Author: Robert Hyatt
Date: 14:10:54 06/19/03
Go up one level in this thread
On June 19, 2003 at 08:11:07, Steve Coladonato wrote: >On June 18, 2003 at 20:12:32, Robert Hyatt wrote: > >>On June 18, 2003 at 11:30:30, Steve Coladonato wrote: >> >>>Dr. Hyatt, >>> >>>Any chance of Crafty doing backward analysis "instead of/in addition to" forward >>>analysis? I use the chess engines mainly for game analysis, and overall Crafty >>>tends to give lines that are not sound compared to other engines. I understand >>>the reasoning behind forward analysis in that it tends to give more consistent >>>results. But starting at the end of the game, where the outcome is generally >>>known, seems to make sense if you want to cache favorable vs. unfavorable lines. >>> >>>Thanks. >>> >>>Steve >> > >> >>I started out doing it backward. But I (and others) didn't like the >>results as it would inconsistently find good moves when the hash didn't >>get clobbered, but there is no guarantee. >> >>I'll look to see if the old code is still around anywhere but I doubt >>it. > >Thanks. > >Steve Lim's suggestion of making it one of the options doesn't sound bad. > >Has the technology improved to the point where the original inconsistencies >would no longer be an issue? > >Steve No... here is the problem. With the current approach (front to back) if you give it (crafty) a minute per move, you will get consistent N-ply evaluations. IE if it can go 12 plies deep, it will see what it should see within 12 plies, on every move, and the comments it gives you will be from that perspective. If (as it did early on) it goes back to front, then you carry info back from the endpoints toward the beginning, and you can see some things earlier in the analysis as a result. Until the hash table can't hold such important stuff, then suddenly the score drops or rises, and it has _nothing_ to do with the search engine and the tree, but it changes due to an important hash entry getting overwritten. Now when you go over the annotated output, you see a comment that doesn't make sense within a N ply search, because it is based on a much deeper search due to the hash entry, but at some random point that hash information gets clobbered, and the score drops (or rises) as a result. It generated lots of questions and complaints, even though I originally thought it was an interesting approach. But now, if a 12 ply search should see something, it does. And it won't randomly act like a 20+ ply search due to old hash values...
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.