Author: Robert Hyatt
Date: 20:11:32 10/24/99
Go up one level in this thread
On October 24, 1999 at 02:39:41, Pete R. wrote: >On October 23, 1999 at 21:02:03, Robert Hyatt wrote: > >>On October 23, 1999 at 19:03:55, James T. Walker wrote: >> >>>Hiarcs has a learn file which is limited to 64K in size. Rebel has an 8k file >>>which appears to be for position learning. Crafty seems unlimited in this area. >>>Q. What is the limiting factor on the size of a "Position learning" file? >>> >> >>Crafty is actually limited to 64K entries. I do this because these are >>sucked into the hash table and marked as 'permanent'. I don't want a bunch >>of those things totally clogging the hash table... >> >> >>>All the Chessbase engines use a "Retro-Analysis" to analyze the games/files >>>starting at the end and working to the beginning. This seems like an excellent >>>way to find mistakes. >> >>retro-analysis is good and bad. It can be bad, because while the entries >>stay in the hash, as you back up move by move, you get better and better >>analysis as you work back. But eventually those older positions get over- >>written, and suddenly the eval makes a big change, and this big change has >>nothing to do with the game, rather it is a hashing issue. I tried this when >>I first did annotate(), but decided that the fluctuations that occur when the >>hash table gets overwritten were more confusing than the deeper analysis was >>useful... >> >> >> >> >> >>>Q. Why can't this be used for the "Position learning" files? >>> >>>Humans can examine games and find "mistakes" made by programs. >>>Q. Why can't humans contribute to the position learning files in the manner? >>> >> >>I am going to do this... someone asked for such a feature for the kasparov >>vs world game... they want to get into analysis mode, but be able to say >>"this position is +9 no matter what you think" and have that as a permanent >>(for the duration of the analysis session) hash entry... >> >>That is doable, in the same context as current position learning. > >That guy that pestered you about this would be me. ;) I think it would have >really helped the game since Crafty was the weapon of choice for most World team >players, including IMs and probably a number of GMs. Meanwhile Kasparov was >apparently using Deep Junior (also on a quad Xeon), so he had us outgunned in >the computer arena as well. That and being able to look at our analysis, so all >in all it's a miracle the World did so well given all the handicaps. A Crafty >that could be selectively trained in critical positions and variations would >have been an awesome tool, easily better than Deep Junior and anything else less >powerful than DB, but in the end the voting format of the game came home to >roost and it ended on a blunder. The world will be much stronger next time. ;) >And there will probably be raging post-analysis debate since Kasparov now has to >back up the claim that he had a win without the final blunder, so we're not done >analyzing this game. Anyway even though the game is over this feature is no >less valuable, it will be great for all serious analysis, and I think it will be >really appreciated by players everywhere. In fact I predict that eventually it >will be standard on all top commercial chess programs, but once again Crafty >will be first. :) E.g., Fritz 6 just got EGTB probing in the search, so right >around Fritz 8 I expect to see this type of analysis learning on the feature >list. ;) Thanks again for adding it to your list. It isn't really on my list. It is in 17.0... and seems to work fine. It works like this: you enter analyze mode, and step back and forth as always. At any point, you can type "store <val>". It saves the current position and the score you supply in the 'position learning' file. Which becomes a permanent part of the hash table also. "val" can be any number you think appropriate, such as -9.0 (queen down) or +9.00 (queen up) or anything else you like. If you do this in a separate directory from your normal play, the position.bin file won't cause odd things to be stored and used during real games... And if you want to clear it out, just delete position.bin and away you go with no stored stuff. Seems to work well in the testing I did...
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.