Author: Amir Ban
Date: 08:05:21 02/28/03
Go up one level in this thread
On February 28, 2003 at 03:11:41, Uri Blass wrote: >On February 27, 2003 at 20:23:39, Amir Ban wrote: > >>On February 26, 2003 at 18:48:53, Stephen Ham wrote: >> >>>Thanks, >>> >>>Dann, Uri, and Robin! I think I understood all of that. That was very helpful. >>>Again, I'm know next to nothing about chess engine programing, so I assumed that >>>the evaluation function guided the search function. I don't know why I assumend >>>that...I just did. >>> >> >>Of course the evaluation guides the search. Since you got here various answers >>that basically say evaluation is not that important I will add a view that >>strongly disagrees. Getting evaluation right is the most important thing for a >>program to do in a position. Under-evaluation will often lead to playing weak >>moves, but over-evaluation is almost always fatal against an opponent who has >>the right assessment. >> >>Amir > >I have some comments here > >1)Evaluation is not only to be more correct in score but to choose better moves. That's almost the same. I'm not arguing with that if there's only one good move in a position a program will choose it regardless of its evaluation, but given half a chance a program with incorrect evaluation will get it wrong. >If you add to your evaluation a random small noise the result may be worse >than if you add a constant. > There are transformations to the scores that do not change the meaning of the evaluation. Any linear mapping is one of them. Actually any mapping that preserves order doesn't matter in principle. This doesn't mean that evaluation is not important. >2)I think that evaluation is one of the main things but it is not clear that >it is more important than search. > >If you make the program 3 times faster thanks to better search rules(I mean >something eqvivalence because I do not expect better search rules to give the >same improvement in every position) then it is a big improvement(I believe that >the potential is practically bigger). > Bigger than what ? >3)If I understood correctly most of your work from Junior5 is about evaluation. > >It may be interesting to know if Junior8 can beat Junior5 with time odds of 3:1 >(it is possible to say +200% for one engine in engine matches) > >I believe that it will be the case only if the time control is long enough >but again I think that better search rules with the same evaluation may give >exponential improvement when the time control is slower. > >4)The problem is that practically you cannot seperate between the 2 and >say that the evaluation gave you x elo and the search gave you y elo because >it is possible that a change in the search rules is productive only thanks to >better evaluation. > Tell me about it :) The fact that things are complicated and interconnected doesn't mean that we should not discuss them. Amir >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.