Author: Robert Hyatt
Date: 20:00:02 11/21/00
Go up one level in this thread
On November 21, 2000 at 17:38:58, Uri Blass wrote: >On November 21, 2000 at 14:48:25, Robert Hyatt wrote: > >>On November 21, 2000 at 14:32:51, ERIQ wrote: >> >>>or is either better ?? your thoughts. >>> >>>I was thinking about buying yet another program for analyses of games. >>>Though I like chess tigers playing ablity I wonder it fritz with it's >>>backward way of doing it and ability to add eval's like (!,!?)automaticaly >>>would be better for just looking at game scores. >> >> >>I don't like back to front. It requires that stuff from the end of the >>game be back-filled through the hash table. Here are the two arguments: >> >>1. back to front is better... because you can use stuff happening at the end >>of the game to influence scores and moves earlier in the game. >> >>2. front to back is better because it gives a true reflection of what the >>computer would have seen had _it_ been playing that game directly. >> >>I go along with 2. 1 means you see some bad things quicker than if you use >>2. But it also means that when the critical hash entries get overwritten, >>the scores jump up significantly, which is also misleading. > >I think that the way of learning is not optimal for analysis. >When I analyze a position I want an intelligent program to learn from the tree >that I generate in the same way that I can do it. > >When I have one forced line there is no problem but my experience is that when >there is a tree of 2 lines of some moves I can get the following behaviour. > >1)I analyze line A with a chess program and go back amd the program understand >that line A is not good and prefer line B. > >2)I analyze line B with the program for a long time and go back. >The program understand that line B is not good but instead of prefering line C >it prefers again line A because it forgot that line A is not good. > >I think that it is better not to forget only the positions in the tree that I >generated with the evaluations of them after search and to forget all the other >stuff. The analysis feature of crafty will let you get around this. Once you find that a move is bad, you can store that entry permanently in hash (permanent means until you exit the program, not forever). Then when you reach that position anywhere in the search, the score you suggest will be plugged in... This has been in a while to avoid the problem you mention. > >The tree that I generate is not too big to remember and the problem is that the >programs that I use do not understand that it is the important tree to learn >from it. > >Uri
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.