SlowChess Blitz Classic 2.0

[Frank Quisinsky]

Hi Jonathan,

if you have it on your list and thinking on it ... super!
A little future project ...

I am thinking that in the next years a lot of interesting new chess-computers will be available.
The energy the chess programmers have will be reflected to new commercial ideas.

I know that companies from India and China have interest and thinking about concepts and good known companies from Germany and The Netherlands are still working. Very interesting private projects are still under development. Absolutly sure, a booster for chess computers in the next 2-3 years.

Enough ...

A strong start for Slow in FCP Tourney-2022.

The tourney will be stronger and stronger ... category 38 or 39 I believe.



Best
Frank

I am thinking on the year 1999.
Around 20 Winboard engines produced an Elo average from ~ 2250.
And on Dual Pentium 733Mhz this was a lot.

[pohl4711]

Slow Chess 2.8 is released.
Download on the Slow Chess Webpage

It scored +30 Elo to Slow 2.7 in a 12000 game gauntlet against similar strength engines (30s + .3s)
I expect rating list Elo difference will be somewhat less depending on conditions. (Also I keep making it slower since my test PCs are fast AVX2)

This time I did try to wait for release and make more improvements, but just ended up not releasing for 3 weeks and not making clear strength progress or working on any of the extras I wanted to. I'm sorta limited in time to spend and computing power, so am happy to just keep making some progress.

My testrun of Slow Chess 2.8 is finished (9000 games played vs 9 opponents) and the progress is above +30 Elo...very impressive. Slow Chess is now stronger than Ethereal and Koivisto!

https://www.sp-cc.de/

[jstanback]

Roughly 2/3rds of the 3.5 MB of neural net data in Slow 2.8 is endgame nets, so endgame, especially late endgame is a strong point.
I mainly spend time training/testing late endgame because it's much more manageable a test-bed than the full game. I only have one computer part-time to test on. Starting in one-piece (per-side) endgames Slow 2.8 was slightly stronger than even Stockfish 12, but from early/mid end-game I still see so many games slip away to losses against the very top.

The sure win endgame values keep getting higher with training, which means attacking pressure is often converted to an obviously won endgame, which looks quite human like, but at some point I want to figure out a way to concentrate training on mating attacks. (Also as the opposition level increases becomes rare to see quick steam-roll games, but very satisfying when they happen.)

I'll add Raspberry Pi compile to list as something to look into later. I've been a bit busy to look too much at chess stuff, luckily the machine learning itself usually eventually shows enough improvement to keep me interesting in trying to make a new version.

I need to work on how to handle endgames. I guess maybe I should start by experimenting with a NN for something like simple KRK endings and maybe generate a bunch of random positions using the TB score for the target and see if the net trains up so that it can mate with only a shallow search. It would be interesting to see what size net and how many training positions it takes to get this working.

John

[Cornfed]

I need to work on how to handle endgames. I guess maybe I should start by experimenting with a NN for something like simple KRK endings and maybe generate a bunch of random positions using the TB score for the target and see if the net trains up so that it can mate with only a shallow search. It would be interesting to see what size net and how many training positions it takes to get this working.

John

You personally (?) need to work on how to handle endgames?

Simple - don't look to an engine, get a good (even basic - don't know your level) book on endgames and read it. There are many which offer specific advise and plenty 'rules of thumb'. Then play a lot...don't take draws and actually try to put what you have read into practice.

As I recall that is what GM Tony Miles once said he did back in the day and it's a tried and true method.

[Madeleine Birchfield]

I need to work on how to handle endgames. I guess maybe I should start by experimenting with a NN for something like simple KRK endings and maybe generate a bunch of random positions using the TB score for the target and see if the net trains up so that it can mate with only a shallow search. It would be interesting to see what size net and how many training positions it takes to get this working.

John

You personally (?) need to work on how to handle endgames?

Simple - don't look to an engine, get a good (even basic - don't know your level) book on endgames and read it. There are many which offer specific advise and plenty 'rules of thumb'. Then play a lot...don't take draws and actually try to put what you have read into practice.

