Author: Eugene Nalimov
Date: 16:41:10 09/19/99
Go up one level in this thread
Similiar approach was used by Jonathan Schaeffer, when he generated checkers endgames for Chinook. I thought about that when I wrote my generator, but decided to use simpler schema that was faster to write and debug. I had all the necessary horsepower for it. Unfortunately, general pawnless 6-man TB will not fit into 4Gb of 32-bit address space, so you'll have to use 64-bit integers for indices and bit vector approach when generating TBs. That can be done, but again - you'll have to write special and sometimes tricky code. I'd prefer to wait a year or two and then use 64-bit system. Eugene On September 19, 1999 at 12:37:06, Peter Karrer wrote: >To reduce the amount of memory needed for generating 6-man tablebases with >pawns, couldn't we try to produce them incrementally? > >First, the tablebases with a pawn fixed on file 7 would be generated, say >KxPa7Kyy...KxPh7Kyy. To generate for instance KxPb6Kyy, we would need only >KxPa7Kyy, KxPb7Kyy and KxPc7Kyy and so on. > >One could devise the indexing scheme in such a way that merging these partial >TBs would be a simple matter of concatenation. > >Memory needed for this approach would be more than for 5-man tablebases, but >most likely below the treshold where you need 64-bit addressing. > >Regards, Peter > >On September 17, 1999 at 12:47:03, Eugene Nalimov wrote: > >>On a 256Mb machine you must be able to generate all 5-man pawnless TBs. Maybe >>you'll be able to generate TBs with pawns, but here I have some doubts - it >>looks that paging will slow it down too much. In any case you can try, and if >>one iteration will took less than (say) two hours you'll be able to do that. >> >>I beleive that you can use compressed TBs during generation; just don't use >>memory-mapping files (-p), use TB caching (-c <cache amount>). Of course it'll >>be slightly slower, but not much. >> >>Theoretically it's possible to modify the generator so it'll produce compressed >>TBs, but I hate to do that. You can try if you wish :-) >> >>Eugene >> >>On September 17, 1999 at 10:54:09, Owen Lyne wrote: >> >>>Continuing to set iup my new Athlon with Crafty, tablebases, Winboard etc. So, >>>I'm at the EGTB stage right now. Thought people might be interested in >>>speed of generation - did all the 3/4 piece tablebases in about >>>90 minutes (Eugene quotes a PII/400 taking about 2 hours, so that >>>sounds about right). >>> >>>Next though is 5 piece, I have 256MB of RAM but Eugene likes >>>machines with 500+, so I may have a problem... >>> >>>Which bases (perhaps pwnless ones?) can I do and which are totally >>>impossible with the code set up as it is? Can I try anything and >>>just put up wirh ridiculous paging to disk and slow performance, >>>or is it not even possible? >>> >>>Other thought is about compression - being able to run Crafty >>>using compressed tablebases is absolutely awesome, and without taking >>>a performance hit is astonishing! Great work Bob, Eugene and Andrew Kadatch! >>> >>>Now - how about generating tablebases, is it possible (either with current code, >>>or to modify code so it is) to generate tablebases where the minors >>>are compressed? That would be incredibly useful, especially again if it >>>was without much performance hit. >>> >>>It's all very well saying compressed only take 6 gigabytes (I have plenty) but >>>If I have to have all 22 gigabytes uncompressed while making them before I can >>>compress any then ouch, I don't think I can manage... I presume I could figure >>>out which minors are used to generate any one base and carefully compress and >>>uncompress accordingly, but directly using compressed minors would be ideal. >>> >>>One step further, generate in compressed form (I know that's probably even more >>>ridiculous, but would hugely help those operating in limited disc space). Of >>>course I know the other thing someone with limited space can do is >>>download not generate them, but when I got 3 or 4 KB/s to Bob's FTP site >>>I knew that wasn't going to be any use for me... Not having to pay >>>the phone bill anyway... From work I get much, much faster FTP, but >>>via my ISP (Freeserve) and modem looks much slower (maybe time of day >>>was also relevant, early morning UK time much better than evening UK time >>>since thg latter is daytime in US). >>> >>>Owen
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.