Author: Ingo Lindam
Date: 00:43:00 10/28/02
Go up one level in this thread
On October 27, 2002 at 20:56:58, Robert Hyatt wrote: >On October 27, 2002 at 17:14:22, Ingo Lindam wrote: > >>Hello... >> >>I just thought about some questions from Bob that lead me to a problem. >>I know the engines should have as much knowledge as the programmer can give to >>it. But ofcourse in may searchtree may appear some positions I can't apply a lot >>(or in worst case any) of that knowledge: material is equal and all the other >>knowledge doesn't fit to current position (in the tree). Will my score 0 (or >>atleast very near to 0)? >> >>If yes...I just ask me whether this could lead the computer to optimize the best >>move sequence into complete knowing nothing about the position the sequence of >>moves end in... as more as I evaluate the positions for my opponent the same >>way? >> >>Shouldn't I better substract something from my score and add something on the >>opponents score in case I know nothing about the position? That was not the solution... that was just a question that should lead in the right direction. As you no I would prefer to have more than one score and strictly distinguish between equal and knowing nothing/too less. >Any such idea is very dangerous. It is amazing how a full-width search to >reasonable depth will mangle your evaluation if you have a big hole in it. >In the example above, the program is going to try to force such positions that >give it a slightly positive score, even if it is wrong. No... I don't want to force the computer going into positions it doen't understand. I want to avoid him to do so. >You basically need to >make sure your eval is covering all the positional things you are aware of, and >then watch as it makes mistakes and continue to add knowledge to fill in each >and every hole. Yes, that's my oppinion...you should fill as much knowledge into that holes that you can get. >If you leave one hole, the search will force the tree into that >"no-mans-land" and cause problems... Yes, that was the problem I talked about. And I guess you have to devide this problem into two different cases: a) computer is playing against someone else b) computer is just used to analyse a position/game In case a) I should be aware of not punishing the unknown (not able to evaluate them right) aspects of a position OR punishing them from the computers AND the opponents view the same way could lead into optimizing the main variation into the wrong direction... into no-mans-land (that might be opponents land, because I have to assume him to have different knowledge, in particular if opponent is human) In case b) computer should be fair and tell me this variation is rather unclear to the computer than it is equal... and ofcourse I would be dissapointed if main variation very often would lead into unclear position. So far I understand that this problem is not seen commonly with same eyes and solved by a common solution in every program the same way. Atleast I think most of you would admit that the chess knowledge of todays engines is not without any holes and having a lot of holes and ignoring them could lead to an optimization towards that holes. Am I right? Ingo
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.