equal' label

"Important Considerations for 6502

* Zero Page: If possible, place your arrays in the Zero Page (first 256 bytes of memory) for faster access.

* Carry Flag: The CMP instruction sets the carry flag if the values are equal. We use this to our advantage.

* No Direct 64-bit Compare: The 6502 doesn't have instructions for directly comparing large values. We iterate byte-by-byte.

* Labels: Replace the comments with actual labels for branching if needed in your larger program.

Let me know if you have any specific modifications or further questions!"


Well not quite perfect: the code seems to be comparing 64 bytes rather than 64 bits - but I've been given a good start!

Of course, having written the code, you'll need somewhere to run it - link.

I think it should be possible to write in a high level language c or c++ for 6502 on a modern pc. I do even think one could compile for 6502.


E.g.
https://www.cc65.org/

Mclane wrote:

I think it should be possible to write in a high level language c or c++ for 6502 on a modern pc. I do even think one could compile for 6502.


E.g.
https://www.cc65.org/

Even so, what remains missing is the code that is needed to function within a dedicated unit, then I need to buy an eprom machine, eproms and a MM5 again. Or maybe it could be done via the UCI protocol and put it into the Phoenix system.

Its is all virtually done today.
One can try out the stuff in the emulators.
And then it can run either on old modules or , as you said, on the machines (mephisto phoenix) that uses the old roms.
Franz Huber on his jimdo website has done a marvellous job bringing almost ALL engines into virtual emulators.

https://fhub.jimdofree.com/

In a team with others they also made it possible to automatize the old dedicated machines in winboard and let them play automatically against each other. They even patched bugs and were able to read out the old H8 cpus (Novag Diamond etc.) that had the rom inside the chip.

All you need is a pc. And you can run the dedicated chess computers on emulators.

For me the major question is: can software progress make better engines on 8 bit hardware with the knowledge we have today ?  

Here you can see the ssdf rating of these units:

https://ssdf.bosjo.net/long.txt

E.g. MM4 5 mhz:
391 Mephisto MM 4 6502 5 MHz                1814

Then…
373 Mephisto MM 5 6502 5 MHz                1883
…
354 Mephisto Nigel Short 6502 5 MHz         1950



Eds Masterpiece 1950 ELO.  I think the Nigel Short and the Milano use the same hardware. And 64 KB ROM.

Rebel 5.0 and MM4, MM5 use 32 KB ROM.

The Milano (maybe due to passive playing style) is
378 Mephisto Milano 6502 5 MHz              1873


So ed was capable to do 1950 ELO out of this 8 bit hardware.
It was a close race with dave kittinger :

379 Novag Super Expert C 6502 6 MHz Sel 5   1868

Fidelitys Spracklen engine was weaker:

404 Fidelity Par Excellence 6502 5 MHz      1745

Thats why they early began on 68000 to win the race against Richard Lang who used 68000 too.

But even with 68000 the highest elo they got on the normal 68000 16 mhz was only:

369 Fidelity Mach III 68000 16 MHz          1900

So eds 8 bit nigel short was 50 elo better then spracklens running on the 68000.
Kittinger then also tried to bring his engine on 68000:

368 Novag Diablo 68000 16 MHz               1911

In 1985 Richard won in Amsterdam on a Motorola:

385 Mephisto Amsterdam 68000 12 MHz         1834
And due to tuning together with Ossi weiner the engine made big progress ….

377 Mephisto Roma 68000 12 MHz              1877

374 Mephisto Dallas 68000 12 MHz            1878
Then came the hash table machines:

364 Mephisto Almeria 68000 12 MHz           1926

But look, it took richard lang many championships to reach ed schröders ELO record with a 68000 machine,

353 Mephisto Portorose 68000 12 MHz         1952

The same level then Ed with bis 8 bit machine.

Richard lang then made enoughwr big jump with the vancouver:

343 Mephisto Vancouver 68000 12 MHz         2009

Ed then changed from 8 bit to RISC Arm CPU with 12-14 mhz depending on the hardware this PC card was used and later it was put into a dedicated unit.

