Critter 1.2 update + new homepage

[kingliveson]

Thanks Richard! I would like to ask you if you have thought about implementing learning for Critter, in the style of Shredder Learning? Currently Critter is my 3rd main engine behind the Rybkas, but I think that it could become the #1 choice for analysis if it had learning.

To be honest, I do not know how exactly Shredder Learning works. Is it like R3 Persistent Hash? While the idea sounds fairly simple, it is very difficult to implement correctly. I mean correctly form the user's perspective - to be really useful. With R3 it was a major pain to use and it was eventually removed in R4.

I will think about it, but can not promise anything yet.

Richard

Thanks for the release/update!

As for learning (an incorrect term in this case), it would require genetic programming, better known as self-modifying code. You are correct that it is more a preserve/persistent hash type feature. Search data is dumped and appended to disk for a unique positional analysis to obtain deeper search. The goal in essence is not learning, but resumption of analyses.

The book-learning feature in Fritz is closer to genetic programming where the books serve as population, and move settings evolve continuously based on statistical probability of 1, 0, and 1/2 results.

[User923005]

Result for Rybka is combined with earlier games against a different opponent.
Rybka / Critter is currently 7-31-28

27-103-96 Rybka v Critter currently on the same contest. (full machine resources, no ponder)

[Uly]

You are correct that it is more a preserve/persistent hash type feature. Search data is dumped and appended to disk for a unique positional analysis to obtain deeper search. The goal in essence is not learning, but resumption of analyses.

I see that learning is perhaps a bad term to use for it, but "Preserve Analysis/Hash" is also a bad term, because:

Preserve Analysis - Is a term that is used by Rybka 4, for a feature that discards everything when the engine is unloaded.

Preserve Hash - Is a term that is used by Naum, for a feature that discards everything when the engine is unloaded.

Persistent Hash - Is a bad term because you don't need to use a hash to store the analysis in hard disk, in fact, Shredder "Learning" uses a linear file to store analysis.

Probably "Store Analysis" or "Auto-save Analysis" or "Resume Analysis".

And after dealing with the "pain to use" Rybka 3 Persistent Hash, I'd be willing to deal with any Critter "Learning" that is a pain to use

Anyway, the thing with "Learning" is that you only store PV Moves and Internal PV Moves (anything where the score is know) in the file (not everything that is in the hash).

[Richard Vida]

You are correct that it is more a preserve/persistent hash type feature. Search data is dumped and appended to disk for a unique positional analysis to obtain deeper search. The goal in essence is not learning, but resumption of analyses.

I see that learning is perhaps a bad term to use for it, but "Preserve Analysis/Hash" is also a bad term, because:

Preserve Analysis - Is a term that is used by Rybka 4, for a feature that discards everything when the engine is unloaded.

Preserve Hash - Is a term that is used by Naum, for a feature that discards everything when the engine is unloaded.

Persistent Hash - Is a bad term because you don't need to use a hash to store the analysis in hard disk, in fact, Shredder "Learning" uses a linear file to store analysis.

Probably "Store Analysis" or "Auto-save Analysis" or "Resume Analysis".

And after dealing with the "pain to use" Rybka 3 Persistent Hash, I'd be willing to deal with any Critter "Learning" that is a pain to use

Anyway, the thing with "Learning" is that you only store PV Moves and Internal PV Moves (anything where the score is know) in the file (not everything that is in the hash).

Hm... you are very serious about this - let's call it "Analysis Save/Restore" feature. You are not the first who asked for some PH-like feature in Critter, but you are the first that seems to know about all the pros & cons. I am very curious about your ideas/needs, and I am more than willing to devote some development time to do something else than pure ELO rising. After all, we make chess engines to help us, not just to beat us

If you don't mind, I would rather discuss with you via PM/emails.

Richard

[ernest]

If you don't mind, I would rather discuss with you via PM/emails.

Beware, Uly is against private correspondence!...

[ThinkingALot]

Actually, IMHO, "Save Hash" and "Preserve Analysis" are exceptionally useful for correspondence chess... But the main hash is usually quite big (>= 1GB) so saving it and loading consumes some time and plenty of disk space. Therefore it would be nice to have smth like "Persistent Hash". Just another Hash but very small, say, 4Mb, and with a replacement strategy by depth only.

[Uly]

Beware, Uly is against private correspondence!...

No, I'm against making private conversations that were public, without a good reason. Richard probably has a good reason to keep this private, and he asked for it, so I agree.

I'm very happy about it, I remember when he made Critter score 2fold-repetition as draws, something that is very useful for games and speeds up analysis (so that on winning positions, Critter avoids repetitions, and in losing ones, Critter seeks them), I will try to help him as much as I can with Critter "learning".

[BB+]

Do we expect STS results (and with the new Komodo), or is Swaminathan more involved with other pursuits now?

[User923005]

Hi,

Critter 1.2 is available for win32/64, linux and android platforms.
It also got a new home: http://www.vlasak.biz/critter/

Thanks to Emil Vlasák for web hosting.

Richard

On I7 with 8 threads, no ponder, against Houdini 15a and time control of 2 min + 1 sec Fischer increment so far (it's early):
17-14-32 (small advantage for critter)

[User923005]

Houdini/Critter currently 39-44-100
For others Critter seems to score much lower for some reason. Perhaps it scales well with many threads