Author: Robert Hyatt
Date: 10:46:56 08/06/99
Go up one level in this thread
On August 06, 1999 at 13:35:09, Terry Ripple wrote: >On August 06, 1999 at 08:51:20, Robert Hyatt wrote: > >>On August 06, 1999 at 00:33:02, Terry Ripple wrote: >> >>> I sometimes see accesses from a total of 7 to 9 pieces still on the board and >>>was wondering how accurate it`s analysis of the evaluation is and is the >>>analysis more accurate when it shows alot more accesses like several thousand >>>compared to 10 to 100 accesses or don`t this matter, as long as it`s showing >>>that it reached the endgame tablebases? Please note that i only installed the >>>four man tablebases in my Hiarcs7.32! >>> I`am using a AMD K6-2/266Mhz with 64 RAM and when playing games other than >>>blitz,i set the hashtables to 44.032KB. >>> >>>Thankyou in advance for any information, >>>Regards,Terry >> >> >>Think about it like this. At that stage of the game, your engine is probably >>searching at least 15 plies deep. For simplicity, think of the tree as a >>graph that is _exactly_ 15 plies deep along every line. With tablebase probes >>in the search, after any capture that takes the total pieces down to 4 or >>less, that branch goes no deeper and gets an _exact_ score at that point. >> >>However, your numbers are way off from what I see. I typically see the first >>EGTB hits (crafty, using 3-4-5 piece files) with 14-15 pieces still on the >>board (at blitz-type time controls). Hiarcs might be more conservative on how >>deep it can probe than Crafty is, I am not sure. But it makes a significant >>difference in the right positions... >--------- >Bob, > I noticed that when there were hits in the tablebases earlier in its search >and first showed only a hundred or so accesses with a particular score and best >move so far in the position and leaving it continue in its search for a few more >ply, the accesses increased dramatically to several thousand hits and the score >plus evaluation of the position also changed which then appeared to be a better >evaluation than early on, so if like you say that once the search starts hitting >the tablebases and these hits are an exact score at this point, then why does >the score and analysis of the position change as it progresses deeper into its >search? I was under the impression that once you start getting hits to the >tablebases, then that score and evaluation will be exact and won`t change its >score of that position any longer, even if you let the analysis continue deeper >because the endgame tablebases once reached are "exact". Help!!! > >Thanks alot for taking the time to explain this, >Best regards, >Terry Think about the tree. Suppose you are currently searching to depth=15. Most of the branches you search reach depth 15 (with more than 5 pieces left) and the program has to make its own evaluation of the position. A few of the branches don't make it to 15 plies because a capture takes them right into an endgame database where that branch gets assigned the _perfect_ score. When you go to depth 16 (next iteration) you have _more_ chances for branches to terminate with database hits, because you have one more ply to do a capture that might take you to 5 pieces. And when you get to 17, even more branches hit. What you are seeing is a 'stepwise-refinement' in action. Every additional ply reveals more exact scores due to database hits. If you can search deep enough so that _every_ branch ends with a hit, you get a perfect score back to the root and you need search no further... I see hits starting with 14-15 pieces still left on the board. By the time I get to 8-9 the disk is buzzing...
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.