On this ARM machine Ed showed that his engine was very good and he won the title.
328 Chess Machine Gideon 512K ARM2 16MHz    2097

Now coming back to the topic:

Ed Schröder was the master in getting the most ELO out of this 8 bit universe.
It took a few years of course, from the first appearance and a selective engine coming written in BASIC with later BASIC combined with machine code for the search.

When Ed began for Hegener and Glaser, they had a nearly brute force  engine by Ulf Rathsmann:


418 Mephisto MM 2 6502 3.7 MHz              1682

Then came eds first commercial engine Rebel 5.0:
406 Mephisto Rebell 6502 5 MHz              1734

1950 (Nigel Short) - 1734 (Rebel 5.0) = 216 ELO software progress on the SAME 8 Bit 5 mhz Hardware !!!!!


This is amazing, isn't it ??


No other programmer was able to do so, IMO.

Even Richard Lang on the 68000 machines (from Amsterdam to Vancouver) had only 175 ELO progress.

2009 (Vancouver) - 1834 (Amsterdam) = 175 ELO software progress


And yes ed, the mephisto phoenix runs shredder , hiarcs and other uci engines !!! And there is no need for the I/O code of HG modules.

Hmmm... are we asking Ed to make washboards in a world of washing machines?




To be fair, there are still dedicated chess computers available - link - but it's a very different market from the golden era of the 1980s.

What the pioneers were doing in the golden era of computer chess was, of course, writing assembly language code to enable a chess board with switches to play a good game of chess. Ordinary people being able to play chess against a computer was a new and novel capability.

So as well as writing assembly language, they were also:

1. Pioneering

2. Inventing

3. Creating art

Given that the authors were not actually assembly language geeks, but that assembly language was the medium through which they expressed their art, maybe we should think more about the pioneering, inventing and artistic aspects of what they were doing.


Newsflash!!!

While chess engines are closing in on becoming unbeatable, there are still lots of ways to apply computing to chess that are well below their maximum level.

Maybe these areas are where pioneers should be looking now.

Example: it takes around 2 decades of dedicating one's life to chess, to the exclusion of other learning (like social skills), to become a GM. With better training skills, I'm sure this could be MASSIVELY reduced. A pioneer/inventor/artist type of person could take a look at the process of training chess skills, and find ways to improve it. Most of the stuff we teach chess students is unnecessary: they will pick up most of the basics as a by-product of learning more sophisticated chess, for example.

My roma 32 bit (richard lang) versus Nigel Short (ed schroeder):




The Roma runs in a handmade motorola 68000 machine with 24 instead of 12 mhz. This hardware is known as GM Petrosian and can switch between:
Mephisto III S Glasgow (1984), Mephisto Amsterdam (1985) Mephisto Dallas (1986) and Mephisto Roma (1987). The later engines by Richard Lang. The first by Thomas Nitsche and Elmar Henne. So this is a 4 in 1 machine.

A little faster with 24 mhz and also none of the known „wait-states“ the old machines modules had.
Because 68020 mainly is 2 times faster then 68000, this is a „68020“.

Nigel Short with normal 5 mhz 6502 CPU.



On the backside you see the switch and the serial number of the machine.

And now tell me that these machines are not fascinating or it makes no fun to use them. I am sitting on the balcony in the sun and have a very nice afternoon with Richard Lang and Ed Schroeders engines . Music in the background (Alan Parsons Projects, Jean Michel Jarre, Barclay James Harvest or Cash) and a good audio-book.

I don't want to let you die without the PGN:

Here the marvellous winning game of Nigel Short:

[pgn] [Event "40/120"]
[Site "SCW"]
[Date "2024.08.17"]
[Round "1"]
[White "Mephisto Nigel Short 5 mhz"]
[WhiteElo "1600"]
[Black "GM Petrosian = Mephisto Roma 68000 24 mhz"]
[BlackElo "1600"]
[Result "1-0"]

