Author: Dann Corbit
Date: 15:28:39 07/25/00
Go up one level in this thread
On July 25, 2000 at 18:23:05, Tom Kerrigan wrote: [snip] >Right. I think the only way to go is binary, i.e., "positive" instead of +0.22. >So the evaluation can be either right or wrong. I don't see any value in throwing away information. Your eval chooses ce=+1 Chess Tiger chooses ce = +340 You got it right? Your eval chooses ce = -20 Rebel chooses ce = -32765 You got it right? >Let's say you have a collection of 10,000 positions where you know which side is >winning. You run your evaluation function on these positions (which should only >take a few seconds) and get some output like: > >Eval function correct for 8,000 (80%) of the positions. > >Then you tweak the eval function and get 82%. You know your tweak was >beneficial. If your evaluation function got close to the right answer all the time, it was doing well. If it missed by 300% on average, it might be lame (see the above examples). I don't think that approach works. Centipawn evaluations are something of a crunchy continuum over the range of a short. To consider only what side of the zero it falls on (and what about zero?) is to throw away almost all of the information. How can you use this to create better decisions?
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.