Author: Dann Corbit
Date: 15:21:59 07/25/00
Go up one level in this thread
On July 25, 2000 at 17:52:08, KarinsDad wrote: >On July 25, 2000 at 14:12:46, Tom Kerrigan wrote: > >>It would be nice to make a change to your evaluation function and get immediate, >>accurate feedback. >> >>So my idea is to get a huge collection of positions of known value (i.e., "white >>has a stronger position") and then see how well the known values correlate to >>the evaluation function's scores. >> >>Does anybody have any ideas for getting a high-quality collection of such >>positions? Or any comments on this approach in general? >> >>-Tom > > >The problem I see with this idea is that although it sounds fine on the surface, >I do not see a way to implement it. > >For example, say your program says that position A is advantage white 0.2 pawns. >What happens when you change your evaluation and it now says advantage white >0.26 pawns. Did you improve your evaluation code or make it worse? > >And what omnipotent being or program decides that position A is really a 0.22 >pawns advantage for white? Ditto for the rest of the positions in the >collection. > >I could see someone doing a lot of work setting this type of thing up and then >either the person using this mis-interprets the data, the data for some >positions is incorrect, or the person changes their evaluation routine, thinks >they have improved their program, and in reality, for some percentage of >positions, has actually worsened their program's playing ability. > >If you could explain a type of implementation that might work, I'm definitely >willing to listen. Amir Ban had an interesting notion a while back. Correlate ce to expected won/loss/draw for a given position. In other words, tie the game theoretic value of a position to its ce. There are clearly dangers with this approach (missing tactical shots that the computers would see, missing positional marvels that the GM's would see, etc.) On the other hand, it should not be to hard to research it and see how it comes out. It may be that with enough data (e.g. played at least 100 times or whatever level is needed for confidence) we can be fairly certain of the value.
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.