Author: Heiner Marxen
Date: 12:43:55 06/30/99
Go up one level in this thread
On June 28, 1999 at 20:18:18, Dann Corbit wrote: >On June 28, 1999 at 20:12:20, Dann Corbit wrote: >>On June 28, 1999 at 19:38:49, Heiner Marxen wrote: >>[snip] >>>>Imagin, I am at some node. Now it has been analyzed to 16 plies and the >>>>children (of which there are 40 possible) have only 14 entries. Of these >>>>two are at 15 plies and 5 at 12 plies and 7 at 9 plies. How can I update >>>>the parent from this? >>> >>>IMO, not at all. The 16-ply result is already build from 15-ply sub-results. >>>Even those cannot improve the 16-ply result. >>>You can look at this in a similar way as when you find a hash hit: smaller >>>depth than wanted is just not usable. (Exception: mate scores) >>I will have to explore this some. I have seen many examples where a parent at >>depth 16 does not see what the child sees at depth 15. Perhaps this is due to >>null move pruning. It may not be as ludicrous as it sounds because we are >>talking about billions of positions examined by this point. If some tiny >>fraction of the null moves pruned are wrong, they can point to a wrong result. >>Can anyone else think of an explanation as to why a 1 ply inferior child >>record should improve the parent result (perhaps a bug?) I do know that I >>have seen this happen time and time again. >>[snip] >Another thing along these lines that probably many of the posters here have >seen is when you are trying to solve a tough EPD position in a test suite. >You might let the computer run all night and it never finds the move. But >if you give the computer the position right after the correct choice, it >finds the answer in short order (much less than even exponential drop-off >would explain). I would call this [cough] a bug. For me, it is one of the most basic properties of minimax and alpha/beta that a search result is constructed from (less deep) sub-searches, i.e. is at least as good as any of the sub-searches. When related searches routinely produce conflicting results, we have a serious problem with their interpretation, especially when we try to combine them. You may try to develop a statistical model and operate with weighted averages... but that may produce more garbage than insight. I've not done that, so I'm guessing. But consider you have such a model, and do an experimental computation: how do you evaluate the result? Against what do you compare it to decide whether the approach is worth anything, when you doubt the soundness of all your data? You would need a well defined and trusted environment to test it. A deterministic and sensible result computation from conflicting basic data cannot be done, IMHO. If anyone can do better... I´d like to learn that I was wrong :-)
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.