1.d4 d5 2.c4 Nc6 3.Nf3 Bg4 4.cxd5 Bxf3 5.gxf3 Qxd5 6.e3 O-O-O 7.Nc3 Qd7 8.d5
Ne5 9.f4 Ng6 10.e4 e6 11.Bh3 Bc5 12.f5 Nf8 13.Bg5 Ne7 14.dxe6 Qc6 15.Qg4 f6
16.Bh4 Qb6 17.O-O h5 18.Qg3 Bd6 19.Qe3 Qxe3 20.fxe3 Bc5 21.Rae1 Rd2 22.e5 Rxb2
23.exf6 gxf6 24.Bxf6 Rg8+ 25.Kh1 Nh7 26.Bh4 Rc2 27.Ne4 Nc6 28.f6 Kb8 29.f7 Rf8
30.Nxc5 Rxc5 31.e7 Nxe7 32.Bxe7 Rd5 33.Rd1 Rxd1 34.Rxd1[/pgn]

Admin wrote:

https://rebel7775.wixsite.com/rebel/simex

Tested most of the top engines at depth=1, thus testing net similarities.

It's a never ending story, apparently the new boys in town are borrowing or voluntarily sharing nets.

Unfair competition again.

Time to stop development.

Are you sure the new boys even think of this as a competition in the same way you do?
I had a brief look at ALL the chessprograms installed on my now oldest notebook, which means the last 3 years:
Arasan 24.1
Crafty 25.6
2 * Chess System Tal
Demolito
2* Dragon
2* Fritz
1 * Komodo
2* LC0
Patricia
ProDeo
8* Rebel
Rodent NN
ca. 35* Stockfish
Toga IV 1.1
Velvet v4.1.0
It is safe to call me a pretty nerdish guy in recent years - I spend quite a lot of time with chessengines again.
It seems I have used none of the in your opinion controversial engines even once.
I had heard about some of the engines you surveyed, but clearly not about even a majority of them.
Besides the guys who do ratinglists no one ( and I mean it !) cares about most of the engines availlable to the least. So why shouldn't they share and do what is fun for them?!
In my case it is obvious that a new and improved Rebel would be way more relevant than any improved version of motor ( what a name!) or the likes of Ethereal.
Take this as a friendly nudge to continue
Peter

Alright, I will give some more information.



This was the situation about one year ago, CSTal and REBEL listed high.

And now we are moving to today.



No sign of CSTal and REBEL, they don't even are in the top-16, an invasion of the new boys in town popping out of nothing.

Am I am bad loser?

Nope.

Let's look at the similarity data.




This is the normal similarity of the rest of the top engines which looks the same since the introduction of NNUE, all original nets with common low similarity, I have nicknamed them as "old boys".

Compare "old boys" with "new-boys" and you will notice a major difference in similarity, the evidence suggest they have a common ancestor.

Now, in the HCE days as a programmer you were supposed to write your own original evaluation function else you would get the death penalty from the community, surely you all remember the Rybka-Fruit case.

And it should not be any different now that HCE is replaced by NNUE, you simply should create your own original NNUE and put all your blood, sweat and tears herein like everybody else.

i think its just that all the top engines use leela data nowadays, including rebel, so you get this

That's way too simple.

But that way too simple has also his limits..they need to find a new way like nnue2 or something else to see nice gains again.

1. There are about 200 billion Leela positions in chunks available for free download spread over a period of 2½ year and the chunks differ a lot in quality. The chance a group of people pick the same data is almost zero.

2. Secondly there is a huge creativity in NNUE programming, a good example is Patricia.

3. Chris and I use different Leela data, see [1], even have the exact same root search + QS and nevertheless pass the test. It's because we both go our own way in NNUE programming.

What I see are nets with a common ancestor, the sim util is about math and math never lies.

Admin wrote:

...math never lies.

10–10=15–15

2(5–5)=3(5–5)

2=3(5–5)/(5–5)

5–5 cuts the other 5–5

So, 2=3

1+1=3

Hi,

it is getting obvious most games in top engines matches end in a draw, up to that point: to get significant results (for strength measurement), some bias need to be introduced (such as unbalanced openings), thus enabling some wins or losses and hoping for some measurable performance discrepancies.

Isn't this an indicator these top chess programs are simply converging towards the very little ideal moves available, whatever the considered position? They are so strong that of course blunders and errors can't anymore happen, but tiny inaccuracies as well, resulting in very small sets of acceptable moves per position?

If so, similarity looks natural to me; unless specific development effort, such as favoring style at the expense of strength (as the author of Patricia does). Isn't similarity the result of lazy convergence led by AI/NNUE, rather than copycat development by authors?

