Author: Vasik Rajlich

Date: 08:44:59 06/21/05

Go up one level in this thread

On June 21, 2005 at 07:13:12, rasjid chan wrote: >On June 21, 2005 at 04:47:27, Vasik Rajlich wrote: > >>On June 20, 2005 at 15:18:05, rasjid chan wrote: >> >>> >>>I have implemented hashing 2 bounds after reading fruit does this. >>>It is fairly complicated to do and I just want to see if it is done wrongly. >>> >>>I measured the figure >>> hash-hit-return-by-2nd-bound/total-hash-hit >>> >>>Per game move it is 0.000 almost all the time. >>>Ocasionally it has 0.001 and almost rarely 0.002. >>> >>>If others have much higher figure then this,I suspect my double-bound is done >>>badly. It is difficult to debug except through usual hit statistic. >>> >>>Vasic, >>>If this 2nd bound statistic is correct, then fail-soft propagation >>>of exactness is far better. For full-search FL exactness,it starts from 2/3% >>>and drops to about 1% for whole game. >>> >> >>I don't understand this sentence. >> >>You'll have exactness when your scores are inside the windows. >> >>You'll also have exactness in some very rare cases very close to the horizon, as >>we discussed in the thread below. (This applies only to your custom algorithm.) >> > >I now have statisic permanent (better) in my log output. >The figure mentioned above is just 1 game(that's me). > >I measure only for full search fail-low and exact(w/o the algorithm,always >upper) and NOT exact when scores within windows; > > x% = FULL_SEARCH_FL_EXACT/TOTAL_FULL_SEARCH_FL x 100.0; > >I summed nominator and denominator cummulatively as game proceeds. >For the whole game, it may vary from 2 - 6%. > Aha, I understand. Keep in mind that these 2-6% will always be either at the horizon or very close to it. If you want, you can take statistics on this. Collect one set of numbers for depth >= 4, and another for depth < 4. You'll find that for depth >= 4, the figure is around 0.00%. >This statistic I think is better then expected and significant. >Of course none can expect something like 10-15%. > >If nullmove adds 150-200 ELO points as some say, then this (beefed-up fail soft) >may add 15-20 points and is free; > Sorry, but it will be much less. By the way, I do like your idea. It's rigorous and correct - the only cost is a few more calculations at each node. A good place to start is to see how many nodes you save. Search some 20 positions to 10 ply with your enhancements, and search the same positions to the same depth without them. If your idea has practical benefits, the first set of searches will need fewer nodes than the second set of searches. By the way - around 1 year ago or so I put in some code to collect a huge number of statistics about the search. I have a bunch of files on my lap top which I'll have in my hands again in a week or so. If you want, I can post these. Vas >Best Regards >Rasjid > > >>Best regards, >>Vas >> >>>Best Regards >>>Rasjid

This page took 0.01 seconds to execute

Last modified: Thu, 07 Jul 11 08:48:38 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.