Arena 3 gaviotaTB vs robbotripleTB

[BorgKing]

I've recently started using ComStock 3, but just noticed that while Comstock seems to be checking everything with the RobboTripleTB (literally millions of hits), Houdini suddenly doesn't seem able to access its GTB "0 TB hits" every time I've watched it at similar depth as CS.

And when running Comstock 3 vs SF 211 GTB, the gaviota hits are only 30-40 while CS 3 still has millions.

Is this a result of the Robbobases taking up too much RAM (I only have 4GB on an old double processored laptop), so the other engine doesn't reach its TB? Or are the default TB setting really so different between these three engines? And if so, what settings would moderate ComStock's TB use, without losing playing strength?

Thanks for any help or advice!

[velmarin]

Possibly just a matter of counting.
Where you put the probe and its formula.

Also with RTB Stockfish can be seen, there is a difference, Comstock provides larger numbers.
But again it may be just a matter of where the probe is located.

And remember, do not give RobboTriplebases are Bitbases total distance, are smaller, faster, and come before the search.
Gaviota Tablebases are similar to Nalimov, give complete information but is more belated entry by search.
Please test with analysis in Arena
and compare.

[User923005]

The gaviota egtb files also have a bitbase feature.

[ernest]

The gaviota egtb files also have a bitbase feature.

Can you explain ???

[velmarin]

That, explain, explain.???

[User923005]

Code: Select all

	/*--------------------------------------*\
	|
	|      	PROBING TBs #4 
	|		(HARD, only win, draw, lose)
	|   
	\*--------------------------------------*/

	/* 
		Probing with the WDL versions of the probing functions
		will return only the info needed to know whether a position
		is a win, draw, or a loss.  
		The Gaviota tablebase library will try to return this info
		with the best performance possible. If the only info needed for
		a position is WDL, this function should be used rather
		than the regular tb_probe_hard() function.
		This function would be the "equivalent" of one that probes a bitbase.
	*/

	tb_available = tb_probe_WDL_hard (stm, epsquare, castling, ws, bs, wp, bp, &info);

	/* print info */
	wdl_print (stm, tb_available, info);

[ernest]

Code: Select all

	/*--------------------------------------*\
	|
	|      	PROBING TBs #4 
	|		(HARD, only win, draw, lose)
	|   
	\*--------------------------------------*/

	/* 

OK thanks!
Do you have a link where this has actually been used, or tested?

[syzygy]

The gaviota egtb files also have a bitbase feature.

Can you explain ???

Most likely this means that cached decompressed blocks store 4 or 5 positions per byte (with W/D/L results) instead of 1 position per byte (with DTM). So it makes caching more effective. A look at the source tells me 4 positions are stored per byte.

It's an improvement over caching decompressed blocks with 1 position per byte, but still far from what is possible with a compression scheme that allows caching compressed data in RAM. The uncompressed gaviota tablebases are 38.5 GB, so with 4 positions per byte more than 9.6 GB of RAM is needed to cache all of them. My tables need 378 MB. The Robbo triplebases need 570 MB it seems.

That the number of probes shown by ComStock is so high suggests that it probes the tables even in the qsearch. Indeed this patch adds a probe to the qsearch. (I don't know if that is the latest code, but I would suggest to whoever maintains it to probe the tables only AFTER probing the TT table.)

[User923005]

Code: Select all

	/*--------------------------------------*\
	|
	|      	PROBING TBs #4 
	|		(HARD, only win, draw, lose)
	|   
	\*--------------------------------------*/

	/* 

OK thanks!
Do you have a link where this has actually been used, or tested?

They are fairly popular, though not all engines use this capability.

https://www.google.com/search?q=tb_prob ... =firefox-a

[ernest]

They are fairly popular, though not all engines use this capability.

Sorry, I don't understand.
Bitbases are of the form *.cmp (Shawul)
Gaviota bases are of the form *.gtb.cp4 and I don't see how they can "contain" bitbases. Do you have a link where the Gaviotas have actually been used as bitbases?