The future of computer chess

[mclane]

That's the origin of my posting.
Is there an algorithmic progress or is the progress being made via hardware progress.

Chris W. is completely right making this point.

I think the only way to find out is by trying it out to beat Ed Schröders Mephisto Nigel short on 6502 5mhz 64/8 ROM/RAM.

Because Ed Schroeder leads the 8 bit engines.

I have spoken with Martin Bryant (Colossus Chess) about this and he also thinks that hardware and software progress is somehow connected because the hardware changes make software changes possible.

[syzygy]

I would like to see how today's programs run on 6502 or z80a or Motorola M6809,
and if they are capable to beat Ed Schröders engines from
1991
https://www.schach-computer.info/wiki/i ... sto_Milano
and 1993
https://www.schach-computer.info/wiki/i ... igel_Short

If an algorithm progress has been made, it should be possible to put it into 64 KB engine size.
It it doesn't fit into 64 KB ROM, 8 KB RAM it is maybe not a software progress but a hardware progress.

The programmers from the 1980s had the unfair advantage of having powerful hardware compared to Alan Turing in 1948.
Ergo, there has not been any software progress since 1948. QED.

Btw, what is your comment on single-core 2020 pre-NNUE Stockfish 15 totally crushing 2005 Fruit 2.1 on the same hardware at 1:1000 time odds? No evidence of software progress? Really?

[jefk]

there has not been any software progress since 1948.

no, not at all, alfa beta ? hah! nnue ? hah !

for the rest not going to react to such ignorance (or trolling);
nor should you mr syzygy (it's blatant nonsense ofcourse)

Example (in general) these AI (LLm's) are just illusions
only hardware !

PS not going to react further into such drivel; most innovations were software;
point. Made more easy with better hardware. Nevertheless; it's about
algorithms; the whole universe is run by algorithmsl but thats something
maybe for later (or preferably for a physics forum) maybe:
Obviously the Mcts engines as Leela also are getting advantage from Gpu
but then again ,this basically was an algorithmic improvement (like alfazero)
PS2 no mr Tf there's no 'hidden algorithm' for chess, its a decision tree and then
you have minimax (von Neumann game theory); if you want to dwell into this further
https://web.math.ucsb.edu/~crandall/mat ... inimax.pdf
But ok then there' s still the problem about the evaluation ofcourse at the end of the tree;
the Alfa beta algoritm was a brilliant invention/innovation but still fully consistent
with the minimax theorem. As for the evals at the end of the tree then we got the Nnue innovation
(inspired by Sjogi) which made another *algorithmic* (software) jump ahead

Best of the engines after a while whether you like it or not was SF although they
didnt' like my prediction that it will end in a draw cq is a draw
(the game of chess)
So now we have chess960, TCEC openings (soon to be tried at top level in
an ICCF tourn as well) etc. And then there's human chess; with beauty, or not
imo we may see some style differences (at least in opening choices) , coming
decades, and nowadays, computer chess is the forefront of such ideas,
once the (usually arrogant) human chess GM's laughed about computer chess,
well not anymore. they have their seconds sorting out correspondence chess
games, and even then, imo usually are not at the forefront of real chess theory
with many new gambits, and so on.

[jefk]

about this (imo awesome) alfa beta algorithm
https://en.wikipedia.org/wiki/Alpha%E2% ... ta_pruning

then we got the null move pruning , not sure if this has ever been validated
for (theoretical/mathematical) consistency but it made impact;
https://www.chessprogramming.org/Null_Move_Pruning
https://en.wikipedia.org/wiki/Null_move


so did many other tweaks.

Sf is a Projectteam whether you like it or not, and it worked, although followed closely by contenders as Obsidian etc.
P5 and so on are interesting distractions showing other ways for human chess players in their styles

But then for slow computer (or ICCF correspondence games we're in the end of
the progress because guess what, chess is a draw with perfect play
(hey heard that before, maybe have you as well ?)

So the future of human chess is now ches 324 and we can
also develop opening books for that btw

[syzygy]

Current hardware would mean 1 cpu since the older chess engines cannot run more then one core.
In a uniform platform tournament, so same hardware, one can try to find out.

I overlooked your reply.
Yes, single core Fruit 2.1 vs single core Stockfish, last release.
Fruit will be crushed.
Do you doubt this?

Already done:

Thanks! And, wow... Fruit 2.1 from June 2005 against single-core pre-NNUE Stockfish from 2020, both on 2020 hardware, with 1000:1 time odds for Fruit 2.1, and Fruit 2.1 got crushed.

And Fruit 2.1 was not a weak engine at all!
One could still "blame" Fruit 2.1 though, because its release as open source certainly contributed to the progress in software that we have seen since 2005.

I feel you are conflating software progress from algorithmic progress with software progress made possible by having large amounts of memory available.

I feel you are being deliberately disingenuous. Pre-NNUE Stockfishs would run perfectly fine on 2005 hardware. It would totally, utterly crush the best engines of that time. Even when giving those engines 1000:1 time odds. The progress in software is truly amazing. And yet here you are, denying reality.

Sure, Stockfish does not run on a 6502. Since when is a 6502 the standard by which everything software-related must be measured?