Warm rgds,
Eric

Eric, the test I did are all on depth=1, similarity then should be low.

Admin wrote:

https://rebel7775.wixsite.com/rebel/simex

Tested most of the top engines at depth=1, thus testing net similarities.

It's a never ending story, apparently the new boys in town are borrowing or voluntarily sharing nets.

Unfair competition again.

Time to stop development.

There is a new and fair competition: Not the strongest engine, but the most aggressive engine. The best Rebel (Rebel EAS) is on rank 35 of my full EAS-Ratinglist. So, much room for improvement and competition here.

                                                          bad  avg.win
Rank  EAS-Score  sacs   shorts  draws  moves  Engine/player
-------------------------------------------------------------------
  1    429481  51.61%  38.03%  05.49%   66   Patricia 3.1 avx2  
  2    384942  52.55%  35.77%  05.19%   69   Patricia 3.0 avx2  
  3    197919  31.18%  29.46%  17.09%   71   Revenge 1.0 avx2  
  4    193706  22.65%  25.49%  08.39%   70   Stockfish 16.1 240224  
  5    192881  21.25%  26.18%  09.40%   70   Stockfish 16 230630  
  6    186267  25.24%  24.69%  13.42%   72   Velvet 7.0.0 avx2  
  7    181381  20.24%  25.04%  10.47%   70   Stockfish 15.1 221204  
  8    161767  25.14%  21.45%  13.73%   72   Velvet 7.3.0 avx2  
  9    158026  16.26%  26.23%  11.07%   69   Torch 2 popavx2  
 10    151721  15.63%  27.12%  13.45%   69   Torch 1 popavx2  
 11    150692  16.75%  22.03%  12.22%   71   Crystal 7 CMB avx2  
 12    149587  17.17%  27.32%  13.10%   69   KomodoDragon 3.2 avx2  
 13    138735  11.40%  40.08%  26.11%   67   Stockfish final HCE  
 14    133142  15.43%  20.88%  12.81%   73   Stockfish 14 210702  
 15    126631  13.20%  19.91%  13.40%   72   Torch 3 popavx2  
 16    126281  15.26%  20.98%  16.17%   75   Uralochka 3.40a avx2  
 17    122289  13.02%  20.75%  14.97%   73   Obsidian 12.0 avx2  
 18    122219  15.26%  21.34%  14.35%   72   KomodoDragon 3.3 avx2  
 19    121885  13.18%  23.70%  15.94%   73   Revenge 3.0 avx2  
 20    120762  23.48%  13.75%  14.59%   80   Arasan 24.2.1 avx2  
 21    116182  10.32%  31.17%  22.16%   71   Slow Chess 2.9 avx2  
 22    115855  16.94%  13.12%  15.76%   80   Lizard 10.5 avx2  
 23    106475  16.26%  19.40%  16.17%   74   RubiChess 230918 avx2  
 24    104928  12.56%  18.44%  15.46%   75   Obsidian 12.29 avx2  
 25    103692  12.95%  18.85%  17.90%   75   Igel 3.5.0 popavx2  
 26    103067  16.83%  13.96%  17.00%   82   Lizard 10.4 avx2  
 27    102118  14.11%  15.47%  15.16%   80   Uralochka 3.41a avx2  
 28    101785  14.93%  11.94%  15.75%   79   Obsidian 9.0 avx2  
 29     98177  12.15%  16.38%  15.87%   76   Obsidian 13.0 avx2  
 30     95585  13.82%  17.32%  16.89%   75   RubiChess 240112 avx2  
 31     89826  11.99%  09.46%  15.93%   80   Obsidian 10.0 avx2  
 32     88339  10.89%  14.68%  16.29%   77   PlentyChess 2.0 avx2  
 33     88254  12.94%  11.61%  15.84%   80   PlentyChess 1.0 avx2  
 34     86688  15.98%  11.30%  16.04%   82   Ethereal 14.38 avx2  
 35     85704  16.14%  12.04%  19.27%   82   Rebel EAS avx2

I have an EAS 220.xxx version, but it loses 50 elo.

I wonder if we slow down these patricia 3.1 it could be an interesting opponent for humans and old dedicated machines.