Author: Dann Corbit
Date: 13:36:47 01/13/05
Go up one level in this thread
On January 13, 2005 at 14:11:32, Dann Corbit wrote: >On January 12, 2005 at 19:18:13, Uri Blass wrote: > >>On January 12, 2005 at 18:37:25, Dann Corbit wrote: >> >>>On January 12, 2005 at 18:27:11, Uri Blass wrote: >>> >>>>On January 12, 2005 at 14:26:59, Dann Corbit wrote: >>>> >>>>>On January 12, 2005 at 14:02:53, Mridul Muralidharan wrote: >>>>> >>>>>>On January 12, 2005 at 13:01:29, Drexel,Michael wrote: >>>>>> >>>>>>>On January 12, 2005 at 12:42:05, Anthony Cozzie wrote: >>>>>>> >>>>>>>>So let me see if I understand this conversation correctly. >>>>>>>> >>>>>>>>1. I state that the 6 man tables are worth 100 elo >>>>>>> >>>>>>>I thought you were joking, but obviously I was wrong. >>>>>>> >>>>>> >>>>>>I have no quantitative way of accurately guessing this too - but "depends on >>>>>>program" maynot be a wrong statement ? >>>>>>And both are definitely agreeing that there is a non-trivial improvement in >>>>>>performance - right ? Then why disagree for the sake of disagreeing !!!! >>>>>> >>>>>> >>>>>>>> >>>>>>>>2. You disagree, and state they are worth 50 elo >>>>>>>> >>>>>>>>3. You do this by pulling numbers out of your *** >>>>>>>> >>>>>>>>4. Since the full 6-man set hasn't been generated, and the elo gain is almost >>>>>>>>certainly different for different programs, we are both guessing. >>>>>>> >>>>>>>Yes, but in this case Uri's guess is much more educated. >>>>>> >>>>>>Hmm , I dont see how - just 'cos there was a "women" reference ? :) >>>>>>Jokes apart - the point to be taken is - they could be a SIGNIFICANT improvement >>>>>>: and would be the world of difference between a loss and a draw (or a draw and >>>>>>a win). >>>>>>Depends on how you eval , and what you do in your search (extensions and qsearch >>>>>>/threat detection). >>>>>>Ofcourse , if you have a junk endgame eval with quiet decent middle game eval - >>>>>>your improvement can be much higher than what both of them quote ! >>>>>> >>>>>>But are we not quibbling over nitty gritty details ?? 50 , 100 , 125 - what does >>>>>>it matter : it would be a substantial improvement !!! >>>>> >>>>>In 20 years, we might be able to memory map the whole 6 man set. >>>>>That would yield a stupendous Elo increase for endgames. >>>> >>>>in 20 years computers will be very fast. >>>> >>>>faster computer mean less blunders without tablebases and mean that tablebases >>>>are less important. >>>> >>>>I suspect that if you wait 20 years more than 90% of the comp-comp games in the >>>>high level will be drawn even without the 6 piece set and 6 piece set will have >>>>smaller influence relative to the influence that it has today. >>> >>>If you memory map the tablebase files, there is almost zero cost for a probe. >>>It may be worthwhile someday also to memory map 7 man bitbase files. With a 64 >>>bit CPU, 18,446,744,073,709,551,616 is the address space (per CPU). >>> >>>An oracle that is 100% certain to have absolutely correct information and which >>>has almost zero cost for the info lookup will be a superior solution. >>> >>>The addition of code to compute endgames will necessarily slow down evaluation >>>and complicate the code base. >>> >>>If the average user had 16 GB Ram systems now (and even now such a system can be >>>purchased for about $12,000) then the existing 5 man tablebase files could all >>>be memory mapped. >>> >>>I predict that in this case, there will be a very large Elo boost, even for the >>>5 man tables. I base this upon the substantial Elo boost which has been >>>measured for bitbase files, which provide inferior information to that of a >>>tablebase. The benefit of a bitbase file is that it can be held in memory much >>>more easily. >> >>What evidence do you have for substantial Elo boost for bitbase files. >>What time control is used? > >I have seen the results of several experiments (at least two of them on the >Winboard forum were discussed) and also have done my own experiments. I have >not seen evidence of substantial gain for tablebase files, but there is clear >evidence for bitbase files. > >You can easily perform your own experiments with patzer and yace. You will see >benefit at every time control. I should mention that my bitbase experiments also had tablebase files activated. I do not know if this is important or not, and probably experiments should be also run with bitbase files only.
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.