[towforce]

Sure, Stockfish does not run on a 6502. Since when is a 6502 the standard by which everything software-related must be measured?

The 6502 is important in computer chess because it was the cheapest of the 8-bit CPUs at the start of the golden age of dedicated chess computers. It's a really bad choice of CPU today, though - whatever you're building.

What Mclane is REALLY impressed with is that the authors from the golden age got such good chess from 8-bit CPUs.

How about... reverse compiling these programs to a high level language so that they can run on today's computers? The resulting programs would be poor, though, because they were written to the constraints of 8-bit CPUs.

[towforce]

Yes Thorsten, but it can be explained with the law of dimishing returns towards perfect play chess engines:

https://en.wikipedia.org/wiki/Diminishing_returns

The closer we get to perfect play, despite more hardware as input, we get less Elo as output.

--
Srdja

This is a good point: the jump to 2000 elo was a lot easier than the jump to 3000 elo - just as getting to the 2 metre high jump was easier than getting to the current WR of 2.45 metres.

[chrisw]

Current hardware would mean 1 cpu since the older chess engines cannot run more then one core.
In a uniform platform tournament, so same hardware, one can try to find out.

I overlooked your reply.
Yes, single core Fruit 2.1 vs single core Stockfish, last release.
Fruit will be crushed.
Do you doubt this?

Already done:

Thanks! And, wow... Fruit 2.1 from June 2005 against single-core pre-NNUE Stockfish from 2020, both on 2020 hardware, with 1000:1 time odds for Fruit 2.1, and Fruit 2.1 got crushed.

And Fruit 2.1 was not a weak engine at all!
One could still "blame" Fruit 2.1 though, because its release as open source certainly contributed to the progress in software that we have seen since 2005.

I feel you are conflating software progress from algorithmic progress with software progress made possible by having large amounts of memory available.

I feel you are being deliberately disingenuous. Pre-NNUE Stockfishs would run perfectly fine on 2005 hardware. It would totally, utterly crush the best engines of that time. Even when giving those engines 1000:1 time odds. The progress in software is truly amazing. And yet here you are, denying reality.

Sure, Stockfish does not run on a 6502. Since when is a 6502 the standard by which everything software-related must be measured?

I fell you are deliberately ignoring RAM availability and the qualitative effect of massive speed. Not just faster, but new effects emerging (eg working up position knowledge through fast search, as opposed to trying to calculate mods to PST tables).

Sure, I agree with you, the comparison to 6502 is nonsensical. They were unable to use multiply and forced into imprecision of 8 bits. The programs needed fitting to the very limited instruction set. Better to compare things written in relatively simple C, certainly pre 2005 with a non-NNUE SF forced to the constraints of 64K useable RAM for tables. SF will win probably, there’s been some software progress, but the bulk of the progress is hardware enabled, it wasn’t done before because it could NOT physically be done before.

If this is too difficult for you, consider your own field. Imagine a future 10 man EGTB. Then add massive more RAM availability and build an 11 man table. Is that hardware or software or hardware availability progress?

[mclane]

There is e.g. a 6502 C compiler called cc65:

https://www.cc65.org/

[syzygy]

I fell you are deliberately ignoring RAM availability and the qualitative effect of massive speed. Not just faster, but new effects emerging (eg working up position knowledge through fast search, as opposed to trying to calculate mods to PST tables).

Single-core speed went up but not massively. From 2005 to 2020 probably less than 10x.
In 2005 I had 2GB of RAM. 512MB should be more than plenty for pre-NNUE SF, certainly in ultrabullet games (from the point of view of SF) against Fruit 2.1. I don't know if it is still escaping your attention that pre-NNUE SF crushes Fruit 2.1 at 1:1000 time odds on equal hardware?

(I think I wrote SF 15 earlier, but the 2020 test reported in the post was with SF-dev from April 29, 2020, i.e. a pre-NNUE development version between SF 11 and SF 12.)

Sure, I agree with you, the comparison to 6502 is nonsensical. They were unable to use multiply and forced into imprecision of 8 bits. The programs needed fitting to the very limited instruction set. Better to compare things written in relatively simple C, certainly pre 2005 with a non-NNUE SF forced to the constraints of 64K useable RAM for tables. SF will win probably, there’s been some software progress, but the bulk of the progress is hardware enabled, it wasn’t done before because it could NOT physically be done before.

64KB pre-2005? This is not a serious discussion, to put it friendly.

If this is too difficult for you, consider your own field. Imagine a future 10 man EGTB. Then add massive more RAM availability and build an 11 man table. Is that hardware or software or hardware availability progress?

This is not difficult at all. Progress in software has been absolutely massive. You can just run pre-NNUE SF on an old machine and see for yourself.

Of course this has been "enabled" by hardware, in particular:
1. The internet.
2. The possibility to run many fast games to evaluate changes.

Again, nobody here is arguing that programmers today are somehow more talented. Nobody is taking anything away from what the 1980s programmers achieved. Do not worry.

But that does not mean we should remain stuck in the 1980s, nor that the 1980s are the measure of everything. Why the 1980s and not the 1950s?

The point is simply that there has been undeniable, spectacular progress in software.
But of course on Talkchess you will find people denying it anyway.