As I recall that is what GM Tony Miles once said he did back in the day and it's a tried and true method.

jstanback is the author of Wasp 5.00, and somebody had just ran an engine match of 10 games between Wasp 5.00 and Slowchess 2.8 earlier on this thread, where Slowchess decimated Wasp 7.5-2.5, 5 wins and 5 draws for Slowchess. Then the Slowchess dev himself mentioned in the thread that Slowchess is very strong at endgames due to his focus on improving the endgame evaluation.

[CMCanavessi]

Is there a way to force the engine to open in uci (or console) mode and not go directly to GUI? Something like -uci or whatever.

[jonkr]

CMCanavessi :

There is no console mode built into the windows version. To create the standard console is only automatic when built for it, so never bothered. The Linux version is compiled for console-mode only though.

John :

I tried once make a net that learned endgame checkmates, but it was still with my usual training code ( positions from test games, values from W/L/D outcome 0, 0.5, 1 with optional partial search re-scoring ) only with a rather small distance-to-mate built into win values. It wasn't particularly effective at learning mate distance, it failed to win BN vs K games even, so I just left my old mating eval function. I think if I set it up as something new trying to learn distance it would have worked fine but been harder to setup/integrate. What I started with was Rook v Rook endgames with any number of pawns on the board, and now that my auto-trainer and eval is better set-up, it takes just a few hours of very fast games to start beating my old hard-coded eval in a specific endgame like that.

[RubiChess]

CMCanavessi :

There is no console mode built into the windows version. To create the standard console is only automatic when built for it, so never bothered. The Linux version is compiled for console-mode only though.

So the better question seems to be (and I also had this question when I tested with SlowChess): How does a GUI like Arena manage to start the slow.exe without displaying its GUI but instead get access to its stdin and stdout streams. And how can a Windows user do this with Windows tools? Maybe not possible.

Regards, Andreas

[jonkr]

CMCanavessi :

There is no console mode built into the windows version. To create the standard console is only automatic when built for it, so never bothered. The Linux version is compiled for console-mode only though.

So the better question seems to be (and I also had this question when I tested with SlowChess): How does a GUI like Arena manage to start the slow.exe without displaying its GUI but instead get access to its stdin and stdout streams. And how can a Windows user do this with Windows tools? Maybe not possible.

Regards, Andreas

On startup Slow checks if there is a pipe setup with fgets on stdin, if there is, it doesn't call CreateWindow and setup the own GUI window. So if an external UI (eg. Arena) supports sending commands to engines you could interact via uci commands that way. (Without a GUI communicating there would be no way to interact with the engine if it didn't show the own GUI).

[CMCanavessi]

CMCanavessi :

There is no console mode built into the windows version. To create the standard console is only automatic when built for it, so never bothered. The Linux version is compiled for console-mode only though.

So the better question seems to be (and I also had this question when I tested with SlowChess): How does a GUI like Arena manage to start the slow.exe without displaying its GUI but instead get access to its stdin and stdout streams. And how can a Windows user do this with Windows tools? Maybe not possible.

Regards, Andreas

On startup Slow checks if there is a pipe setup with fgets on stdin, if there is, it doesn't call CreateWindow and setup the own GUI window. So if an external UI (eg. Arena) supports sending commands to engines you could interact via uci commands that way. (Without a GUI communicating there would be no way to interact with the engine if it didn't show the own GUI).

Thanx for the explanation. So here's request time x 2

a) can you add an option to start it in console or uci mode? -c or -uci or whatever
b) when the engine loads, i see the following

info string 9 bitbases loaded.
info string 14 neural nets loaded.
info string 34237 opening positions loaded.
readyok

it will show "9 bitbases loaded" wether I use syzygy or not, which is fine because the bitbases file is in the engine's directory. But it would be nicer (maybe) if when using syzygy, it would print another message like "xxx syzygy bases loaded" or something like that, just to be sure they are being correctly found and loaded

And another question, if both bitbases and syzygy are loaded, does it use both? if not, which ones are used? and if yes, when does the engine use one or the other?